No title

OXFORD UNIVERSITY PRESS

JOURNAL OF SEMANTICS AN INTERNATIONAL JOURNAL FOR THE INTERDISCIPLINARY STUDY OF THE SEMANTICS OF NATURAL LANGUAGE MANAGING EDITOR: PETER BOSCH (IBM Germany) REVIEW EDITOR: BART GEURTS (IBM Germany) EDITORIAL BOARD: PETER BOSCH (LBM Germany) SIMON C. GARROD (Univ. of Glasgow) BART GEURTS (IBM Germany) PAUL HOPPER (Carnegie Mellon Univ., Pittsburgh) LAURENCE R. HORN (Yale University) STEPHEN ISARD (Univ. of Edinburgh) HANS KAMP (Univ. of Stuttgart) LEO G. M. NOORDMANN (Univ. of Tilburg) ROB A. VAN DER SANDT (Univ. of Nijmegen) PIETER A. M. SEUREN (Univ. of Nijmegen)

CONSULTING EDITORS: R. BAKTSCH (Univ. of Amsterdam) D. S. BREE (Univ. of Manchester) G. BROWN (Univ. of Cambridge) 0. DAHL (Univ. of Stockholm) G. FAUCONNIER (Univ. of California, San Diego) P. N.JOHNSON-LAIRD (MRC, Cambridge) SIR JOHN LYONS (Univ. of Cambridge)

J. D. MCCAWLEY (Univ. of Chicago) B. RICHARDS (Imperial College, London) H. SCHNELLE (Ruhr Univ., Bochum) M. STEEDMAN (Univ. of Pennsylvania) Z. VENDLER (Univ. of California, San Diego) Y. Waits (New Mexico State Univ., Las Cruccs) J. VAN BENTHEM (Univ. of Amsterdam)

H. E. BREKLE (Univ. of Rcgcnsburg) H. H. CLARK (Stanford University) H.-J. EIKMEYER (Univ. of Bielefeld) J. HOBBS (SRI, Mcnlo Park) D. ISRAEL (SRI, Mcnlo Park) E. L. KEENAN (Univ. of California, Los Angeles) E. LANG (Univ. or Wuppcrtal) W. MARSLEN-WILSON (MRC, Cambridge)

H. REICHGELT (Univ. of Nottingham) A.J. SANFORD (Univ. of Glasgow) A. VON STECHOW (Univ. of Konstanz) D. VANDERVEKEN (Univ. of Quebec) B. L. WEBBER (Univ. of Pcnnyslvania) D. WILSON (Univ. College, London).

EDITORIALADDRESS: Journal of Semantics, IBM Germany Scientific Center, IWBS 7000-75, Postfach 800880, D-7000 Stuttgart 80, W. Germany. Phone: (49-711-) 6695-559. Telefax: (49-711) 6695-500. BITNET: bosch@dsolilog. New Subscribers to the Journal of Semantics should apply to the Journals Subscription Department, Oxford University Press, Pinkhill House, Southfield Road, Eynsham, OX 8 iJJ. For further information see the inside back cover. Volumes 1-6 are available from Foris Publications Holland, PO Box 509, 3300 Am Dordrecht, The Netherlands. Published by Oxford University Press

Copyright by NIS Foundation

ISSN 0167-5133

JOURNAL OF SEMANTICS Volume 7 Number 3 CONTENTS JOCHEN MUSSELER and GERT RICKHEIT

The Cognitive Resolution of Anaphonic Noun References

221

WILLIAM CROFT

A Conceptual Framework for Grammatical Categories

245

KEES VAN DEEMTER

Forward References in Natural Language

281

ERICA C. GARCIA

A Psycho-Linguistic Crossroads: Frequency of Use

301

Journal ofSemantics 7: 221-244

© N.I.S. Foundation (1990)

The Cognitive Resolution of Anaphoric Noun References JOCHEN MOSSELER and GERT RICKHEIT University ofBielefeld, Federal Republic of Germany

Abstract

INTRODUCTION In order to understand a text it is important to relate different parts to larger semantic units. This process is controlled by various cognitive mechanisms in which reference processes within and between sentences play a central part. These references are based on the knowledge of the reader or listener and are responsible for the integration of text information into the cognitive representations (Rickheit, Schnotz and Strohner 198 5, Schnotz 1988). In text processing research, different referential relations have been studied. For the cognitive resolution of anaphoric noun references it was shown by Haviland and Clark (1974) and Clark (1977) that additional time was needed in processing the so called 'bridging inferences' (see also Tanenhaus and Scidenberg 1981, Sanford and Garrod 1981, Dell, McKoon and Ratcliff 1983, Miisseler, Rickheit and Strohner 1985): (ia') Horace got some beer out of the car. (ib) The beer was warm.

(close inference)

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The cognitive resolution of references depends on whether an anaphor has previously been mentioned in the antecedent text or whether a relation between the anaphor and the antecedent has to be constructed by means of an inference. This inference process in comparison with a mere concept repetition produces an increase in reading and comprehension time. In the present paper aspects of this processing difference are examined. Recognition data are used in addition to reading and comprehension times in order to compare the mental representations generated by the different processes. Experiment 1 shows that the differences in processing time not only depend on the repetition of the concept, but are also produced by varying the semantic distance between concepts. Differences in recognition performance were not observed, indicating that the resulting text representations were similar on a semantic level. In Experiment 2 the point of rime of the inference process during text processing was determined by a word-by-word presentation. The results show that the inference process starts immediately after the reception of the critical reference concept and ends with the completion of the proposition. Finally, Experiment 3 examines whether inference processes are different after specification or generalization of a previously mentioned concept. Here the inference effect occurs with specification rather than generalisation. The recognition data indicate that it is questionable whether an inference process takes place when a concept is generalised.

222 The Cognitive Resolution of Anaphoric Noun References

(ia") Horace got some picnic supplies out of the car. (ib) The beer was warm.

(distant inference)

EXPERIMENT 1 In this experiment the inference effect is replicated under controlled conditions in order to eliminate the following possible explanations of the reading and comprehension time differences.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The referential distance between the first and the second sentence varies only through a specification that is expressed in the concepts 'picnic supplies' vs. 'beer'. In the second example there is no immediate relationship between these concepts. The relationship has to be constructed in the knowledge of the reader who has to identify 'beer' as a 'picnic supply'. This assumed inference process is not necessary in the close inference condition because of the repetition of the word 'beer'. The postulated inference is empirically supported by comprehension and reading time differences. In the distant inference version the processing times of the second sentence differ by about 180 msec (Haviland and Clark 1974) and 270 msec (Miisseler, Rickheit and Strohner 1985) compared with the close inference condition. Although it is plausible to characterise these processing time differences as differences in inference processes there are several open questions: contrary to the example mentioned above, a variation in the semantic distance between the concepts to be linked seems to be more informative (see also Garrod and Sanford 1977), especially to exclude a mere word repetition effect caused by the close inference condition. For example, when substituting 'beer' by 'drinks' in sentence (ia'), the semantic distance between 'drinks' and 'beer' still seems to be smaller than between 'picnic supplies' and 'beer'. The time difference should be preserved. Additionally it has to be checked that the sentences in both conditions are combined and integrated in an appropriate way. Only then should the reception of the close and distant inference condition lead to comparable representations for the text as a whole. In the following Experiment 1 we therefore combined on-line (i.e. the reading and comprehension times) with off-line measurements (i.e. a recognition test). Another question is at what point is the inference process performed during the reception phase. Is it already initiated and completed when reading the concept 'beer' in the second sentence or does the inference process lag behind the encoding of this concept? Experiment 2 is concerned with this question. In the last experiment different text situations are examined in which the inference process becomes necessary after specification or generalisation of an antecedent.

J. Miisseler and G. Rickheit 223

Method Subjects Thirteen female students and 17 male students of different faculties of the University of Bielefeld took part. They were paid for participating. Their mother tongue was German and their average age was 23.23 years.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

One can say that the relevant concepts in the distant inference condition are not combined in a suitable way so that time differences are derived from the two concepts existing side by side or 'in parallel' in which case they are not necessarily connected. The prolonged processing time is then due to a change of topic 0ust and Carpenter 1980, Haberlandt 1984) and/or an intended inference process, which is not necessarily performed. The characteristic integration could be missing. Even if one assumes the integration process to have taken place, it is not clear whether the text representations due to the different inference processes are comparable. Reading and comprehension time differences can only be of interest if one can assume that comparable text representations are produced after encountering both text versions. That is to say, on reception of two totally different texts one would never think of comparing reading and comprehension time differences. Furthermore, if we assume that encountering both text versions leads to comparable text representations, it has to be clarified that shorter comprehension time is not due to the repetition of an anaphor (see also Haviland and Clark 1974, Haberlandt and Graesser 1985). Accordingly, comprehension time differences are not necessarily due to a prolonged inference process but rather to a word repetition effect. In other words, short processing times would then result because of word repetitions in the close inference condition. The following experiment tries to exclude these possible explanations. By varying the semantic distance between concepts, the word repetition effect is controlled (as done by Garrod and Sanford 1977). To make sure that the integration process has taken place and the resulting comparable text representations have been attained, additional data were collected from a recognition test. However, in order to guarantee that the relevant inference process did not take place during the recognition phase, subjects were only confronted with the recognition tests after the presentation of several experimental text blocks. It should then be impossible for the subjects to remember semantic or syntactic details of the text surface. Additionally, in this experiment, we examined whether the inference process only affects the local processing time of the critical sentences or whether it is also carried over to the following sentences.

224 The Cognitive Resolution of Anaphoric Noun References

Reception Texts Sixty four-sentence texts were constructed. For each text there was a distant and close inference version (altogether 120 texts). The sentences 1, 3 and 4 of the two text versions were exactly the same; only the second sentence was modified in such a way that the inference to the third sentence was either more difficult or more easy to perform. This was achieved by using a concept which was immediately related to the word, or a concept which belonged tc the presented scenario. In contrast to the experiment of Haviland and Clark (1974), the concept repetitions are not inherent for the close inference condition. The following text examples (translated from German) make this clearer:

distant inference version: (2a) Karin drove with her car through the countryside. (2b") Karin glanced across to the mountain range. (2c) The grain was ripe and was being harvested by the farmers. (2d) The people were busy in the fields. The first sentence is the setting. The concept 'field' in the second sentence of the close inference version (2b') should be related easily to the concept 'grain' of the third sentence (2c); this is not the case with the concept 'mountain range' in the distant inference version (2b"). The definite article 'the' of the sentences (2c) marks the anaphoric relation to the previous sentence.

Recognition Sentences For each of the 120 texts three different types of recognition sentences were constructed from the second sentence: identical sentences, performed inferences and elaborations. Identical sentences represented a mere repetition of the second sentence: (2b-i) Karin glanced across to the

field.

(identical sentence)

or for the distant text version: (2b-2) Karin glanced at the mountain range.

(identical sentence)

Alternatively, the performed inferences represented a text content probably resulting from the inference process between the second and third sentence. If

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

close inference version: (2a) Karin drove with her car through the countryside. (2b') Karin glanced across to the field. (2c) The grain was ripe and was being harvested by the farmers. (2d) The people were busy in the fields.

J. Miisseler and G. Rickheit 225

'Karin glanced across the field' (or 'the mountain range') and it was stated that 'The grain was ripe', then this implies that (zb-}) Karin glanced at the grain.

(performed inference)

The third type of recognition sentences contained aspects which were clearly elaborative. With such sentences it should be tested whether the subjects were ready to relate text concepts within the recognition phase. For example, in the text it is not mentioned (although possible) that 'Karin, leaving the car, glanced at the field'. This makes it possible that (2b-4) Karin glanced at the car.

(elaboration)

Design One-half of the subjects were given 30 distant inference texts and 30 close inference texts. The other half of the subjects were given the complementary versions. This results in a two-way factorial design with the factors text version (close vs. distant) and the processing times of the third and fourth sentences following the critical second sentence. The dependent variable was the reading time for each sentence. To each text two recognition sentences were given: an identical sentence, a performed inference, an elaboration or a distractor. Altogether 60 recognition sentences and 60 distractors were presented. Half of these were mere repetitions of the text, the other half had been modified to perform inferences or elaborations. Accordingly, the recognition phase was based on a two-way factorial design with the factors text reception (distant vs. close version) and type of recognition sentence (identical, performed inference, elaboration). Each cell of the design was filled with 10 recognition sentences. The dependent variable was the recognition frequency.

Apparatus and Procedure The texts were displayed on a personal computer. The first sentence was presented in the middle of the screen. As soon as the subject had read the text, he or she had to press a button for the presentation of the following sentence. This appeared below the first sentence but without the previous sentence disappearing. In this way all four sentences were on display at the end of each text.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

In addition to these three types of recognition sentences, distractor sentences were used. They were taken from other parts of the texts (sentences (2a) or (zd)). All recognition sentences were true with respect to the text.

226 The Cognitive Resolution of Anaphoric Noun References

The subjects were not informed that reading times were measured. They were instructed to read as 'normally' as possible and to avoid repeated reading. The texts were presented in six blocks with ten four-sentence texts respectively. After each text there was a pause of 4 seconds before the next text appeared. After a text block the recognition items were given in random order. For this task the subjects had to press two other buttons (recognised vs. not recognised). They received no feedback on the correctness of their answers.

Results

Extreme reaction times (M ± 2*SD) were replaced by the average reaction time of the subject per cell of the design. In accordance with Clark (1973), reading and comprehension times were dependent variables in two ANOVAs. Text version and sentence position were 'fixed effects'. 'Random effects' were in one ANOVA the subjects, in the other the texts. The resulting F-values were transformed to min F'. Figure 1 shows the mean reading times. The analyses with subjects (FS(I,2Q) = 4.52, MSE = 1408308.43, p < 0.05) and with texts (F T (I,II8) = 64.01, MSE — 198925.22, p < 0.001) show a significant increase in reading times between the third and fourth sentence; accordingly, the min F'(i,33) = 4.22, p < 0.05 was

-

p close

3700 /

3300

:

M distant

/?

2900

2500 • 3

rd

sentence

.th 4

sentence

Figure I Average reading times of the critical third and the following fourth sentence in the distant and close inference condition.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Reception Phase

J. Miisseler and G. Rickheit 227

Recognition Phase The average recognition frequencies were dependent variables in two ANOVAs one with subjects, the other with sentences. The results are shown in Figure 2. An effect can be noted with the different types of recognition sentences (Fs(2,58) =353.13, MSE = 0.018, p < 0.001; FT(2,354) - 275.66, MSE = 0.047, P < 0.001 and min F'(2,238) = 154.18, p < 0.001). Identical sentences were more easily recognised than performed inferences and the latter were better than elaborations. As in the reception phase, the factor text version had no effect

distant close identical

performed inference

elaboration

Figure 2 Mean frequencies of the recognised sentences determined by the presented texts (close vs. distant inference condition) and the three different types of recognition sentences (identical, performed inference and elaboration).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

also significant. The main factor text version had no effect on the results (Fs(i,29) - 1.32, MSE - 35667-55, n.s.; F T ( I , I I 8 ) = 0.47, MSE = 197870.31, n.s. and min F'(i,6s) — 0.17, n.s.). The interaction of the factors proved to be significant (Fs(i,29) = 15.14, MSE = 3577I-3O, p < 0.001; F T ( I , I I 8 ) = 5.45, MSE = 198925.22, p < 0.05); min F'(i,i43) — 4.01, p < 0.05). This is due to the increased reading times of the third sentence and the shorter reading times of the fourth sentence in the distant text version in comparison with the reading times of the close text version. The Scheffe Test shows a hybrid interaction: the two inference versions of the third sentence differ (p < 0.01 with subjects, p < 0.05 with texts), whereas in the fourth sentence only a tendency can be noted (p < 0.10 with subjects, n.s. with texts).

228 The Cognitive Resolution of Anaphoric Noun References (F s (i,29) — 0.73, MS E — 0.022, n.s.; F T (i,354) = 0.68, MS E — 0.047, n - s F ' ( i , H 4 ) = o.35, n.s.).

Again, interaction between the factors is critical. Even though the 5 per cent level of significance was exceeded in both ANOVAs (Fs(2,58) — 4.02, MSE — 0.023, p < 0.05 and FT(2,354) — 3.89, MSE = 0.047, P ^ 0.05), there is no significant effect with the min F' value (min F'(2,iO4) — 1.98, n.s.). With reference to this result, the recognition frequencies of the performed inferences and the elaborations do not differ in the two inference versions close and distant.

Discussion

EXPERIMENT 2 One of the results of the first experiment was that the inference process measured in reading and comprehension times is local and does not extend to the following sentence. The question is still open as to at what point of the critical third sentence the inference is performed.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The critical reading time difference of approximately 173 msec equals the results of Haviland and Clark (1974). Apart from this, the time difference is not only due to the concept repetition but is also realised when the conceptual distance is varied. This basically replicates in German an earlier finding by Garrod and Sanford (1977) using English texts. Secondly, it was shown that the inference effect was carried over only to the third sentence and could not be found in the fourth. The general increase in reading rime of this sentence is probably due to the end of an episode, i.e. the end of the text (Haberlandt 1980, Haberlandt, Berian and Sandson 1980). Thirdly, if the recognition phase can actually be judged as a criterion of mental representations, it can be concluded that an integration process has taken place, resulting in similar representations after the reception of distant and close text versions. However, this last interpretation has to be restricted: it is only acceptable if the inference process is not ascribed to the recognition phase (see above). In addition, it can only concern the retrieval of semantic memory traces. Nevertheless, differences may exist on a text surface level as indicated by the difference between the recognition of identical sentences and the performed inferences. Finally, there is a marginally significant interaction between the type of recognition sentence and the close vs. distant text reception. Although not reliable by min F', this could indicate a weak difference in representation as a function of presentation condition. In the following experiments we will keep this interaction in mind.

J. Miisseler and G. Rickheit 229

Method Subjects In this experiment 10 male and 20 female students of the University of Bielefeld participated and were paid afterwards.

Reception Texts and Recognition Sentences Exactly the same texts as in Experiment 1 were used, except that 30 distractor texts were interspersed. The distractors had the same structure as the other texts and served to prevent the subjects from predicting the stop in the critical third sentence (see below). Accordingly, the distractor texts never stopped in this sentence. The experimental texts stopped in three controlled positions in the third sentence: (1) after the noun phrase (position 1); (2) after the verb phrase (position 2); and (3) at the end of the sentence (position 3). Each subject only encountered one of the three stops in this sentence. Additionally all texts were stopped

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Just & Carpenter (1980) interpret their results from eye movement experiments in relation to a multiple processing levels model. The information is encoded, referred to semantic contents and stored in memory. Firstly, they assume that the processing of a word starts immediately on all levels after reception [immediacy assumption). Secondly, it is postulated that the processing time of a word is reflected in its fixation time (eye-mind assumption). The eye fixation of a word is as long as one needs to process it. Even the reference process to other previously read parts of the text are reflected in the fixation time. These assumptions have been empirically and theoretically criticised. Ehrlich and Rayner (1983) postulate that eye movement merely serves to supply information to a memory which can then be referred to during complex text processing. In this way the semantic processing of a word and the resolution of reference can lag behind the encoding of a word (cognitive lag, Rayner 1978). With regard to the inference process, it can be postulated that following the first assumption the inference process starts immediately after reading the critical concept and is also finished at that point. If we follow the second assumption this inference process takes place later. These opposing hypotheses were tested with a word-by-word presentation called the moving-window technique. The text presentation was continuous and stops at critical text positions. It is assumed that the reaction times at the point of restart after a stop should reflect the necessary processing times.

230 The Cognitive Resolution of Anaphoric Noun References

unsystematically up to four rimes in different places per text. The following example shows the position of stops in the critical third sentence (marked with *). (2c) The grain* was ripe* and was being harvested by the farmers*.

Design

Procedure The texts were presented word-by-word with the moving-window technique on a computer screen. Each letter of a word was masked with a ' + ' to relieve eye fixation. One word was presented for 300 msec1 and subsequently replaced by its mask. Immediately following was the presentation of the next word. As mentioned above, the text stopped at critical positions. At the point of stopping the subject had to press a button to continue with the reading of the text. The reaction time for this task was registered. The texts were presented in six blocks containing 10 experimental texts and five distractors. The recognition test was between each block as in Experiment 1.

Results

Reception Phase Again the extreme reaction times were replaced by their average reaction time. The ANOVA shows a significant result for the factor inference (F S (I,2Q) — 8.33, MSE = 17066.41, p < 0.01; FT(i,59) = 8.08, MSE = 34969.32, p < 0.01 and min F'(i,79) = 4.11, p < 0.05). With the text versions of the type distant inference the reaction times are slower than with the close inference (see Figure 3). The position factor shows only a tendency in the analyses for subjects (Fs(2,j8) - 2.41, MSE — 17797.09, p < 0.10) and for sentences a statistical

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The position of the three stops in the third sentence was varied in the close and distant inference versions. All subjects were randomly confronted with all combinations of the two (2 X 3—) factors. The design of the recognition phase was extended by one additional factor resulting from the different positions of stops in the critical third sentence. This turned out to be a three-way factorial design with the factors text version (close vs. distant), the position of the three stops (after the noun phrase, after the verb phrase, and at the end of the sentence) and the type of recognition sentences (identical, performed inference, elaboration).

J. Miisseler and G. Rickheit 231

~ 950 u 0) «

_ __

a

distant

£ 900 ^~^^~~^-^ ^

5 850

^ /

^>o close

-^

c o

5 800

1

1

I

stop position Figure 3 Mean reaction times of the critical stops (position 1: noun phrase, position 2: verb phrase and position 3: at the end of the sentence) determined by the two versions of close and distant inference.

significance (F T (2,u8) — 3.81, MSE — 22499.94, p < 0.05). The min F' value is not significant (min F'(2,i29) = 1.48, n.s.). A similar result shows the interaction of the two factors (Fs(2,58) = 3.47, MSE — 7011.27, p < 0.05; F T (2,u8) — 2.91, MSE — 16605.93, P < °- 1 0 a n d min F'(2,i64) = 1.58, p < 0.25). However, Scheffe Tests for the stopping positions show significant results in position 1 (for subjects and texts p < 0.01) and 2 (for subjects and texts p < 0.05) with the distant and close inference condition. Position 3, however, does not.

Recognition Phase The frequency data in this experiment were based on only three to four recognition sentences per cell and subject, because we used the same recognition sentences as in Experiment 1 despite one additional factor (i.e. three positions of stops). For this reason the mean recognition frequency of about 100 recognition presentations of all subjects and sentences served as estimated values in an ANOVA which used only one observation unit per cell (see Bortz 1977, p. 396). This statistical procedure uses as error variance the residual variation in which the interaction variance and error variance of the second order are confounded. With the use of the additivity test of Tukey (1949) it was examined whether there was any interaction at all. This seems to be the case

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

750

232 The Cognitive Resolution of Anaphoric Noun References

distant close identical

performed inference

elaboration

Figure 4 Mean frequencies of the recognised sentences determined by presented texts (close vs. distant inference condition) and the three different types of recognition sentences (identical, performed inference and elaboration). The factor 'position of stops' was averaged.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(F(i,3) — 89.54, MSBa!ancc — 0.28, p < 0.001), although interpretations on the main factors and any interaction effects should be more conservative. However, the different types of recognition sentences show a significant effect as in Experiment 1 (^(2,4) = 513.50, MSResidual = 6.47, p < 0.001, see Figure 4). Additionally, the tendency to recognise sentences from the close inference condition rather than distant inference condition is higher than chance (F(i,4) = 8.79, MSU — 6.47, p < 0.05). These main effects, however, should be interpreted with their interaction (F(2,4) — 13.71, MSK — 6.47, p < 0.05). As in Experiment 1, there is a tendency to recognise elaborative sentences of the distant inference version rather than those of the close inference. Surprisingly, this effect is reversed with the identical sentences. For the critical performed inferences there seems to be no difference in the two conditions. Additionally, the interaction of the position factor and type of recognition sentence shows significance (F(4,4) — 11.82, MS]( = 6.47, p < 0.05, see Figure 5). With the recognition of the identical sentences there is no evidence that the stop position is in any way relevant, whereas with the performed inference it seems that they are recognised less when stopped at the second position and hardly ever recognised when stopped at the end of the third sentence. On the other hand, this tendency is reversed with the elaborations. There is no difference in recognition frequencies when stopped at the third position with both the performed inferences and elaborations.

J. Miisseler and G. Rickheit 233

performed inference

elaboration

Figure 5 Mean frequencies of the recognised sentences determined by the position of stops (positions 1-3) and the three different types of recognition sentence (identical, performed inference, elaboration). The factor 'inference version' was averaged.

Discussion Looking at the results of the reception phase, the inference is performed immediately after reading the critical concept (position 1) as well as after the verb phrase (position 2). The reaction time differences of about 97 msec and 56 msec at these two positions added together give almost the time that is needed for reading the whole sentence as seen in Experiment 1. This result complies with the tendency of the recognition phase to perform the inference when stopped in the first positions of the sentence. This implies that the inference is only performed by the subject when the necessary processing time is allowed. With our means of presentation this only seems to be the case when the text is stopped at the appropriate place. Accordingly, one can conclude that the process of inference is performed immediately after the reception of the critical word concept and finishes with the construction of the first semantic proposition. However, one problem with the procedure is that readers in a free situation may spend different amounts of time fixating words in different parts of the sentence. This is not taken into account with our procedure. Word-by-word presentation can possibly produce artificial effects, making it uncertain whether positions 1 and 2 reflect the time of the inference in normal reading. Indeed, results from experiments based on artificial reading methods have

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Pos. 1 identical

234 The Cognitive Resolution of Anaphoric Noun References

always to be treated cautiously (Giinther 1989). On the other hand, different studies show that a word-by-word presentation is comparable with normal reading (Just, Carpenter and Woolley 1982, Ward and Juola I982,juola, Ward and McNamara 1982). Additionally, we demonstrated in a previous study that a presentation similar to the one used here led to comparable text representations on a propositional level (Miisseler and Nattkemper 1986). EXPERIMENT 3

(4a) A robin would sometimes wander into the house. (4b) The bird was attracted by the larder. (5 a) A bird would sometimes wander into the house. (5b) The robin was attracted by the larder. They found that sentence (4b) was read faster than sentence (5b). This would seem to indicate a processing advantage for the anaphoric generalisation. However, the two sentences (4b) and (5 b) differ so that the reading times may not be comparable at all. One could proceed with the variation in a different way: when a concept is specified (in the following 'zooming-in' version), for example, 'animal-fishshark', the semantic distance of'fish' and 'shark' is shorter than that of'animal' to 'shark'.2 On the word level there are a number of experiments and theoretical discussions which have dealt with such semantic distance effects (Wilkens 1971, Meyer, Schvanefeldt and Ruddy 1975). For the present experiment examples of the following sort are taken: 'zooming-in', word level: (6a') The schoolgirl admired the fish. (6b) The shark was very aggressive. (6a") The schoolgirl admired the animal. (6b) The shark was very aggressive.

(close inference) (distant inference)

Processing time differences are to be expected because the unspecific concept animal can be split into many subcategories. The number of subcategories in the close inference version is clearly more restricted.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

In this experiment texts are introduced which make inferential references of different qualities necessary. One can vary the semantic distance of concepts when the antecedent is specified differently. The texts of Experiments 1 and 2 were principally constructed in this way. On the other hand, concepts can be generalised differently. This gives rise to the question of how manipulations of this sort induce different information processing. Garrod and Sanford (1977, Exp. 1) followed up this variation with the following pairs of sentences:

J. Miisseler and G.'Rickheit 235

If one reverses the concepts and proceeds by going from the specific to the general category 'shark-fish-animal' (in the following 'zooming-out' version), it is realised on a textual basis in the following way: 'zooming-out', word level: (7a') The schoolgirl admired the fish. (7b) The animal was very aggressive. (7a") The schoolgirl admired the shark. (7b) The animal was very aggressive.

(close inference) (distant inference)

'zooming-in', spatial scenario: (8a') Karin glanced at thefield. (8b) The grain in the field had grown. (8a") Karin glanced at the mountain range. (8b) The grain in the field had grown. 'zooming-out', spatial scenario: (9a') Karin glanced at the field. (9b) The mountain range was impressive. (9a") Karin glanced at the grain. (9b) The mountain range was impressive.

(close inference) (distant inference)

(close inference) (distant inference)

Comparing the examples (8) and (9), enough new information in the second sentences is introduced in order for the reader to make a spatial focus change of the following discourse or theme necessary. However, this focus change should be performed more easily in the close inference versions independent oi the zooming condition.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

A processing time difference can be expected only when the next highest category is activated; in this case 'fish' is associated with 'animal' and 'shark' only with 'fish'. However, because the number of main categories is restricted in generalisation as opposed to when they are specified, one does not necessarily anticipate processing time differences. With the concept 'animal' no new information is introduced at the point of text reception which the processing system does not already know. In comparison with the specification, the text does not introduce something new because it is already in the world knowledge of the reader. The concept 'fish' semantically implies 'animal'. In other words, the definite article and the concept 'the animal' are a signal for focus maintenance in the 'zooming-out' version and do not announce anything new, whereas in the 'zooming-in' version a mental focus tracking (e.g. Sidner 1983, Schnotz 1986) is probably necessary because of specification. To control this type of influence an additional variation is introduced that is neutral in this respect: the 'zooming-in' or 'zooming-out' in a spatial scenario ('mountain range-field-grain' vs. 'grain-field-mountain range'):3

236 The Cognitive Resolution of Anaphoric Noun References

Method Subjects Forty students of the University of Bielefeld took part in this experiment. They were paid for participating.

Reception Texts

Recognition Sentences For each of the 40 experimental texts recognition sentences of the type 'performed inference' were constructed. In addition 60 distractors were used.

Design Half of the subjects were given the 'zooming-in' versions, the other half the 'zooming-out' versions. Within subjects the inferential factor ('close' vs. 'distant') and the referential context factor ('word level' corresponding to examples (6) and (7) vs. 'scenario' corresponding to examples (8) and (9)) were varied.

Procedure The texts were read as in Experiment 1 sentence-by-sentence. The same procedure was used except for the fact that the 60 four-sentence texts were divided in four experimental blocks of 15 texts.

Results Reception Phase All extreme reading times were excluded from the ANOVA. As in Experiments 1 and 2 the inference effect between the close and distant condition is evident in the 'zooming-in' version (Figure 6). In the 'zooming-out' version it can only be found in the spatial scenario condition.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

From Experiment 1 and 2,40 texts were modified to correspond to the example texts of (6) to (9). Mainly the second and third sentences were changed. The first and last sentences remained. In addition, 20 distractor texts were introduced.

J. Miisseler and G. Rickheit 237

23001-

Z 2200

' zooming in1

'zooming out'

-

distant

a

in

5 2100 0 E

C •H

-"

1900

0

0

close

o--//

0 close

1800 1700 1

1600 word level

scenario

1

1

word level

scenario

Figure 6 Mean reading times of the critical third sentence determined by the inference conditions (close vs. distant), the referential context (word level vs. spatial scenario), and the 'zooming' factor (in vs. out).

Statistically one can expect an interaction of all three factors, but the results only show a weak tendency (Fs(i,38) — 2.71, MSE = 116690.43, p < 0.15; Fs(i,38) — 3.82, MSE — 60997.27, p < 0.10 and min F'(i,78) = 1.59, n.s.). Instead there is an interaction between the 'zooming' factor and the referential context in the ANOVA of subjects (Fs(i,38) — 6.10, MSE — 60116.86, p < 0.05) and of texts (FT( 1,38) = 3.89, MSE — 120331.25, p — 0.056), but not for the min F' value (min F'(i,76) — 2.38, n.s.). The main factors close and distant inference were tendentiously significant (Fs(i,38) — 3.93, MSE — 179470.63, p — 0.055 an< i FT(i,38) — 15.75, MSE — 53587.00, p < o . o o i yielding a min F'(i,59) — 3.11, p < 0.10).

Recognition Phase As in the preceding experiments the recognition frequency of the two inference conditions hardly show a difference in the 'zooming-in' versions (Figure 7). There are differences, however, in the 'zooming-out' text versions. Here, for the word level more sentences were recognised after the reception of the close inference condition. This effect is reversed in the scenario condition. The ANOVA in effect shows a significant interaction of the three factors (Fs(i,38) — 5.76, MSE = 210.20, p < 0.05, FT(i,38) — 5.30, MSE — 228.36, p < 0.05), but only a tendency with the min F'-value (min F'(i,8o) = 2.76, p < 0.15). Additionally,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

a

/

2000

•a ID

distant 0

238 The Cognitive Resolution of Anaphoric Noun References

Figure 7 Mean frequencies of the recognised sentences (type: performed inference) determined by the referential context (word level vs. scenario), the 'zooming' factor (in vs. out), and the inference factor (close vs. distant).

the interaction between the inference factor and the referential context is significant (Fs(i,38) = 10.00, MSE — 210.20, p < 0.01; FT(i,38) = 8.37, MSE — 251.25, p
Discussion The reading time differences in the reception phase prove the expected results. Garrod and Sanford's (1977, Exp. 1) findings show that concepts are processed by about 200 msec faster when generalised than when they are further specified. This is in accordance with our results on the word level. However, the findings of the recognition phase show results which lead to the assumption that the text representations after the reception of the 'zooming-out' version are different. Let us have a closer look at the critical recognition sentences of the performed inference type: they arise through the exchange of the last-named concept of the second sentence with the firstnamed concept of the third sentence. With respect to the 'zooming-out' version on the word level, this exchange went more unnoticed when they were lined up closer in the hierarchy of concept categories. On the other hand, the concepts in the distant inference condition which did not border on the hierarchy of word categories had a greater tendency to be rejected. This result would generally have been expected

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

word level scenario word level scenario

J. Miisseler and G. Rickheit 239

GENERAL DISCUSSION AND C O N C L U S I O N In summary, the above experiments give the following results. Firstly, the inference effect measured by reading and comprehension times is not bound to any concept repetition, but can also be found when the semantic distance between concepts is varied. This basically replicates an earlier finding in English (Garrod and Sanford 1977), but with slightly different form of semantic distance manipulation. Secondly, the recognition test proved not to be different between the close and distant inference condition (with one exception, see below). So the text representation constructed by the close and distant inference process seems to be comparable for the text as a whole, at least on a semantic level. This is a

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

if the subjects had not performed any inference process particularly in the distant inference condition. Experiments 1 and 2 and the 'zooming-in' versions show that the contrary is true due to the fact that reading and comprehension time decreases and no important differences were found in the recognition phase.4 In the 'zoomingout' version there is no difference in the reading and comprehension time on the word level but there is a difference in the recognition phase. This result points to the fact that the inference had not necessarily been performed in the reception phase so that different mental text representation results, and hence the subjects can distinguish between the recognition sentences. In other words, if the inference is not performed or there is no necessity for it in the texts used, as could be the case in the 'zooming-out' versions on the word level, then different text representations are still available. Consequently the inference process leads to a qualitative change in the text representations, at least in those that are retrieved in the recognition phase. For the spatial scenarios the findings are different: in this case the inference variation leads to increased reading and comprehension times in the 'zoomingout' as well as the 'zooming-in' versions. Performed inferences of the 'zoomingout' versions are better recognised when the spatial focus is extended further as is the case in the distant inference version. With the 'zooming-in' version this did not happen. A speculative interpretation would have to be that the 'zooming-out' version requires a qualitative change of the mental scenario representation when the spatial mental model is extended. This representation would not equal the antecedent one. With the spatial 'zooming-in' this change is not necessary because the new concept incorporated into the spatial mental model is only more detailed information. In both cases the reading and comprehension times can increase. Nevertheless, their mental text representations cannot necessarily be compared.

240 The Cognitive Resolution of Anaphoric Noun References

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

necessary assumption to secure the integration process combining the two critical sentences. Thirdly, the inference process starts immediately after the reception of a critical concept and ends with the construction of the first propositional unit. This result is consistent with a weak version of Just and Carpenter's (1980) immediacy hypothesis. In this way, the identification of the antecedent in the present study seems to be a similar recursive mechanism to that used, for example, in pronominal resolution (for differences between noun and pronoun resolution, see Cloitre and Bever 1988). With the encoding of a referent a search-and-match procedure is initiated which establishes the reference's relation to the preceding discourse and identifies the object it refers to (e.g. Carpenter and Just 1977, Ehrlich and Rayner 1983, but see also, for an alternative view concerning pronominal resolution, Musscler and Terhorst 1990). Fourthly, with the specification ('zooming-in') of a concept on the word level the inference effect occurs more easily than with generalisation ('zooming-out'). With the latter the inference process as such is questionable when looking at the recognition performance; at least there are differences in the semantic memory traces. Finally, both the spatial 'zooming-in' and the spatial 'zooming-out' lead to increased reading and comprehension times which reflect the inference effect; again, as far as the recognition performance is concerned, only the specification seems to lead to similar mental text representations. This indicates that the 'zooming-out' version requires a qualitative change of the mental scenario representation when the spatial mental model is extended. With the spatial 'zooming-in' this change is not necessary, because the new concept incorporated into the spatial mental model is only more detailed information. Nevertheless, the question remains as to whether the increased reading and comprehension time reflects a greater processing expense of the same process or whether it reflects an additional process. In this regard, previous assumptions from Clark and Haviland (1977) have to be made more clear: Clark and Haviland assumed a syntactic parser which dissects the present sentence into 'new' and 'given' components. The 'given' component signalled by the definite article connects the concept to the antecedent which should be no problem when the concept is repeated. The actual inference process is only accomplished in their distant inference condition because here the antecedent can only be identified when it is connected with the world knowledge of the recipient (Rickheit, Schnotz and Strohner 1985). With regard to the first example, 'beer', it should be known to be a 'picnic supply'. The 'bridging' process is then an additional process, which is not necessary in their close inference condition or at least is processed in a different manner. In the present texts this inference process is accomplished in both cases: the antecedent is not more easily identified in the close inference condition by

J. Miisseler and G. Rickheit 241

Acknowledgement The research reported here was supported by grants from the Deutsche Forschungsgemeinschaft (Ri 314/9). Thanks arc due to Julie Carson and Andrea Reuthcr for many insightful stylistic suggestions. Requests for reprints should be sent tojochen Miisseler and Gert Rickheit, Fakultat fur Linguistik und Literaturwisscnschaft, Arbeitseinhcit Psycholinguistik, Univcrsitat Bielefeld, Postfach 8640, D-4800 Bielefeld 1, Federal Republic of Germany. JOCHEN MUSSELER AND GERT RICKHEIT University ofBielefeld Postfach 8640 D-4800 Bielefeld 1 Federal Republic of Germany

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

concept repetition but rather by its semantic distance. Accordingly, one can assume a gradual change in the information process. The different processing times are then due to semantically closer distances to an already constructed mental model, whereas the processing cost becomes greater when one has to relate it to a semantically more distant mental representation. Garrod (1985) called thefirsttype of inference a pseudo-inference which tends to be processed easily and in general automatically, in contrast to true inferences which are derived from a propositional representation and need intensive processing time. It is, however, questionable whether one need assume increased processing; the effect could also be interpreted as a reduction in processing time with the close inference condition. Results from experiments on the word level show that by varying the semantic distance, different processing times can be found (e.g. Wilkens 1971, Meyer, Schvanefeldt and Ruddy 1973). A semantic facilitation or a reduction in processing time in contrast to a neutral condition can be stated even during the resolution of word ambiguities (Swinney 1979,1982). In contrast to these findings the inference effect mentioned in the above experiments could be due to a processing reduction or extension, because one cannot decide empirically between these two alternatives. Reduction or extension can only be decided with respect to neutral conditions. Whatever these look like, one can only postulate theoretically. The findings of the last experiment in particular allow us to assume that reference also depends on structural features of the critical concepts. It obviously makes a difference whether in the following sentence a concept is specified or generalised. At least the inference effect could only be seen after specification on the word level. As indicated above, one could interpret these findings with hierarchically constructed network models (see also Collins and Quillian 1972). Here one could assume that the processing time is determined by the amount of subcategories. In the specification case the number of subcategories is increased, whereas with generalisation their numbers decrease.

242 The Cognitive Resolution of Anaphoric Noun References

NOTES model. On the word level only a hierarchical class relation was realised. For example, the 'grain' is part of the 'field', but a 'fish' is not a part of an 'animal' but rather a membership of the class 'animal'. 4 But note that in Experiment 1 a marginally significant interaction (though not reliable in min F ) between the recognition test type and the close vs. distant text version was reported. This may be caused by the confounding of 'zooming-in' and 'zooming-out' in these texts.

REFERENCES BortzJ. (1977), Lehrbuch derStatistikfurSozial- language comprehension', in L. W. Gregg, wissenschaftler (Textbook of statistics for (ed.), Cognition in Learning and Memory, the social sciences), Berlin, Springer. New York, Wiley, 117-37. Carpenter, P. A. & Just, M. A. (1977), 'Reading Dell, G. S., McKoon, G. & Ratcliff, R. (1983), comprehension as eyes see it', in P. A. Just 'The activation of antecedent information & M. A. Carpenter, (eds), Cognitive Processes during the processing of anaphoric referin Comprehension, Hillsdale, NJ, Erlbaum, ence in reading', journal of Verbal Learning 109-39. and Verbal Behavior, 22, 121-32. Clark, H. H. (1973), 'The language-as-fixed- Ehrlich, K. & Rayner, K. (1983), 'Pronoun effect fallacy: a critique of language statisassignment and semantic integration tics in psychological research', Journal of during reading: eye movements and Verbal Learning and Verbal Behavior, 12, immediacy of processing', Journal of Verbal 335-59Learning and Verbal Behavior, 22, 45-87. Clark, H. H. (1977), 'Inferences in compre- Garrod, S. C. (1985), 'Incremental pragmatic hension', in D. LaBerge & S. J. Samuels interpretation versus occasional inferencing during fluent reading', in G. Rickheit (eds), Basic Processes in Reading: Perception & H. Strohner (eds), Inferences in Text Proand Comprehension, Hillsdale, NJ, Erlbaum, cessing, Amsterdam, Elsevier, 161-81. Clark, H. H. & Haviland, S. E. (1977), 'Com- Garrod, S. C. & Sanford, A. J. (1977), 'Interprehension and the given-new contract', preting anaphoric relations: the integrain R. O. Freedle (ed.), Discourse Production tion of semantic information while and Comprehension, Norwood, NJ, Ablex, reading', Journal of Verbal Learning and 1-40.

Cloitre, M. & Bever, T. G. (1988), 'Linguistic anaphor, levels of representation, and dis-

Verbal Behavior, 16, 77-90.

Giinther, U. (1989), 'Lesen im Experiment' (Reading in experiments), Linguistische course', Language and Cognitive Processes, 3, Berichte, 122, 283-320. 293-322. Haberlandt, K. (1980), 'Story grammar and reading time of story constituents', Poetics Collins, A. M. & Quillian, M. R. (1972), 9,99-116. 'Experiments on semantic memory and

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

1 The mean fixation time in normal reading is about 250 msec (e.g. Rayner 1978) depending on text difficulty and reading ability. In a pilot experiment we increased the presentation rime until a comfortable reading was achieved. 2 Note that in German 'a shark' or 'a fish' is taken as 'an animal' even in some nontechnical contexts. 3 In the spatial 'zooming-out' and 'zoomingin' an obvious part-of-relation is established with respect to the spatial mental

J. Miisseler and G. Rickheit 243

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Strohner, H. (eds), Inferences in Text ProcessHaberlandt, K. (1984), 'Components of sening, Amsterdam, Elsevier, 247-7' • tence and word reading times', in D. E. Kiereas & M. A. Just (eds), New Methods in Miisseler, J. & Terhorst, E. (1990), 'ProReading Comprehension Research, Hillsdale, nominale Besetzung: ein alternativer NJ, Erlbaum, 219-51. Mechanismus neben der rekursiven Auflosung?' (Pronominal occupations: an Haberlandt, K., Berian, C. & Sandson, J. alternative mechanism in addition to the (1980), 'The episode schema in story comprehension', Journal of Verbal Learning and recursive resolution?), Sprache und Kognition, 9, 3 7-49. Verbal Behavior, 19,635-50. Haberlandt, K. & Graesser, A. C. (1985), Rayner, K. (1978), 'Eye movements in reading 'Component processes in text comprehenand information processing', Psychological Bulletin, 85, 618-60. sion and some of their interactions', Jowrna/ of Experimental Psychology: General, 114,Rickheit, G., Schnotz, W. & Strohner, H. (1985), 'The concept of inference in dis357-74course comprehension', in G. Rickheit & Haviland, S. E. & Clark, H. H. (1974), 'What's H. Strohner (eds), Inferences in Text Processnew? Acquiring new information as a proing, Amsterdam, Elsevier, 3-49. cess in comprehension', Journal of Verbal Learning and Verbal Behavior, 13, 512-21. Sanford, A J . & Garrod, S. C. (1981), Understanding Written Language, Chichester, Juola, J. F., Ward, N. J. & McNamara, T. Wiley. (1982), 'Visual search and reading of rapid serial presentations of letter strings, words, Schnotz, W. (1986), Koharenzbildung beim and texts', Journal ojExperimental Psychology: Aufbau von Wissensstrukturen mil Hilfe von General, 111,208-27. Lehrtexten (Coherence in the construction of knowledge on the basis of textbooks), Just, M. A. & Carpenter, P. A. (1980), 'A Tubingen, University, Institut fur Ferntheory in reading: from eye fixations to comprehension', Psychological Review, 87, studien, Forschungsbericht Nr. 39. Schnotz, W. (1988), 'Textverstehen als Aufbau mentaler Modelle' (Text underJust, M. A., Carpenter, P. A. & Woolley, J. standing as construction of mental (1982), 'Paradigms and processes in readmodels), in H. Mandl & H. Spada (eds), ing comprehension',yowrMa/ of Experimental Wissenspsychologie, Munich, Psychologic Psychology: General, m , 228-38. Verlags Union, 299-330. Meyer, D. R, Schvanefeldt, R. W. & Ruddy, M. G. (1975), 'Loci of contextual effects on Sidner, C. L. (1983), 'Focusing and discourse', visual word-recognition', in P. M. A. Discourse Processes, 6, 107-30. Rabbit & S. Domic (eds), Attention and Swinney, D. A. (1979), 'Lexical access during Performance, New York, Academic Press, sentence comprehension: (reconsideraVol. V, 98-118. tion of context effects', Journal of Verbal Learning and Verbal Behavior, 18, 645-59. Miisseler, J. & Nattkemper, D. (1986), 'Visuelle sequentielle Darbietung und Swinney, D. A. (1982), 'The structure and "normales" Lesen: ein Vergleich zweier time course of information interaction experimenteller Darbietungsarten' (Visual during speech comprehension: lexical sequential presentation and 'normal' readsegmentation, access, and interpretaing: a comparison of two modes of pretion', in J. Mehler, E. C. T. Walker & senting texts), Psychologisches Beitrdge, 28, M. F. Garrett (eds), Perspectives on Mental 107-19. Representation, Hillsdale, NJ, Erlbaum, 151-69. Miisseler, J., Rickheit, G. & Strohner, H. (1985), 'Influences of modality, text diffi- Tanenhaus, M. K. & Seidenberg, M. S. (1981), culty, and processing control on inferences 'Discourse context and sentence percepin text processing', in Rickheit, G. & tion', Discourse Processes, 4, 197-220.

244 The Cognitive Resolution of Anaphoric Noun References Tukey,J. W. (1949),'One degree of freedom Wilkens, A. (1971), 'Conjoint frequency, category size, and categorization time', for nonadditivity', Biometrics, 5,523-40. Ward, N.J. &Juola,J. F. (1982), 'Reading with Journal of Verbal Learning and Verbal and without eye movements: reply to Just, Behavior, 10, 382-5. Carpenter, and Woolley, Journal of Experimental Psychology:General, i l l , 239-41.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Journal of Semantics 7: 245-279

© N.I.S. Foundation (1990)

A Conceptual Framework for Grammatical Categories (or: A Taxonomy of Propositional Acts) WILLIAM CROFT University ofMichigan

1 INTRODUCTION

Nouns: woman, dog, car, house, tree,flower Verbs: run, hit, break,go Adjectives red, large, tall

If we were testing hypotheses concerning the grammatical behavior of nouns, verbs or adjectives in English, it is most likely that we would choose example sentences containing words from this list or similar words: words denoting human beings, animate entities, or physical objects for nouns; words denoting simple translational motion or physical force for verbs; and words denoting colors or parameters of spatial dimensions for adjectives. Likewise, if we were writing grammatical descriptions of exotic languages, and trying to determine the characteristic grammatical behavior of nouns, verbs and adjectives, we would likely use the semantic translation equivalents of these words. More broadly, we would use the traditional semantic categories of persons and things for nouns, actions for verbs and (physical) properties for adjectives at the beginning of grammatical analysis. There are a large number of other words that are normally called nouns, verbs or adjectives, that we would probably not want to use to elicit exemplary nominal, verbal and adjectival behavior, however. They fall into several categories. The first and least problematic category are derived words that denote persons, things, actions and properties. Their grammatical peculiarity is that the derivational process often shifts the grammatical category or function to which they belong, producing nominalizations, participles and denominal adjectives: 'Former' verbs: explosion (v. > n.), motion (v. > n.), broken (v. > adj.), sleeping (v. > adj.) 'Former' adjectives; redness (adj. > n.), height (adj. > n.), justice (adj. > n.) 'Former' nouns: doglike (n. > adj.), industrial (n. > adj.), commercial (n. > adj.)

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Most of us have an intuition that the following words constitute 'typical' or 'prototypical' nouns, verbs and adjectives, respectively:

246 A Conceptual Framework for Grammatical Categories

It is not unnatural to assume that the behavior of these words will be related partly to their 'former' (root) category and also the actual category they fall into. For example, gerunds such as eating, as in I resent Tim's eating all the cookies, takes a genitive modifier Tim's, like a noun, but a direct object all the cookies, like a verb (see section 2 and Croft 1991). Another less-prototypical class are derived words which differ significantly in meaning from their root, but still denote persons, things, actions and properties; a few examples follow (we will call them type-changing derivations to distinguish them from the function-changing derivations just described). runner (< v.), writer (< v.), musician (< n.) buzzer (< v.), retiree (< v.) theft (< n.) stapler (< n.), [door] knocker (< v.) treasury (< n.), mint (< v.) scratch (< v.), gash (< v.)

These two classes of words, despite their non-prototypical character, still behave more or less 'normally' in terms of most grammatical inflections, due to the fact that they denote the same types of entities that are denoted by prototypical nouns, verbs and adjectives. In addition to derived words, however, there are nouns, verbs and adjectives not systematically derivable from other more prototypical terms, that nevertheless do not fit our intuitive prototype for these categories: Nouns??: day, edge, middle, amount,piece, dozen, time ('five times') Verbs??: begin, wonder, suppose, precede, try Adjectives??: same, imaginary, other, numerous, next Linguists definitely do not use such words (or their translation equivalents in other languages) for exemplary nominal, verbal and adjectival grammatical behavior, for reasons which are intuitively obvious but remain to be elucidated. If anything, these non-prototypical nouns, verbs and adjectives show a semantic affinity to other underived words which are sufficiendy different in their grammatical behavior that linguists place them into the m i n o r categories, in particular the 'leftover' category of adverbs: Prepositions: Connectives: Auxiliaries: Demonstratives: Articles: Numerals:

in, on, inside, on top of before, because, if can, must, will this, those the, a one, five

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Agent: Undergoer: Action: Instrument: Place: Result:

W. Croft 247

Quantifiers: Adverbs:

many, lots, most, any here, together, now, inside, almost, again, twice, back (as in she hit him back)

And, via the minor categories, these words also show some semantic affinity to various inflectional endings associated with the major syntactic categories: Number: Tense: Aspect: Case:

singular, plural present, past, future perfective, imperfective nominative, accusative, oblique

2 PROPOSITIONAL ACTS AND THE MAJOR CATEGORY CONCEPTS The conceptual framework presented in this paper is based on similarities in semantic (conceptual) function between grammatical and minor lexical morphemes associated with nouns, verbs and adjectives. Since this tripartite major category distinction is central to the analysis, its conceptual basis must be outlined and evidence presented in favor of it. This has been done in detail in other work (Croft 1984, 1991), and will be summarized here. In section 1, I described the prototypical nouns, verbs and adjectives as persons/things, actions and properties of qualities, respectively. This is not all there is to the definition of prototypical nouns, verbs and adjectives: the person/ thing, action or property must be used in its typical 'syntactic' functionreference, predication and modification respectively. Thus, man is a

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The aim of this paper is to provide a general conceptual framework that will accommodate all of the non-prototypical lexical items and inflectional morphemes, in addition to the prototype categories. In particular, broad parallels are drawn between 'function' words and inflections associated with nouns and those associated with verbs (and, to a lesser extent, adjectives). The purpose of this paper, then, is not to provide a direct account of linguistic structure and behavior, but instead to provide an analysis of linguistic function that can in turn be used to explain linguistic structure and behavior. This analysis has benefited from a large amount of prior research on the semantics of grammatical categories, and on prior attempts to provide an overall conceptual organization (e.g. Talmy 1978, 1988; Morrow 1986;Jackendoff 1983), although it differs in significant respects from their goals and results. It should go without saving that the framework sketched here is tentative. Nevertheless, its purpose is to provide a starting-point for the construction of a conceptual framework that is sufficiently rich, comprehensive, and well defined to use for semantic and pragmatic analyses of grammatical phenomena.

248 A Conceptual Framework for Grammatical Categories

SYNTACTIC CATEGORY PRAGMATIC FUNCTION SEMANTIC CLASS (prototypical)

Noun Referring Object

Verb Predicating Action

Adjective Modifying Property

Table 1 Reference

Modification

Predication

OBJECTS

core nouns

genitives adjectivals, PP modifiers

predicate nominals

PROPERTIES (QUALITIES)

abstract de-adjectival nouns

core adjectives

predicate adjectives

EVENTS (ACTIONS)

nominal izations, infinitives, gerunds, complements

participles, relative clauses

core verbs

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

prototypical noun only when it is the head of a referring expression such as the matt; ran is a prototypical verb only when it is the main predication, as in the man ran; and tall is a prototypical adjective only when it is a modifier, as in the tall man. If the semantic class is found in a different syntactic function, then it falls into the first of the non-prototypical categories described in section i, namely the function-changing derivations such as explosion, broken, redness and doglike. The latter terms are grammatically marked, in the sense of typological markedness (Greenberg 1966; Croft 1990); this is manifested in, among other ways, the presence of non-zero derivational morphology on the marked terms. In fact, the marked combinations of semantic class and 'syntactic' function provide the raison d'etre for a large class of inflectional categories and grammatical constructions as demonstrated in Table 1.' The unmarked forms are the prototypical or core forms, and there is overwhelming cross-linguistic evidence in support of the prototypical correlation of semantic class and 'syntactic' function (sec Croft 1991, Chapter 2 for documentation). I have used scare quotes around the phrase 'syntactic' function because these grammatical functions are definitely not formal. Instead, it is best said that they are pragmatic functions. The pragmatic functions can be defined roughly as follows: referring indicates what one is talking about, predicating indicates what is being said about it, and modifying indicates a secondary referring or predicating function (restrictive and non-restrictive modification respectively). Thus, the major syntactic categories may be defined as a prototype correlation of pragmatic function and semantic class, as follows:

W. Croft 249

Her red cheeks radiated youth and good health. Her cheeks were red.

Second, comparing attributive verb forms in prenominal and postnominal head position, the usual positions for adjectives and verbs respectively, also produces an inherent/transitory effect (Bolinger 1967): a responsible man (cf. the tall man) the man responsiblefor the accident (cf. the man sleeping on the bench)

The full differentiation can be found in comparing nominal, adjectival and verbal forms of fuss in the following examples (Bolinger 1980b: 79): Jill isfussing. Jill isfussy. Jill is afussbudget.

These observations illuminate the major category distinctions but do not provide their source. A pragmatic account, however, gives a justification. This account is based on Searle's original work on speech acts (Searle 1969). Speech acts are normally taken to describe illocurionary acts, such as requesting, promising, asserting, etc. Illocurionary acts have the whole proposition in their scope, and so they do not tell us anything about the division of the proposition into main predication, referring expressions, modifiers, etc. For instance, knowing the illocurionary force of Did Fred fix the hose? accounts for the interrogative sentence type, but not the choice of fix as the main predicate, the hose as a referring expression, etc. Searle's 1969 book is not chiefly about illucutionary acts, however. Searle identifies two other levels of speech acts that happen when a sentence is uttered (Searle 1969: 23-4). Illocurionary acts form the highest level of speech act. The lowest level are utterance acts, that is, simply producing words and morphemes. The intermediate level acts are prepositional acts, which structure the sentence

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

But what exactly are the pragmatic functions? Using a word as a noun, adjective or verb has a semantic effect in that the entity denoted is conceptualized in a certain fashion (Bolinger 1967; Wierzbicka 1986). For example, Wierzbicka notes that being Polish (adj.) represents just one of many properties of the individual in question, and so is compatible with being an American citizen, for instance; but being a Pole (n.) categorizes an individual as a whole, and so is incompatible with being an American citizen. Nouns classify the entities they denote, while adjectives merely specify a feature. Verbs represent a more transitory or fleeting event than adjectives. This can be demonstrated in two ways, first by comparing attributive adjectives to predicate adjectives, where the former is more likely to indicate a lasting property (Wierzbicka 1986: 376):

250 A Conceptual Framework for Grammatical Categories

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

and are constituted by reference and predication: 'the characteristic form of propositional acts are parts of sentences: grammatical predicates for the act of predication, and proper names, pronouns and certain other sorts of noun phrases for reference' (ibid.: 25). The act of reference is that of identifying some entity, that is, the entity that the speaker intends to talk about (ibid.: 8 5). The act of predication is the act of ascribing a property to a referred-to entity (ibid.: 100). These are exactly the pragmatic functions of reference and predication described in this section. Searle's work, based in the philosophical tradition, restricts itself to reference and predication, and does not attempt to provide an exhaustive taxonomy of propositional acts in the same fashion that Searle later provided a taxonomy of illocutionary acts (Searle 1976). In fact, the cross-linguistic analysis of the major syntactic categories already provides a third propositional act, that of modification, which appears to serve the purpose of'enriching' or adding to the description of an entity being referred to (Wierzbicka 1986: 374, and below). The purpose of this paper is to outline the remaining propositional acts; in so doing, I will have provided the promised conceptual framework for grammatical categories. Before doing so, however, we must describe the grammatical categories associated with the three major proposirional acts, referring, predicating and modifying.2 First, we have already seen that function-indicating morphemes and grammatical constructions serve to indicate the major proposirional acts. More precisely, function-indicating morphosyntax indicates marked or nonprototypical instances of the three primary propositional acts; the basic simple grammatical constructions of phrases and clauses indicate the unmarked primary propositional acts. Type-changing morphosyntax, on the other hand, serves to express a metonymic semantic relationship between two concepts. For example, agent nominalizations relate the semantic type of an action, say run, with a semantic class of individuals characterized by habitual involvement in that action (runner). Likewise, a causative construction such as the English NPi make NP2 VP relates the semantic type of an action to the semantic class which denotes actions that involve an external entity causing the first type of action to take place. A significant typological question worth exploring is what type-changing derivations are found in natural languages—that is, what metonymic conceptual relations are reflected in word formation—and why are those derivations and no others found. For example, one very frequently finds derived nouns denoting major participants in the action denoted by the root, and derived verbs which are actions systematically related to the action denoted by the root—the two types of examples just given. This question has not been systematically examined to my knowledge (but see Comrie 1985 and Comrie and Thompson 1985), so it is premature to explore the conceptual role that

W. Croft 251

3 CATEGORIZING ENTITIES The analysis summarized in the preceding section accounts for the existence of the three major syntactic categories, noun, verb and adjective, and the

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

type-changing derivations play. At any rate, type-changing derivations chiefly indicate lexical conceptual relations. As a consequence, they do not indicate a propositional act, that is, the conceptualization of experience for packaging in grammatical constructions. The patterns of relationships between semantic class and major propositional acts has accounted now for function-changing derivation and suggested a role for type-changing derivation. Turning to inflectional morphosyntax, we find that it can also be divided into two types in functional semantic terms. Stephen Anderson (1985: 172) has called these two types relational and inherent. Inherent grammatical elements characterize semantic properties of the entities denoted by the lexical roots that make up the major syntactic categories. Examples include number, gender, tense, aspect, and modality. Relational morphosyntax, inflectional or lexical, puts together lexical items of the major categories to form phrases and clauses. The morphosyntactic devices which perform these functions are as follows. Agreement, case marking or adpositions, and word order link predicates to their arguments. Voice also plays a major role in identifying argument roles. Agreement (of two types), case marking/adpositions, invariant linking particles of various types and word order link modifiers to the heads of argument phrases. The semantics and pragmatics of these forms are quite complex, as the vast literature devoted to them indicates; but it is clear that the function they perform with respect to propositional acts is to glue the major propositional acts together, so to speak: linking modifying expressions to referring expressions, and linking referring expressions to predicating expressions. We now turn to the rest of this cognitive-functional research programme, the description of the remaining propositional acts of human language. The semantic categories to be described are manifested linguistically in, among other things, relational grammatical elements other than those described in the last paragraph, inherent grammatical and minor lexical elements, and also the major category lexical roots themselves. I propose that these elements fall under no more than three additional propositional acts, which I will call the MINOR propositional acts of categorizing, situating and selecting.3 Each predication and referring expression, and possibly each modifying expression, is categorized, situated and selected, either explicitly or implicitly (i.e. from context). The following sections describe the various grammatical categories that perform these propositional acts in one way or another.

252 A Conceptual Framework for Grammatical Categories

Nouns: gender, classifiers (found with numerals and verbs) Verbs: stem categorization by transitivity and starivity Adjectives: intensifiers (?) The categorizing function that these inflectional markers play must be understood in terms of the concept of a (physical) background dimension. The hypothesis behind the notion of a background dimension is that human beings understand entirites only in the context of some structuring dimension of experience. Those structuring dimensions are therefore of a different ontological character from entities themselves. Instead of coming into existence and passing away, or changing, they are always present and always used as the background for conceptualizing entities (though the degree to which such reference is grammaticalized varies from language to language).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

derivational (function-changing and type-changing) and inflectional (relational) morphosyntax associated with it. However, it does not account for the pragmatic function of the lexical items themselves, that is, noun, verb and adjective roots. Although this problem appears to be as trivial as its answer is (I hope) obvious, it must play a foundational role for any conceptual or functional theory of language (or any theory of language, for that matter). The answer that should be obvious is that the function of the basic lexical vocabulary of human languages is to categorize persons, things, actions and properties in the world. This function is identical across the three major syntactic categories, though of course the conceptual features by which this categorization is made differ quite substantially.4 The categorizing function provides a conceptual explanation for the creation of the vast taxonomy of objects, events and properties that makes up the large part of the vocabulary of human languages. The categorization function has been discussed extensively in the psycholinguistic and anthropological linguistic literature (for a summary of some of the more significant research, see Rosch 1978 and Berlin 1978). The basic grammatical similarity is that all three categories consist of root morphemes. There are few if any languages in which the morphological nature of noun roots, verb roots and adjective roots is consistently and significantly different. Even languages with unusual root morphological structure tend to display the structure across syntactic categories. For example, the consonantal root structure of Semitic, in which the root consists of a sequence of two to four consonants and inflection and derivation are accomplished by the insertion of vowels and additional consonants, is found in all major syntactic categories. In English, stem vowel change is found in both nouns and verbs, though this is more a relic of ancient historical processes rather than a characteristic of contemporary English root structure. In addition to the existence of root morphemes, there are a number of inflectional categories that also appear have a categorizing function:

W. Croft 253

(a) 'stick-like' (one-dimensional) vs. 'flat object' (two-dimensional) vs. 'round object' (three-dimensional); (b) rigid vs. flexible, e.g. 'clothlike object' (two-dimensional flexible), 'ropelike object' (one-dimensional flexible); (c) 'solid' (count) vs. 'liquid' (mass) (see section 5); (d) single vs. plural or scattered objects (plurality implies distribution through space); (e) augmentative vs. normal vs. diminutive (size, though see section 3.1). This is also true of some noun gender systems that categorize inanimate objects, to the extent that the semantic basis of a categorization can be recovered. However, it should be noted that noun classification systems also categorize in terms of animacy (human vs. animate vs. inanimate) and function (implement vs. other) as well as spatial properties. These categories involve the classification of objects by the type of experiential domain they belong to—e.g. human interaction, earing and drinking, manipulation of objects, etc.—and suggest that Langacker's broader conception of background dimension must be invoked to account for these patterns.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The three major categories of noun, verb and adjective each has its own background dimension; this serves to distinguish them conceptually as well as pragmatically. The primary background dimension(s) for objects (prototypical nouns) is space. Langacker (1987a) argues that the background dimension (or base, in his terminology) of a noun is any experiential domain, not only space; however, space is clearly most basic, and this is manifested in the grammatical categories to be described below. The background dimension for actions/states (prototypical verbs) is slightly more complex; it is both time and causality, since events arc conceptualized in terms of linear causal chains (see Croft 1991, in press, for the evidence for this point of view). Prototypically, causal sequence coincides with temporal sequence, and even vice versa; but of course it need not be the case that the two coincide. The background dimension for physical properties (prototypical adjectives) is more abstract: it is quantity itself, that is, a quantity scale. All of the prototypical adjectives are scalar adjectives (even the color terms fall on a color scale, though in a more complex fashion). The notion of a background dimension will be crucial for the organization of the conceptual functions of situating and selecting; but it also sheds light on the grammatical categories that serve as classifiers in some way or another. This is perhaps most clear in the classifier constructions found with numerals in many languages and with some verbs in some languages, such as Atsugewi (Talmy 1972) and Navajo (Witherspoon 1977). Friedrich (1970) has called these 'shape categories', because they represent a conceptualization of objects in terms of spatial dimensionality and other properties that make crucial reference to spatial features (cf Allan 1977):

254 A Conceptual Framework for Grammatical Categories

Categorization systems for verbs are more difficult to come by than for nouns. Most verbal categorization is covert. However, cross-linguistically it appears that the two dimensions that are encoded overtly are transitivity (transitive vs. intransitive) and stativity (process vs. state). (Again, an originally semantic categorization may become more arbitrary as it breaks down through language change.) These generally are not independent categories; they appear to represent a conceptualization of actions in one of the following three causalaspectual types (see Croft, in press, for more extensive discussion):

Although transitive states exist (e.g. love, see, resemble), cross-linguistically they are extremely variable in their expression, suggesting that they do not fit well into the conceptual structure of events (in some languages, they are expressed by adjectives or nouns). An example of more or less overt categorization of verbs can be found in Quiche, a Mayan language (other Mayan languages have similar systems). If verbs occur in final position, they must take a suffix, which is -Vh for transitive verbs and -ik for intransitive ones. In addition, many intransitive verbs subdivide into stative verbs with -(V)l and non-stative verbs without -(V)l. Finally, there is a resultative passive which also has a stative -(V)l form. The significance of verbal categorization for our present purposes, however, is that it is based on causal and temporal structure. This supports our hypothesis that the categorizing function, at least as it is expressed in non-lexical morphemes, is based on the background dimensions. There is no clear candidate for adjective classifiers in human languages. However, if we take the concept of background dimension seriously, then we can argue that intensifiers such as very (hot), and their opposites, downtoners such as a little (hot), play the same role for adjectives as noun and verb categorization does. We suggested that the background dimension for adjecrives is the quantity scale itself, since physical properties are defined in terms of a scalar dimension in a 'property space' (not unlike the phase space of physicists). If that is true, then a categorization of properties on the quantity scale could be made in terms of where the property is found on the scale: high, low, or in the middle (i.e. the standard value in the relevant context). This is indicated by intensifiers and downtoners: no matter what dimension of the property space is indicated by the adjective, very indicates a more extreme value and a little or somewhat indicates a less extreme value. There are some crucial weaknesses with this argument, however. First, the adjecrives themselves indicate a position on the scale: tall is higher than the standard on the vertical dimension and short is lower. This suggests that adjecrives themselves categorize the property by its position on the scale. The

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Causative (transitive): / broke the plate. Inchoative (intransitive process): The plate broke. Stative (intransitive state): The plate is broken.

W. Croft 255

Table 2 Objects

Events

Properties

LEXICAL

noun roots

verb roots

adjective roots

GRAMMATICAL

gender, classifiers (numeral, verbal)

causal-aspectual verb classification

adjective roots (?) intensifiers/ downtoners (?)

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

adjectives also categorize the property by which dimension in the 'property space' is being referred to: height, width, weight, color, etc. This is more apparent in the use of one of the members of an antonym pair with measure terms (six feet tall, two inches thick) or the constructions of the form 'Numeral Measure Preposition Property-Noun' found in languages without adjectives (akin to die English expression i 00pounds in weight). This evidence suggests that the lexical roots themselves are defined straightforwardly in terms of the background dimension of quantity (and also the dimensions of the 'property space'). Intensifiers and downtoners only serve the purpose of elaborating the categorization of properties found in adjective roots. This raises the issue of whether the categorization performed by noun and verb roots themselves is ultimately based on shape and causal-temporal structure respectively. It is clear that to a great extent it is, but to some extent it is based on an enormous number of other conceptual features. Thus, although we may assert that the root morphemes of prototypical major category words perform the conceptual function of categorization, we can safely assert that only grammatical morphology classifies largely on the basis of the relevant physical background dimension. We summarize the organization of categorizing morphemes in Table 2. Although space and time have a privileged status in conceptual structure because of the role that they play as background dimensions to ordinary objects, they too can be conceptualized as entities themselves, and thus categorized. There are a large number of common nouns that refer to elements of space and time. The largest category of common nouns that refer to space and rime are measure terms such as foot, mile, day, week and year, they function to individuate units along spatial and temporal dimensions, and that function is discussed in section 5. Other than those, however, there are very few terms which categorize regions or points of space or intervals inrimetaxonomkally. There are a few significant exceptions. The clearest exception is the categorization of days of the week or periods of the year according to work or ritual: weekend, workday, holiday, vacation. The only other category of common nouns which appears to play a role in categorizing regions are the common nouns of landscape and celestial phenomena whose features are used to categorize

256 A Conceptual Framework for Grammatical Categories

regions in space and are found in locative expressions with great frequency (Greenberg 1974): mountain, meadow, valley; earth, sky, ocean. These are entities

that by their nature happen to be suitable for categorizing spatial regions due to their (relatively) permanent position over rime.

4 SITUATING ENTITIES 4.1 Situating in a Physical Dimension

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The next minor propositional act is that of situating an entity in a background dimension. There are a number of different ways in which situating is performed in natural languages, and it turns out that in this case, the grammatical similarities across entity type (object and event) are, with one exception, quite strong. We begin with objects and events, and turn later to the more questionable case of properties. There are a large number of semantic strategies to situate objects and events. We may begin to analyse the strategies by classifying the situating functions into two types, those that situate the entity 'directly' in the background dimension and those that situate the entity relative to another entity. The former type will be called deictic. For objects, the deicrics are the deictic demonstratives, such as English this and that. The deictics modify an object, or refer to an object, in terms of its spatial location, defined by the location of the participants in the speech act. Other languages involve more sophisticated sets of distinctions, including near speaker, near both speaker and hearer, above or below speaker, or visible or not visible from the vantage point of the speech situation. All of these have in common the use of the spatial location of the speech act as the 'origin' for the coordinate system for situating the object in space. The second situating function is relational, in which there is a separate expression naming some other entity with respect to which the referred-to entity is located in the background dimension. The 'other entity' can be the background dimension itself, e.g. space in the case of objects. Space can be referred to deictically, with locative deictics such as English here and there, or in terms of some point whose position is established absolutely in space. The latter are manifested grammatically as proper names of points or regions of the landscape, geographical or cultural, such as San Rafael, Lake Michigan, the Sears Tower, etc. We will refer to these as deictic-location and absolute situating, respectively. Finally, an entity can be situated with respect to another entity which is not part of the background dimension. This function is performed by spatial adpositions or case markers, such as to, in, over, etc.3 The grammatical parallels between situating objects in space and situating

W. Croft 257

follow. Unfortunately, there is a significant exception to this grammatical parallelism between objects in space and events in time in the pure deictic category: there is no obvious grammatical parallel between demonstratives and tense markers, the pure deictic indicators of temporal location. The linguistic reason for this is clear: tense markers almost always arise from aspectual verbal markers, while demonstratives are ancient and do not have any particular knowifesource. Nevertheless, their semantic function is the same. An. apparent difference between objects and events arises due to the role of causal structure in events. In addition to being temporally related to each other, events are also causally related to each other. Although semantically distinctevent x can follow event y without y having caused x—causality and time are naturally correlated in that frequently the expression of a temporal sequence of events in a narrative invites the inference of some causal link between the two events. The nearest analog to the time/causality distinction for objects is the distinction between simple co-location or adjacency of objects and support or attachment. An object x can be at or near object y, without contact or attachment, but if object x is on or in y, x may be attached to y or at least y supports x, and so if something happens to y, such as being moved in space, then x is also affected (it moves too, or it falls off and is no longer on or in y). Hence, contact or attachment leads to a causal as well as spatial relationship between objects. In fact, contact vs. non-contact is a significant distinction among spatial prepositions (Stassen 1985: 32); and Herskovits (1986) notes that support and attachment play a significant role in defining the English preposition on [the book on the table, the picture on the wall) and also some uses of in(the light bulb in the socket).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

events in time have been frequently noted (see e.g. Traugott 1978). This is abundantly clear in the case of the relational situating of events. Temporal deictics (now and then) are parallel to locative deictics, often based on the same morphemes historically (as in then). There are also proper names for points or intervals of time, based on natural time cycles (moon and sun), or cultural time cycles (days of the week, the counting of the years) and significant events [the Civil War, the Cezanne Exhibition, the Fall ofRome). Finally, one can temporally or causally situate events relative to other events by means of temporal connectives such as before and after and also causal connectives such as because. Just as one situates a box with respect to the table by saying the box under the table, one situates the event of washing the car with respect to the event of it raining by saying / washed the car after it rained. Temporal and causal connectives, as well as temporal adpositions, are very frequently derived from spatial ones (see e.g. Genetti 1986), demonstrating the cognitive parallels between the two domains and also inspiring 'localist' theories of non-spatial relations (e.g. J. Anderson 1971; cf. Langacker 1982). In addition to the usual temporal connectives, there are also verbs which express a temporal relationship between events: precede and

258 A Conceptual Framework for Grammatical Categories

Table 3 Situating an entity in a physical dimension

Objects (Space)

Events (Time)

Deictic

demonstratives

tense

Relative to: Background points: Deictic-Location

locative deictics

temporal deictics

names of places

names of times

spatial adpositions

temporal/causal connectives and adpositions

Absolute Other entity

9\^

In the case of proper names, again one finds proper names almost solely for human beings, and that will be accounted for in section 4.2 in the same way as I will account for personal pronouns. The use of proper names for pets, deities, natural forces, etc. appears to be an anthropomorphization of those entities, that is, a sort of metaphorical extension of the use of proper names chiefly for human beings. Thus, there is a loose correlation between taxonomically categorizing lexical items and entities on the one hand, and deictic terms/proper names and points, regions or intervals of the relevant background dimension on the other (though

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Hence, causal or potential causal relations play a role in spatial conceptualization as well. Thus, we have the characterization of situating an entity in a physical dimension, accounting for the subsequent subclasses of nouns and verbs, minor syntactic categories, and inherent inflections as depicted in Table 3. At this point one can observe an interesting grammatical correlation. Deictic terms and proper names tend to name points or regions in the background dimensions, while common nouns describe objects or events—as one would expect from the categorizing function. However, there are a number of exceptions to this correlation. Demonstrative pronouns refer to objects, not points in the background dimension, in contrast to locative deictics. However, it is worth noting here that in many languages, demonstrative pronouns are either locative deictics or demonstrative adjectives with an empty or vague noun such as 'person', 'thing', 'body', or 'one' compounded with it. In other words, they do not appear to actually name objects deictically. Likewise, tense never occurs except as a modifier of a verb, and in fact is frequently absent from the grammatical categories of a language. (Another apparent exception, personal pronouns, will be dealt with in section 4.2.)

W. Croft 259

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

see the end of section 3). The background dimension functions as known or given reference points, hence the use of deictics and proper names. The entities being referred to are not necessarily given, hence their identification through categorizing and situating. Now, let us turn to properties and see how well we can extend this framework to them. It turns out that there are parallel forms for all of these examples if one takes the physical background dimension for properties to be quantity or degree, although die parallels are not quite perfect. But perfect parallelism is not to be expected, at any rate: there are important ways in which space, time and quantity differ, and there are important ways in which objects, events and properties differ. Also, adjectives are in some sense not as fundamental or 'major' a category as nouns and verbs are, and so it might be expected that the inflectional apparatus and minor categories that serve adjectives may not be as rich as those for nouns and verbs. What is the adjectival counterpart to deictic situating? Deictic situating is the location of an entity in its physical background dimension with respect to the speech act. We can find a counterpart only by extending the notion of deictic situating outside the speech act to other context-based factors than simply the speech act's location and rime. The adjectival counterpart to deictic situating terms seem to be the basic adjective forms themselves, that is, the relative extent terms such as big/small, high/low, fast/slow, young/old, that form the core of the category 'adjective' (see Dixon 1977 and below). As has been often noted, these terms describe the degree or measure of the value of the property only with respect to some standard: a big mouse is not necessarily bigger than a small elephant, or someone may be tall for an adult but short for a professional basketball player. In other words, there is some context-dependent standard which determines the point at which short becomes tall, slow becomes fast, etc. The basic property antonyms simply name a direction away from the standard point: tall means towards the upper end of the scale from the standard point and short means towards the lower end of the scale from the standard point. Hence there is a deictic element (or if one does not like this extension of the term, a 'context-determined' element) to the basic property antonyms, as well as an element indicating what property is being attributed (size, height, speed, age, etc.). This analysis suggests that adjectives should be minor categories or even grammatical inflections (our proposed analysis for adjectives as categorizing elements also suggested this). The semantic parameters that we are analysing here are supposed to be just those which give rise to closed-class syntactic categories and inherent inflections. In many languages, basic relative extent adjectives are just that. Many languages do not have an open class of adjectives as English does; instead, there is a small class of lexical roots which have distinct syntactic behavior that merits their being called 'adjectives'. This closed class

26o A Conceptual Framework for Grammatical Categories

4.2 Situating in Mental Space Up to this point, I have described situating with respect to the physical background dimension. However, this is not the only kind of situating that is found in natural languages. A speaker also situates objects and events with respect to

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

has been found to consist primarily of just those relative extent antonyms which require a contextually determined standard for evaluation: large/small, long/short, new/old, slow/fast, etc. (Dixon 1977).6 In addition to the actually rather widespread occurrence of a closed class of adjectives, there is an obviously related inflectional category which otherwise does not fit into this conceptual framework, namely diminutives and augmentatives. Although they are often associated with social and emotional connotations, diminutives and augmentatives appear to have a source in what is perhaps the most basic adjectival antonym pair, big/little. Despite the semantic similarities, the adjectival analog to pure deictic situating appears to be quite different from its nominal and verbal counterparts. In the case of adjectives, the deictic situating semantic feature is lexically fused with the adjective itself: one could imagine a language with words that meant simply 'size', 'height', 'age', etc., which each took affixes or particles that meant 'greater than the contextual standard' or 'lesser than the contextual standard'. However, to my knowledge there is no natural language which morpholexically separates these two components of meaning. This is similar to the quite common phenomenon of the fusion of demonstratives and categorizing elements, such as the gender/class markers, mentioned above. If anything, the odd inflection out is verbal tense: it does not resemble adjectives or demonstratives, and in fact its most common grammatical source is from aspect marking. Nevertheless, the semantic parallels between verbal tense and nominal deixis are quite clear. The relational situating expressions are more straightforward, being the most characteristic of adjectival inflections: the comparative and equative constructions. In an utterance such as She is as tall as/taller than Bill (is), the speaker is specifying the measure of her height relative to the measure of another height, namely Bill's height. This is the situating of Bill's height on the quantity scale relative to another height on that scale. The situating of a property in terms of an absolute or deictic measure on the scale is done by means of a construction indicating the exact value or some deictic term pointing to a gesture: lam 6feet tall or Joey is this [gesturing] tall. The proper names for values on an adjectival scale are the numerals, which provide a unique specification of a point on the quantity scale. Thus, we may add the minor subclasses, closed classes and inherent inflections or constructions to our framework, as shown in Table 4.

W. Croft 261

Table 4 Situating an entity in a background dimension

Adjectives (Quantity)

Pure Deictic

core adjectives, diminutives intensifies?

Relative to: Background point: Deictic-Location Absolute

numerals comparatives, equatives

the metaphysical frame of reference, that is, the epistemological, ontological, and psychological status of the state of affairs that the speaker is describing with respect to the speaker's mental state. These phenomena are those for which the mechanism of possible worlds has been invoked in formal semantic analysis. More recently, the notion of a 'mental space' has been developed by Fauconnicr for the same phenomenon (Fauconnier 1985). Here also, there is a semantic parallel between nominal and verbal grammatical categories, which Fauconnier exploits in his analysis, and there is a distinction between deictic and relational situating. I will begin with a brief discussion of the relevant verbal categories, which are typologically the most developed, and then argue that there is a semantic and typological parallel between at least some of the verbal categories and certain nominal categories. The verbal category of modal auxiliaries or inflections express the deictic situating of events in mental space. The 'point of origin' is taken to be the known, actual, existing world according to the speaker (or better, mutually believed by the speaker and hearer). In John might be at the station by now, the uncertainty expressed by might is uncertainty of the speaker's knowledge, not of anyone else (in fact, the hearer might know for a fact ifJohn is at the station). Conversely, the simple sentence John is at the station now expresses the speaker's certainty that John is at the station. The same is true of the other epistemic modals and the epistemic adverbs (possibly, probably, etc.). Turning to the deonric modals, in the utterance You ought to quit yourjob, it is the speaker's idea of the way the world ought to be, not anyone else's (including the hearer). Finally, in the attitude markers found in some languages, and the attitude adverbs of English, such as fortunately, foolishly, etc., it is the attitude of the speaker that is being expressed (by default, see below). Thus, the modal auxiliaries and adverbs and related constructions (optative, and also hortative, imperative, and other speech act types, in which the speaker is expressing

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Other entity

(demonstratives)

262 A Conceptual Framework for Grammatical Categories

the 'owner' of the mental space (e.g. It seems to me that you are getting upset), and

above all, they take complements denoting events rather than arguments denoting objects. The use of world-creating predicates for the metaphysical situating of events with respect to some mental state of a person appears to fit into the conceptual framework outlined here quite well. They also have an obvious semantic affiliation with modals, attitude markers, and speech act types; for example, these predicates require an irrealis modal verb form (such as the subjunctive) in the complement clause. Finally, it is worth noting that in the case of English adverbs of attitude, there is the possibility of a relational expression for situating the attitude in the mind of someone other than the speaker: Fortunatelyfor Harry, he left before anything happened. In some cases, cor-

responding 'attitude-creating' predicates are possible in English: Harry was lucky/fortunate to leave before anything happened.9

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

his/her suggestion, command, etc.), are all 'deictic' in that it is the speaker's own viewpoint that is being expressed through the modal, attitude, speech act, etc. form. Finally, from a linguistic point of view, all of these categories are frequently collapsed into a single grammatical category or cluster into related categories (Palmer 1986). Mental spaces are based on human beings, that is, individuals who are capable of mental states; so the 'points' or 'locations' of mental space are the individuals who possess the beliefs, desires, attitudes, etc., or who produce the commands, requests, etc. Thus, the 'points' in the background dimension of mental states are the human beings themselves, and the deictic and absolute terms that refer to them are the personal pronouns and proper names for human beings respectively. For this reason, personal pronouns and proper names for human beings are parallel to deictic terms and proper names for places, times, and numerical values on a scale. Human beings function as reference points in the mental spaces they create as well as functioning as entities to be talked about.7 This is part of the reason that deictic terms and proper names for humans exist. Of course, it is not the only reason these words exist", as entities to be talked about, individual human beings and especially the speaker and hearer are highly salient. I believe that this is a case of multiple motivation, and so neither explanation should be invoked to the exclusion of the other.8 The relational terms that appear to be used to situate events with respect to points on the background dimension are the complement-taking verbs which create possible worlds, the 'world-creating' predicates (McCawley 1981: 32640; see below for complement-taking verbs that are not world-creating predicates). For example, in the sentence John believes that a pig will be elected President ofthe United States, the complement-taking verb creates a mental space which is established relative to the subject, John. World-creating predicates are by no means core verbs: they do not describe physical processes, they arc frequendy stative, they frequently exhibit non-prototypical case-marking for

W. Croft 263

Table 5 Situating an entity in a mental space

Events (Metaphysical)

Deictic

modals, attitude markers, speech acts, sentential adverbs

Relative to: Background: Deictic-Location Absolute Other event

world-creating predicates personal pronouns names of humans conditionals

We now turn to the parallels between the mental situating of events and the mental situating of objects and properties.-The natural language category which appears to be that most closely associated with the mental situating of objects is that of definiteness and referentiality, particularly as expressed by

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Finally, we turn to the relational situating of events with respect to other events. This turns out to be the slot in our framework for a construction type which odierwise tends to dangle in semantic classifications of grammatical constructions: conditional constructions. The conditional construction situates an event, the one denoted by the consequent, with respect to another event, the antecedent, in a mental space (Fauconnier 1985: 31-2). The status of the consequent is dependent solely on that of the antecedent and the perceived relationship between the two of them, since the antecedent has no direct connection to the real world, being possibly hypothetical or even counterfactual.10 We thus have an outline of the metaphysical situating of events illustrated in Table 3. The significant difference between this table and the preceding tables for background physical dimensions is that world-creating predicates are used only with relational situating with respect to points in the mental space, while conditionals are used for relational situating with respect to other events. Perhaps a better analogue to spatial adposirions and temporal connectives is the inflection of mood: the realis/irrealis (indicative/subjunctive) forms of verbs which are used in bodi the complements of world-creating predicates and in conditionals. The complements of modal auxiliaries may also inflect for realis and irrealis, making mood cover deictic as well as relational constructions; and mood inflection plays an important role in the antecedents as well as the consequents of conditionals. One can hope that these asymmetries can be explained in such a way as to leave intact the overall parallels between situating in a physical dimension and situating in mental space.

264 A Conceptual Framework for Grammatical Categories

Anyone can do this. * Anyone HA this. A falcon is a predatory bird.

*A falconjlew overhead. [* for generic reading of subject] The closest analogs for objects to relational situating in a mental space (in English, at any rate) fall into two types: the so-called syncategorematic adjectives such as fake, imaginary, alleged, hallucinated, sought-after, and so on (as opposed to real, actual), and the fictional context nouns (also known as 'picture nouns'), such as picture (of), book/movie/story (about), dream (about), etc. (Fauconnier 1985: 10-13). We also find the use of articles to indicate status in mental space, of course, very much like the realis/irrealis mood distinction which cross-cuts all of the verbal metaphysical situating types. Possessive pronouns and proper names can be used for relating entities to points in the background dimension, namely human beings: my amplifier or Tina's dog to refer to the stereo/dog that I and Tina respectively want, are looking for, etc. Possessive pronouns in their more prosaic uses can perhaps be thought of as setting up a socially defined space of ownership; but as with regular personal pronouns, their existence is probably motivated by other factors as well. There seems to be no analog to that particular use of the conditional. The more limited variety of nominal mental situating terminology is listed in Table 6. When we turn to adjectives, there seems to be no obvious parallel between events and properties. Instead, adjectives appear to take on the characteristic verbal metaphysical inflections if they are predicated, and as modifiers they 'inherit' their metaphysical semantic features from those of the head noun. The reason for this is that metaphysical semantic features arc properties of the noun phrase or sentence as a whole, rather than just the head noun or main verb, although they arc frequently morphologically expressed on

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

articles and indefinite pronouns. Consider the semantic foundation of the definiteness categories: definite, specific indefinite, and non-specific indefinite (including intensional objects, and generic indefinites). The definite/specific indefinite distinction involves an epistemological criterion: if the object is identifiable (known) to speaker and hearer, then the definite form is used; if it is not, the specific indefinite form is used. Thus, the use of the and (specific) a is based on the knowledge spaces of speaker and hearer, and moreover are 'deictic' (context-determined) because of their reference to speaker/hearer. The distinction between specific and various non-specific indefinites is largely a matter of other types of mental spaces: the indefinite article of Gladys wants to buy a house indicates (in its non-specific reading) that the object of the search or desire is intensional, existing only in or Gladys' wish mental space. Similarly, the use of the so-called 'universal any' and 'generic a' is restricted to contexts which are not specific realis events (see e.g. Davison 1980):

W. Croft 265

Table 6 Situating an entity in a mental space

Objects (Metaphysical)

Deictic

articles

Relational Background:

syncategoremaric adjectives, fictional context nouns possessive pronouns

Relational

names of humans (in possessive form) N/A?

Other entity

the head noun and the main verb. Predicate adjectives are Treated like other predicates, namely main verbs. Attributive adjectives function as modifiers. Modifiers perse, as opposed to predicates and arguments, do not appear to have a metaphysical status independent of predicates and arguments. Hence one does not find independent natural language grammatical categories expressing the metaphysical status of modifiers. In sum, there does seem to be a semantic parallel between the category of definiteness in nouns, or more precisely noun phrases, and modal or modalityrelated categories in verbs and complex sentence constructions such as worldcreating predicates and conditional constructions. There is one syntactic typological parallel between auxiliaries and articles, the pure deictic metaphysical situating forms. Modal auxiliaries tend to be fixed in their syntatic position either next to the main verb or in characteristic clitic positions—first, second or last in the utterance (Steele 1975). Definite articles also tend to be found either next to the head noun or in characteristic clitic positions—first, second or last in the noun phrase. This typological parallel is most notable in the cases in which the auxiliary or the article is found in second position (see Table 7). These positions—next to the head and 'clitic positions'—seem to be the Table 7 Bulgarian articles: dve -te

novi

knigi

two-the new.PL book.PL 'the two new books' Papago auxiliaries (Zepeda 1983: 20): S-hottam

'ait

cipkan 'ani

quickly IMPF.iSG walk I 'I am/was walking quickly.'

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Deictic-Location

266 A Conceptual Framework for Grammarical Categories

ones favored by grammatical morphemes which semantically are modifiers of the argument/proposition as a whole; and those semantic categories are just the metaphysical or mental space-creating categories we have just discussed.1'

5 SELECTING ENTITIES

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

As we stated in section 3, speakers possess a large vocabulary which effectively classifies the entities of the world. Under normal situations, that means there is both a method of categorization, giving rise to what is called a 'type', and some non-zero, non-singular set of entities, called 'tokens', which are classed under the type. Thus, any common open-class lexical item taken by itself is potentially polysemous: it may be used to denote the type or a token of the type (cf. Nunberg 1979). In addition, in the normal situation in which there exists more than one token of the type, the common open-class lexical item, may denote any one of those tokens or sets or subsets thereof. A speaker must SELECT just which of these possibilities is actually being denoted by a major-category term. This is the third and last minor prepositional act. The first distinction that must be made is between reference to type and reference to a token. The former is generally described as 'generic'. The type/ token distinction is just as valid for events as for objects: one can refer to the generic activity of eating as well as a specific event of eating. Generic uses of nouns are sometimes expressed by articles (or their absence) in those languages which have them. The use of the articles (or lack thereof) varies considerably: for example, English prefers to use the plural form without articles or the socalled generic a for generic reference, while French frequently uses the definite article. Thus, generic expression of nouns manifests itself in the article system, possibly as the unmarked member. In the verbal system, one finds generic inflections sometimes related to durative/iterative aspect markers. In English the generic is expressed by the simple present tense (in contrast to the progressive), another instance of the least-marked inflection being used for generic expression. It is difficult to see what the comparable category for properties is, though in English one finds unmarked adjective forms that appear to be referring to types: Red is my favorite color. (On the other hand, red may be functioning as a noun in this construction). The expression of reference to type as opposed to token is an area which merits further typological exploration. Selecting a token rather than the type implies a more complex range of choices. There are important semantic parallels between objects and events with respect to both the selection of object tokens and event tokens. To discuss them, we must digress briefly to describe certain lexical characteristics of nouns and verbs which interact with the semantic parameters we are about to describe. Nouns and verbs fall into roughly parallel lexical semantic subclasses.

W. Croft 267

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Nouns can be divided into count and mass. This 'syntactic' distinction is a manifestation of a semantic one between what we may call 'naturally individuatable' and 'not naturally individuatable' objects (more precisely, object types). Naturally individuatable object types are those whose tokens come in natural units, such as children, dogs and tables; these appear to be the prototypical nouns. Not naturally individuatable object types are those whose tokens do not come in natural units, such as liquids and gases. There is an unclear intermediate area of finely particulate objects such as sand and berries, which sometimes are treated as count and sometimes as mass; and there is also room for cognitive manipulation of the distinction, so thztfurniture and people involve a conceptualization!of coarsely particulate objects as unindividuated masses.12 There is a comparable lexical semantic distinction among verbs. Verbs have what is called inherent aspect or Aktionsart. This includes the state/process distinction, which does not concern us here, and also the telic/atelic distinction, that is, an action which is bounded vs. an action which is not bounded. Most simple state and.process verbs.denote events that are not bounded in time: like, read, run, carry, etc. Most examples of bounding of actions involve the existence of definite direct objects such as read 'War and Peace' which bounds the activity of reading by the end of the book. However, punctual verbs such asflash,snap, and chip, and state transition verbs like break (intr.) and die, are naturally bounded. The semantic parallel holds between count/mass on the one hand and telic/ atelic on the other: the first subclass in each pair is naturally individuated in its physical background dimension and the second is not. The difference between the two is in the physical background dimension in which the bounding occurs (see Mourelatos 1981; Langacker 1987a; Croft and Hobbs, in prep.). For objects, individuation is in space: liquids are not individuated because they don't naturally occupy a fixed region in space and they can 'fall apart' in space (i.e. they can split and occupy two disjoint regions). For events, individuation is in time: a bounded process is bounded in time, and an unbounded process is one that in theory can go on forever. Thus, we must again make recourse to the physical background dimension for objects and events. It is worth noting that with respect to the background dimension of quantity, properties are unbounded: 'tall' is open-ended in the degree of height beyond the standard point it represents (for complications, see Clark 1973). On the other hand, the anomalous category of color terms is bounded on its scale, although the boundaries are 'fuzzy'. Of course, having a property, like being a member of a class of objects, is unbounded on the time dimension (in linguistic terms, predicate adjectives and predicate nominals are atelic). The first selecting parameter applying to tokens that I will discuss is that of individuation: the existence and use of subclasses and inherent inflections to

268 A Conceptual Framework for Grammatical Categories

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

individuate otherwise unindividuated objects and events. In the case of objects, that means terms that 'convert' mass nouns into count expressions. This is the raison d'etre for unit phrases such as a glass ofwater, a cup of coffee, a sheet of paper, or a blade or grass (Leech and Svartvik 1975: 45), where certain ordinary count nouns are associated with particular mass nouns to the point that they become semi-standard individuating terms. In many cases, we find non-individuatable objects individuated by unit terms for salient properties that they possess; these are what Leech and Svartvik call 'measure words': a gallon of gas (volume), a yard of cloth (surface area) or a ton of sand (weight). This also illustrates the application of individuation to properties: measure words divide the scale of a propertyvolume, area, weight, etc.—into individuated units. In a number of languages, unit terms are used for count nouns as well; they are called classifiers or counters, e.g. Garo borarj te-gini [tree.house hollow.thing-two] 'two tree houses' (Burling 1961: 53). In the case of verbs, there are a variety of devices that individuate events. The inflectional category of perfective bounds an action in time. In addition, as we mentioned before, definite direct objects have the effect of bounding atelic processes. However, there are also special adverbial bounding expressions which by virtue of the parallel with this side-effect are usually expressed as direct objects: vague unit terms such as (He read) a bit/a while,]i and also measure terms of time, such as He read (for) an hour/a day/all night. There are also two non-standard individuating functions. One may wish to speak of a unit other than the one that the type naturally comes in, that is, one may wish to refer to a unit larger than the token (a group), or to a unit smaller than the token (a part). I will call these group and part selecting respectively. Aggregation of objects is expressed by group terms such as group, bunch, flock, herd, army, etc. This results in the creation of units at a different level than that of a naturally individuated unit, which can themselves be quantified (three divisions of British troops; two bunches of grapes). The closest counterpart to group terms for verbs is the collective, expressed in English again by an adverb, together (as opposed to separately, each or by X-selves), but in other languages it is sometimes expressed by a verbal inflection. Collective inflections of verbs refer to cooperative execution of a single action by an aggregate of agents, and these can also be quantified: They carried the piano upstairs (together) three times last year (cf. Talmy 1985 b). Again, there is no obvious counterpart for properties. The other non-standard individuation is decomposition. One may wish to select a part of an individual object or event that has internal structure to refer to rather than the whole, and so terms of decomposition are required. The decomposition of objects involves the special subclasses of part terms, such as half of the cake, or unit terms again such as a piece/slice of the cake (Leech and Svartvik 1975: 44). Three common special classes of part terms are body part

W. Croft 269

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

terms: orientation terms such as top, bottom, front, back, and right /left side {right and left also being deictic situating terms), and shape-based part terms such as edge, end, and face. The decomposition of events is at least as complex. Aspectual inflections such as the progressive and imperfective can denote a part of a bounded action. Complement-taking aspectual verbs (or, in other languages, verbal inflections) allow one to refer to specific parts of an action: inceptive or ingressive [start, begin), cessative or terminative [finish, complete, end, stop), continuative [keep on . . . ing) and inchoative [become, get). Finally, volitional actions can be decomposed in terms of the sequence of mental events that leads up to the real-world event; these are what are called 'agent-oriented' modal verbs (Bybee 1985: 166), such as intend, try, manage to, succeed in (and the negative counterpart fail). There is no decomposition of properties apparently because properties do not have internal structure. The best-known selection parameter applying to tokens is quantification. Properly individuated objects and events can be quantified. The quantification of objects involves the inflectional category of number (singular, dual, paucal, plural, and so on), quantifiers, and the syntactic category of numerals, already met with in the case of naming points on the background scale of properties. These forms are used for individuated measures of properties as well, that is, expressions like fivefeet, since measure terms are normally nouns.14 In the case of events, there are aspectual inflections that refer to singular vs. plural occurrences of an action (semelfactive vs. iterative, and possibly also habitual), as well as event-counting expressions such as English twice and three times (cf. French troisfois, Russian tri raza), event quantifiers such as always, sometimes, rarely, and multiple-occurrence adverbs such as repeatedly. A third selection function is distinguishing. This subsumes terms which are used to mark the use of the same lexical item to denote more than one entity in the same discourse, in a contrastivc fashion. For nouns, this is the function performed by the adjectives same, different, other, next, last X, and else as in someone else; and by the ordinal numerals [a second dogcame out and attacked the first one; cf. Croft 1985). This function is also found among verbs. There are two adverbs in English that appear to function to distinguish events that arc described by the same verb: again and back, as in Gerry paid Mark again/back. The paying event is distinct from a previous event also denoted by paid: with again, it was a previous event of Gerry paying Mark, and with back, it was a previous event of Mark paying Gerry. These adverbs correspond to inflections in other languages such as Fula: 'yam-t- [ask-REPETITIVE-] 'ask again (verbal extension II)'; lat-it- [kick-RETALIATIVE-] 'kick back (verbal extension IV)' (Arnott 1970: 341-3). Otherwise, in English the verbal forms for distinguishing use the nominal modifiers on the event unit term time: the last time, another time, etc. Again, there seems to be no clear counterpart in the case of properties, except in locutions

270 A Conceptual Framework for Grammatical Categories

such as Fred is the same height as Mark, in which the nominal distinguishing terms are used to apply to a nominalized form of the property (contrast the distinctly adjectival equative and comparative constructions discussed above.) However, this dimension, like others concerning adjectives and properties that I have referred to in this paper, merits more cross-linguistic examination which may reveal some parallels not evident in English. We summarize the parameters of selection in Table 8. Table 8 Objects (Space)

Events (Time)

Properties (Quantity)

Generic

articles habitual

generic,

[generic?]

unit terms

perfective, duration phrases collective

measure terms

Specific: Individuate: Unit Group Part

Quantify

Distinguish

group terms, collectives part terms, body parts, orientation terms number, numerals, quantifiers same/other, ordinals

N/A

imperfective phasal verbs, agent-oriented modals

N/A

iterative, quantifiers, cognate objects repetitive, retaliative

[sec objects]

(see objects]

6 COMPLEX RELATIONSHIPS AMONG MINOR PROPOSITIONAL ACTS The descriptions in sections 3-5 account for most of the concepts that are typically expressed in human languages by peripheral members of the major syntactic categories, minor syntactic categories and inflectional categories. A summary table of the major and minor propositional acts, and the types of grammatical categories that employ them, is given as an appendix to this paper. However, this is not the entire story, because of the semantic structure of the three major semantic classes and the background dimensions associated with them. The correlation of minor propositional acts with the three major semantic classes of objects, events and properties is only a prototypical one. For

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Selecting an entity

W. Croft 271

Deictic-Location: 1 ofeet away 10 years ago /from now Relational: 1 ofeet above thefloor 10 years after/before the quake Finally, gradable events as well as properties may be categorized, situated and measured on the quantity scale, as in She eats too much (categorizing by intensifier), She eats as much as/more than I do (situating on quantity scale), She eats five thousand calories a day (measuring quantity). The non-prototypical character of these constructions can be observed in English by the fact that they arc grammatically parasitic on the corresponding prototypical constructions: the intensifier and comparative constructions are very close to the adjectival constructions, and the measuring the quantity of the event uses a quantified NP construction. Another type of 'peripheral' construction, the let alone construction described by Fillmore et al. (1988), represents a scalar measurement of situating: He didn't make captain, let alone general. He wouldn't make captain, let alone general.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

example, events may be located in space: In the park, children were playing.15 English uses the same construction—an adpositional phrase—to indicate the situating of events, but in a different syntactic position.16 More different grammatically is the temporal situating of objects. Here a deictic strategy yields constructions such as his ex-wife, his former wife, his future wife, his wife-to-be, with the unmarked his wife referring to present 'tense'. Some languages use an inflection to indicate the temporal situating of objects (or more precisely, . descriptions of objects), e.g. Carib paru.ru e:p'i-mbo [banana stem-SUFF] 'former banana stem [now cut down and stripped of its leaves]' (Hoff 1968: 223). These examples imply that both objects and events, that is, argument NPs and predicates, are situated in both space and time as well as mental space (see Fauconnier 1985 and Croft, in prep., for evidence supporting this position). In addition, the background dimensions of space and time can be conceptualized as entities in their own right, which means they can be categorized, selected and even situated. We noted in section 3 that words categorizing parts of space and time exist (sky, mountain, ocean, hour,'year, etc.). In addition, space and time can be selected, which means measured in this case, since they are scalar properties. (In fact, the most prototypical adjectives except for color terms refer to size in space and time—big/little, tall/short, old/new, fast/slow, etc.) The measurement of space and time is found in situating expressions:

2J2 A Conceptual Framework for Grammatical Categories

7 CONCLUSION In this paper, I have proposed a coherent conceptual framework for organizing rhe semantics of virtually the full range of grammatical morphemes of natural language. This is, of course, an ambitious proposal, and there are certainly a number of grammatical morphemes or constructions with unusual semantic descriptions that do not readily fit into the framework. However, I believe that most if not all of these will fit fairly easily into the six prepositional acts described above, summarized in Table 9. The reason why we may express hope in developing an adequate conceptual framework is because the framework is not intended to characterize the full richness of experience—certainly an impossible task—but rather the way that experience is employed in the construction of an utterance. The classification of morphemes presented here is by prepositional acts, that is, the processes by which human beings organize information for expression. These prepositional acts are more likely to be finite and small in number. In fact, the prepositional acts represent a fairly 'late' stage in the cognitive process of structuring information for expression. The prepositional acts presuppose that the speaker has already taken a perspective on the scene to be described (Fillmore 1977, Talmy 1988) and that the speaker's attention has been

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The let alone construction situates one event relative to another in a mental space of evaluation, but implies a scale in which the first conjunct (making captain in this case) is lower than the second conjunct (making general).17 These examples suggest that not only do both time and space apply to both objects and events, but that scalar quantity should apply to events as well. More generally, the prepositional acts of categorizing, situating and selecting (measuring) apply to each major semantic class with respect to any background dimension that is applicable—time, space, mental space and quantity—although certain combinations of background dimension and semantic class are prototypical in the type of grammatical construction that is employed. From a conceptual semantic point of view, however, the most important fact is that each concept that is employed in a predicating, referring or modifying prepositional act must be categorized, selected and situated, either explicitly through the types of grammatical constructions described here or implicitly through context. For example, in the sentence The water evaporated, although no morpheme situates the object or event in space, and no morpheme specifies how the relevant water was selected, the speaker has already implicitly situated the object and event in the context of the utterance, and the definite article in The water refers back to a previous occurrence of water in which the selection of the relevant amount was explicitly performed.

W. Croft 273

Table 9 Major propositional acts:

Referring Predicating Modifying Minor propositional acts:

Selecting Generic (type) Specific (token): Individuating (group, unit or part) Quantifying (including measuring on a scale) Distinguishing

drawn to certain elements of the scene, namely those that will be expressed (Talmy 1988's 'distribution of attention'). The propositional acts themselves represent the structuring of the elements of the scene being attended to. Unfortunately, at this stage of cognitive linguistic research, the 'earlier' processes of adopting a perspective on a scene and focusing attention on elements of the scene to be expressed have been little explored18—perhaps because the principles underlying them probably have more to do with cognitive psychology than they do with linguistics. We may nevertheless close by discussing the reasonableness of assuming that the propositional acts described here suffice in describing the construction of propositions. The three major propositional acts appear to represent a basic division of communicative function of the concepts attended to by the speaker (for detailed discussion, sec Croft 1991, section 3.2). Referring involves the opening (or reopening) of a cognitive file, to use DuBois's term (1987: 817), for some entity that will continue to be referred to. Predicating on the other hand is instead the registering of something to be sequentially scanned by the mind of the hearer, to use Langacker's phrase (1987a: 72). Modifying appears to be an enrichment of the image evoked by the opened cognitive file for a referred-to entity (Wierzbicka 1986: 374). Our major concern in this paper has been the minor propositional acts of categorizing, situating and selecting. They can best be understood by recalling the 'earlier' cognitive processing that the speaker must undertake. In particular,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Categorizing Situating (in space, time and mental space) Deictic Relational, with respect to: Background frame of reference (deictic-location or absolute) Other entity

274 A Conceptual Framework for Grammatical Categories

Acknowledgement I would like to thank Dwight Bolinger, Joseph Greenberg, Jerry Hobbs, Suzanne Kemmer, Roland Langacker, Elizabeth Traugott, and Tom Wasow for comments on earlier versions of this paper. Portions of the research described herein were funded by the Defense Advanced Projects Research Agency and the Kellogg Foundation. WILLIAM CROFT Program in Linguistics 1076 Frieze Building University ofMichigan Ann Arbor, MI 481 og-i2&$ USA

NOTES 1 Certain terms have to be defined in order to clarify the analysis that is embodied in this chart. I will use entity as a cover term for a member of any of the semantic classes of objects, qualities and events or actions. Obviously, 1 am taking the position that events and qualities have an ontological status comparable to objects: they exist as much as objects do, though by their nature they are somewhat differ-

ent semantically. I will use the term denote to indicate the relationship between the word and what it designates, regardless of 'syntactic' function; reference on the other hand is restricted to the 'syntactic' function. Thus, eats in Fred eats tripe denotes the activity of eating, although the sentence refers to Fred and tripe, but predicates the activity. Prototypical objects are defined as stative,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

the speaker distributes his/her attention over the scene, ultimately settling on those elements that he/she will express in the proposition. The speaker's attention is determined by his/her specific purpose in mind, and also to a sort of natural human attunement to particular configurations in the scene (as noted by the Gestalt psychologists). The hearer must then reconstruct the scene intended by the speaker; so the speaker must bring the hearer's attention to the same elements of the scene that the speaker's attention was directed to. That is the function of the minor propositional acts. Categorization provides a qualitative image of the element categorized. Selecting the entity gives a more specific identity to the element. Situating the entity allow the hearer to anchor the element in a background—space, time, mental space—so that the hearer can better fit the element, and the scene in general, into his/her existing experience. It seems intuitively obvious that categorizing, situating and selecting are the best means to accomplish the goal of drawing the hearer's attention to the 'right' elements of the scene in order to reconstruct it. If any other minor propositional acts exist, then they presumably perform the same function.

W. Croft 275

2 I switch to the gerund forms of these terms here to emphasize the fact that these are propositional acts. 3 This still leaves one large and important class of morphosyntactic devices found in natural language grammars unaccounted for the so-called 'discourse particles' such as well, so, anyway, connectives like and,

but, thus and many others, as well as comparable devices in other languages. I will not address this aspect of grammar here; nor will I discuss other inter-clausal and inter-sentential phenomena such as anaphora, although they undoubtedly play a major role in the grammar of natural languages. It should be observed that the goal of this paper is to account for only the grammar of the clause, and possibly the sentence as a certain type of multiclausal unit. Presumably, the conceptual-pragmatic approach illustrated here for syntactic categories, and illustrated below for non-prototypical lexical and grammatical elements, can also be used to analyse the structure of larger units of discourse. 4 I am using the term 'categorizing' here differently than Wierzbicka (1986) uses it. Wierzbicka contrasts the term 'categoriz-

ing', what nouns do, with the term 'describing', what adjectives do. Bolinger (1980a), in a similar analysis of the difference between nouns and adjectives, uses the terms 'classifying' and 'describing'. I follow Bolinger's terminology here because of the general use of the term 'categorization' in psychology for the . categorization of objects, properties and events. 5 In many languages, spatial adpositions may also be used to relate an entity to a point in the background dimension, e.g. in here, in San Rafael. There is some variation across languages as to when an adposition is necessary. 6 One exception to this are the color terms. Color terms are in the core adjective set but except for the simplest two- or threeterm systems (black/white/red, which are perhaps better translated by the relative extent terms dark/light/bright), they are not relative extent terms and their focal points are universal, not context-based (Berlin and Kay 1969). 7 One also finds proper names for social/ tribal units and the languages that they speak, e.g. the English, English [language] (I am grateful to Tom Wasow for pointing this out to me). A social/tribal unit can function as a mental space, namely in the social/cultural 'world' that it defines, analogous to Fauconnier's domain spaces (Fauconnier 1985: 31). This explanation would predict that proper names for social/tribal units (and for languages) are used chiefly or exclusively as the background dimension for a situating propositional act. 8 This would be committing the exclusionary fallacy in linguistic explanation (Langacker 1987b: 28). 9 English tends to express as subject the person who functions as the reference point for the mental space; other languages (e.g. those in the Caucasus and South Asia) express the person in the dative instead. 10 Conditional subtypes refer to the status of

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

zero valency (that is, not inherently relational; see Croft 1991, Section 2.3), object-level predicates (that is, roughly, permanent or persisting; see Carlson 1979, Croft 1991). Prototypical events or actions are processual, non-zero valency, stage-level predicates (that is, roughly, transitory) predicates. Prototypical qualities are stative, unary valency, transitory, gradable predicates (neither core objects nor core processes are gradable). These definitions will serve to make more precise the semantic definition of'object', 'event' and 'quality', all vague terms in common use. However, these definitions are fairly narrow, and so exclude many items that are considered 'lexical' rather than 'grammatical', such as relational nouns (e.g. father) or stative verbs (e.g. love).

276 A Conceptual Framework for Grammatical Categories

11

12

13

15

16

REFERENCES Allan, Keith (1977), 'Classifiers', Language, 53,

and Los Angeles, University of California Press. Anderson, John M. (1971), The Grammar of Bolinger, Dwight (1967), 'Adjectives: attribuCase: Towards a Localistic Theory, tion and predication', Lingua, 18, 1-34. Cambridge, Cambridge University Press. (1980a), Syntactic Diffusion and the InAnderson, Stephen R. (1985), 'Inflectional definite Article, Bloomington, Indiana morphology', in Shopen (1985), 150-201. University Linguistics Club. Arnott, D. W. (1970), The Nominal and Verbal — (1980b), Language, the Loaded Weapon, Systems ofFula, Oxford, Oxford University London, Longman. Press. Burling, Robbins (1961), The Can Language, Berlin, Brent (1978), 'Ethnobiological classiPoona, Linguistic Society of India. fication', in Rosch & Lloyd (1978), 9-26. Bybee, Joan (1985), Morphology, Amsterdam, — & Kay, Paul (1969), Basic Color Terms: John Benjamins. Their Universality and Evolution, Berkeley Carlson, Greg (1979), 'Generics and atem285-311.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

14

playing is the same as that in The children in the antecedent with respect to the known the park were playing. Since the locative in world. the latter sentence is situating only the Occasionally one finds clitic pronouns in children rather than the whole scene, it second position, as in Serbo-Croatian. I has a slightly different pragmatic funchave no explanation for this anomaly, tion, namely modification. except to observe that clitic pronouns are often associated with the auxiliary. 17 Another construction described by Fillmore et al. (1988), the 'comparative conSee Wierzbicka (1985a, 1985b) for a ditional' construction exemplified by The detailed analysis of the 'natural' indimore he drinks, the less he works, is handled viduation patterns of various entities, and more straightforwardly in the conceptual Langacker (1987a) for additional disframework described in sections 3-5. In cussion. the comparative conditional, a class of A similar effect is achieved by the havea V events, defined by a scale, is situated in construction in English: have a drink, have mental space with respect to another class a look (Wierzbicka 1982). of events; so the construction represents a This category includes of course vague combination of an act of selecting with an numerals such as a Jew, several, and many. act of situating in mental space. Also, it is worth noting that English has developed terms for quantifying mass 18 For example, Talmy has not published nouns without their being properly detailed studies of his imaging systems of individuated, such as much, a lot of, and so 'perspective' and 'distribution of attention'; his published work refers chiefly to on (Leech and Svartvik 1975: 51). 'schematization' (Talmy 1988), which In fact, these locatives, at least when precorresponds roughly to the semantic posed, perform the 'scene-setting' funcnotions described here, and 'force tion remarked on by various discourse dynamics' (Talmy 1985 a), a generalized analysts (e.g. Chafe 1976). concept of causality, which corresponds Of course, the location of an event is to the analysis of relational morphosyntax parasitic on the location of the particimentioned in section 2. pants, so the 'objective' situation described by In the park, the children were

W. Croft 277

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

poral when, Linguistics and Philosophy, 3,DuBois, John A. (1987), 'The discourse basis 49-98. of ergativity', Language, 63, 805-55. Chafe, Wallace L. (1976), 'Givenness, contras- Fauconnier, Gilles (1985), Mental Spaces, tiveness, definiteness, subjects, topics, and Cambridge, Mass, MIT Press. point of view', in Charles, Li (ed.), Subject Fillmore, Charles (1977), 'The case for case and Topic, New York, Academic Press, 25reopened', in Peter Cole & Jerrold M. Sadock, (eds), Grammatical Relations (Syntax 55and Semantics 7), New York, Academic Clark, Herbert H. (1973), 'Space, time, Press, 59-82. semantics and the child', in T. E. Moore — (ed.), Cognitive Development and the Acquisi- , Kay, Paul and Kay O'Connor, Mary (1988), 'Regularity and idiomaticity in tion of Language, New York, Academic grammatical constructions: the case of let Press, 28-64. alone', Language, 64, 501-38. Comrie, Bernard (1985), 'Causative verb formation and other verb-deriving Friedrich, Paul (1970), 'Shape in grammar', Language, 46, 379-407morphology', in Shopen (1985), 309-48. — & Thompson, Sandra (1985), 'Lexical Genetti, Carol (1986), 'The development of subordinators from postpositions in Bodic nominalization', in Shopen (1985), 349-98. languages', Proceedings ofthe Twelfth Annual Croft, William (1984), 'Semantic and pragMeeting of the Berkeley Linguistics Society, matic correlates to syntactic categories,' Parasession on Lexical Semantics, Twentieth Vassiliki Nikiforidou et al. (eds), 387Regional Meeting of the Chicago Linguistic 400. Society,' D Teston etal (eds.), 53-71 Greenberg, Joseph H. (1966), Language Universals, with Special Reference to Feature — (1985), 'Determiners and specification', in Jerry Hobbs etal., Commonsense Summer: Hierarchies Qanua Linguarum, Series Minor, 59), The Hague, Mouton. H«a/Rfporr,CSLIReportNo.CSLI-85-35, Center for the Study of Language and — (1974), 'The relation of frequency to semantic feature in a case language Information, Stanford University. (Russian)', Working Papers in Language — (1990), Typology and Universab, Universab, 16, 21-47. Cambridge, Cambridge University Press. Herskovits, Annette (1986), Language and — (1991)' Syntactic Categories and Grammatical Relations: The Cognitive Organization of Spatial Cognition, Cambridge, Cambridge University Press. Information, Chicago, University of Chicago Press. Hoff, B. J. (1968), The Carib Language, The Hague, Martinus Nijhoff. — (in press), 'Possible verbs and the structure of events', in S. L. Tsohatzidis (ed.), Jackendoff, Ray (1983), Semantics and CogniMeanings and Prototypes: Studies on Linguistiction, Cambridge, Mass., MIT Press. Categorization, London and New York, Langacker, Ronald N. (1982), 'Space Routledge. grammar, analysabiliry and the English passive', Language, 58, 22-80. — (in preparation), 'The conceptual unities of the clause (with apologies to Aristotle)'. — (1987a), 'Nouns and verbs', Language, 63, — & Hobbs, Jerry (in preparation), 'Aspect, 53-94countability and granularity'. — (1987b), Foundations of Cognitive Grammar, Davison, Alison (1980), 'Any as universal or Vol. I: Theoretical Prerequisites, Stanford, existential?' in Johan van der Auwera (ed.), Stanford University Press. The Semantics of Determiners, London, Leech, Geoffrey & Svartvik, Jan (1975), A Croom Helm, 11-40. Communicative Grammar of English, Dixon, R. M. W. (1977), 'Where have all the London, Longman. adjectives gone?', Studies in Language, I, McCawley, James D. (1981), Everything that 19-80. Linguists Have Always Wanted to Know about

278 A Conceptual Framework for Grammatical Categories Logic (but were ashamed to ask), Chicago, nition: a synopsis', Proceedings ofTINLAP2, David Waltz (ed.). University of Chicago Press. Morrow, Daniel (1986), 'Grammatical — (1985 a), 'Force dynamics in language and thought', Papers from the Parasession on morphemes and conceptual structure in discourse processing', Cognitive Science, 10, Causatives and Agentivity, Twenty-First Regional Meeting of the Chicago Linguistic 423-55Society, William H. Eilfort etal. (eds), 293Mourelatos, Alexander P. D. (1981), 'Events,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

337processes and states', in Philip J. Tedeschi & Annie Zaenen (eds), Tense and Aspect — (1985 b), 'Lexicalization patterns: semantic structure in lexical forms', in Shopen (Syntax and Semantics, 14), New York, (1985), 57-149Academic Press, 191-212. Nunberg, Geoffrey (1979), 'The nonunique- — (1988), 'The relation of grammar to cognition', in Brygida Rudzka-Ostyn (ed.), ness of semantic solutions: polysemy', Topics in Cognitive Linguistics, Amsterdam, Linguistics and Philosophy, 3, 143-84. John Benjamins, 165-205. Palmer, F. R. (1986), Mood and Modality, Traugott, Elizabeth (1978), 'On the expresCambridge, Cambridge University Press. sion of spatio-temporal relations in Rosch, Eleanor (1978), 'Principles of categorlanguage', in Joseph H. Greenberg, Charles ization', in Rosch & Lloyd (1978), 27-48. A. Ferguson & Edith A. Moravcsik (eds), — & Lloyd, Barbara (eds) (1978), Cognition Universals of Human Language, Vol. 3: Word and Categorization, Hillsdale, NJ, Lawrence Structure, Stanford, Stanford University Erlbaum Associates. Press, 369-400. Searle, John (1969), Speech Acts, Oxford, Wierzbicka, Anna (1982), 'Why can you have Oxford University Press. a drink but not 'havean eat?', Language, 58, — (1976), 'The classification of illocutionary acts', Language in Society, 5, 1-24. 753-99Shopen, Timothy (ed.) (1985), Language — (1985a), Lexicography and Conceptual Typology and Syntactic Description, Vol. Ill: Analysis, Ann Arbor, Karoma. Grammatical Categories and the Lexicon,— (1985b), '"Oats" and "wheat": the fallacy Cambridge, Cambridge University Press. of arbitrariness', in John Haiman (ed.), Stassen, Leon (1985), Comparison and Universal Iconicity in Syntax, Amsterdam, Benjamins, 311-42. Grammar, Oxford, Basil Blackwell. Steele, Susan (1975), 'On some factors that — (1986), 'What's in a noun? (or: how do nouns differ in meaning from adjectives?)', affect and effect word order', in Charles Li (ed.), Word Order and Word Order Change, Studies in Language, 10, 353-89. Austin, University of Texas Press, 199-268. Witherspoon, Gary (1977), Language and Art in the Navajo Universe, Ann Arbor, UniverTalmy, Leonard (1972), 'Semantic structures sity of Michigan Press. in English and Atsugewi', Ph.D. dissertaZepeda, Ofelia (1983), A Papago Grammar, tion, University of California, Berkeley. Tucson, University of Arizona Press. — (1978), 'The relation of grammar to cog-

W. Croft 279

Appendix Semantic morpheme types by proposirional act Entity Type: Background Dimension:

Events

Properties

Space

Time, Causality

Quantity

nominalizations, infinitives, gerunds, complements (verb) participles, relative clauses

abstract deadjectival nouns predicate adjectives (adjective)

verb roots causal-aspectual verb classification

adjective roots adjective roots (?), intensifies/ downtoners (?)

Major proposirional acts: Referring: (noun)

Predicating: Modifying:

predicate nominals genitives, denominal adjectives, PP modifiers

Minor propositiona 1 acts: Categorizing an entity Lexical

Grammatical

noun roots gender, classifiers (numeral, verbal)

Situating an entity in a physical dimension:

Deictic

demonstratives

tense

core adjectives, diminutives intensifies?

Relational Background Deictic-Loc. Absolute Other entity

locative deictics names of places spatial adpositions

temporal deictics names of times temporal/causal connectives and adpositions

[demonstratives] numerals comparatives, equatives

articles, definiteness

modals, attitude markers, speech acts, sentential adverbs

N/A?

syncategorematic adjectives, fictional context nouns possessive pronouns names of humans N/A?

world-creating predicates

N/A?

Generic

articles

generic, habitual

[generic?]

Specific Individuate Unit

unit terms

perfective, duration phrases

measure terms

Situating an entity in a mental space:

Deictic

Relational Background

Deictic-Loc. Absolute Other entity

personal pronouns names of humans conditionals

Selecting an entity:

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Objects

280 A Conceptual Framework for Grammatical Categories

Appendix (cont.) Entity Type: Background Dimension: Group Part

Quantify

Events

Properties

Space

Time, Causality

Quantity

group terms, collectives part terms, body parts, orientation terms number, cardinal numerals, quantifiers same/other, ordinals

collective

N/A

N/A imperfective, phrasal verbs, agent-oriented modals [see objects] iterative, quantifiers, cognate objects [see objects] repetitive, retaliative

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Distinguish

Objects

Journal of Semantics 7: 281-300

© N.I.S. Foundation (1990)

Forward References in Natural Language* KEES VAN DEEMTER Institutefor Perception Research (IPO), Eindhoven, The Netherlands

Abstract

1 INTRODUCTION Recent approaches to anaphora such as Kamp's Discourse Representation Theory (DRT) (Kamp 1981) and Heim's file-change semantics (Heim 1982) adhere to the procedural principle of familiarity. According to this principle, expressions whose denotation is dependent on other material (i.e. anaphoric expressions) may only depend on previously processed, and therefore 'familiar', expressions. A particularly interesting variety of familiarity is obtained if it is assumed that processing operates in the same 'direction' as speech, that is—in the western tradition of writing—if it operates from left to right. Henceforth, this variety of the familiarity approach will be called the left-to-right, or briefly the 1-t-r approach.' It is sometimes thought that theories such as Kamp's and Heim's are instances of the 1-t-r approach, and this appearance may have added considerably to their intuitive appeal. However, this appearance is deceptive, as we will show.2 At least four parallel levels of text processing can be distinguished: speech recognition, syntactic analysis, the algorithm that constructs Discourse * I thank Johan van Benthem.Jeroen Groenendijk and Martin Stokhof for helpful discussions on the subject of this paper. Robbert-Jan Beun and Sieb Nooteboom provided useful comments on an earlier draft.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

This paper deals with forward references (also called kataphoric references) in natural language. In order to calculate truth conditions for sentences that involve kataphoric references, an extension of Discourse Representation Theory, PATIENT DRT, is proposed, inspired by socalled backpatching techniques for the parsing of programming languages. The main idea is that a kataphoric element introduces an incomplete discourse entity, to be completed by subsequent material under certain conditions. This approach is applicable to pronominal as well as complex Noun Phrases, and has no special difficulties with crossing co-references. The main virtue of this approach is that it allows parsing of kataphors from left to right, which makes it suitable for on-line language processing by computer and plausible as an element of a theory of human language processing as well. However, the approach suggests that a left-to-right treatment of kataphoric constructions is hard to reconcile with the requirements of compositionality.

282 Forward References in Natural Language

Representation Structures (DRSs), and finally truth conditional interpretation: Interpretation: > > DRS construction: >—>—>—> > Syntax analysis: ->->—>—>—>->-> TEXT:

blabla

>

> >

> > > blabla

(Arrows depict steps in the processing of the TEXT: recognition may result in words, syntax analysis produces constituents, DRS construction results in partial DRSs, and interpretation results in truth conditions.) Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

In its unrestricted form, the 1-t-r hypothesis would require that all four processing stages proceed from left to right. Applied to recognition, for instance, this requirement is intuitively quite plausible. For interpretation, it would amount to the highly desirable property of incremental (i.e. on-line) interpretation. Note that there must be several non-trivial dependencies between the four processes. For instance, it is widely acknowledged that speech recognition needs 'higher' linguistic cues (e.g. Lea 1980 on HEARSAY). Consequently, the higher processes must have the same direction as recognition and cannot be allowed to lag too far behind it. In short: the 1-t-r principle is a highly attractive hypothesis for language processing in general. For theories of anaphora, the most important stage of processing is DRS construction, where discourse entities are introduced and subsequently picked up by anaphors. Consequently, the processing direction of DRS construction is our central concern. We will largely leave aside the feasibility of 1-t-r processing in the other areas of text processing, and the same holds for questions of synchronisation between the four processes.3 Note that especially the explanatory value for familiarity-based theories of anaphora would be greatly enhanced by 1-t-r DRS construction, since it would explain why certain material can be considered as 'familiar' at a certain moment. Also, it would help to legitimise the often claimed role of recency in anaphora resolution (Chafe 1976, Sanford and Garrod 1981, more recently e.g. Pinkal 1986). The problem addressed in this paper is, how can the hypothesis of 1-t-r DRS construction be reconciled with the empirical phenomenon of kataphoric expressions (cf. Buehler 1934)—anaphoric expressions whose denotation depends on material to their right? For an 1-t-r analysis of kataphoric expressions would only be possible if an antecedent can already be familiar before it has been processed, and it is hard to see how this is possible. After a brief exposition of the linguistic data (section 2), we shall discuss some previous treatments of kataphora (section 3). Then (section 4) we outline a modified, patient version of DRT which deals with kataphoric reference, after

K. vanDeemter 283

which (section 5) we will apply Patient DRT to some problematic sorts of kataphors. In the concluding section, we try to answer some hard questions prompted by the solutions proposed in the body of the paper. Throughout, the word 'anaphora' will—contrary to Greek grammar—be used to denote both forward and backward cases of anaphora, that is both kataphoric and strictly anaphoric reference. Similarly, an 'antecedent' may either precede or follow the anaphor. No ambiguity will arise from this usage, I hope.

2 THE P H E N O M E N O N OF KATAPHORA

(1) Ever since her childhood, Dorit has been extremely lazy. (2) Ever since Dorit's childhood, she has been extremely lazy.4 Now one might argue that 'her' must be anaphoric rather than kataphoric: that it can only refer to Dorit if she was introduced earlier. But even if this is true—a supposition that fits in neatly with the tendency for pronouns to refer to the focus of a discourse (e.g. Sidner 1983)—the felicity of the use of the pronoun 'her' depends also on future material. Compare: (3) Mary, Dorit and Bill are a strange lot. She is weird. The others. . . (4) Mary, Dorit and Bill are a strange lot. Ever since her childhood, Dorit has been extremely lazy. The others. . . 'Her' in (4) can refer to Dorit, while 'she' in (3) cannot. The explanation must be that 'her' is linked to the second, rather than the first, occurrence of the proper name 'Dorit' in (4). Consequently, 'her' must be a kataphoric pronoun.3 Kataphoric constructions can be considered more involved than the examples provided so far. One complicated kind of kataphora is 'mutual anaphora', where two Noun Phrases (NPs) mutually depend on each other (or on an NP embedding the other) for their denotation (cf. Bach 1970, Karttunen 1971):

(5) [A woman who works in his, department], was interviewed by [the manager who hired herj2 (subscripts indicate co-reference). Further kataphors are not restricted to pronouns. For instance, if we label the relation between 'the parents' and 'these puppies' in: (6) These puppies were born this spring. The parents took good care of them

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Kataphoric reference has sometimes been depicted as a rather marginal phenomenon. In some cases, however, kataphoric reference (1) is decidedly more felicitous than a 'backward' anaphoric analogue (2):

284 Forward References in Natural Language

as relational anaphora, then there is also such a thing as relational kataphora. For instance, in... (7) Whenever the parents sleep, these puppies do not eat 'the parents' can be described as kataphoric to 'these puppies'. (For an analysis along the lines of DRT, see van Deemter 1989.) Cases of extrasentential kataphora are reported as well, but their grammaticality seems debatable. A case in point is (8) First he lost his wallet. Then his car got stolen. Fred was having a bad day

(9) He was an old man who fished alone in the Gulf Stream.. . (Hemingway: The Old Man and the Sea) Note that this is the opening sentence of a book, so a backward anaphoric reading of the pronoun 'he' is excluded. In section 5 we will return to the topic of kataphors that cross coordination boundaries, as well as to the other phenomena just described. While dealing with kataphoric reference, it will not suffice to indicate possible antecedent-anaphor (kataphor) pairs, since the possibility of an anaphoric relationship will also depend upon analysis. For instance, in (10) Everyfarmer who admires her courts a widow a kataphoric relationship is only possible if 'a widow' has wide scope over the universal quantifier in 'every farmer', as Kamp observed (Kamp 1981). Scope phenomena motivate much of the complexity of the rules in the main section of this paper.

3 KATAPHORA IN DRT AND RELATED APPROACHES In the introduction the 1-t-r approach to anaphora was advertised. This approach may be argued to consist of the following three principles: (Pi) Parsing operates from left to right on surface structure; (P2) Discourse entities are introduced in the context during parsing; (P3) Pronouns pick up existing discourse entities from the context. Together, these three principles rule out kataphora. Consequently, actual proposals for the treatment of kataphora have departed from one or more of

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

in which, according to Asher and Wada (1989), 'he' and 'his' can be kataphoric to 'Fred', although one may doubt whether this can amount to a completely natural reading. Some literary texts provide interesting examples:

K. van Deemter 285

(11) (John walks.) He talks. He keepsforgetting the time.

Then, obviously, the processor should not be precluded from making any sense of this story—as Kamp's theory would have it. This observation suggests that the pronoun 'he' should introduce a Reference Marker (RM) of its own. Given the interpretation rules of DRT, this means that (11) is true if and only if there is at least one person who talks and keeps forgetting the time. Kataphoric and incompletely perceived discourse should be interpreted along similar lines. Therefore, we propose to maintain P i and P2 but to abandon P3. Recently, some steps towards a solution along these lines were taken in Asher and Wada (1989). There, pronouns introduce their own RMs. Resolution of the pronouns is postponed until there are no more reducible conditions left. Several rules constrain resolution, the most central one being that the antecedent must either be accessible for the anaphor, or be a definite NP, or z wide scope

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

them. Thus, in Heim (1982), P i is amended to apply only to a level of logical form, to be obtained from surface structure by a preprocessing stage that puts all the definite NPs in front. Unfortunately, this means that parsing works in two 'passes'. As a result, on-line interpretation of a sentence is impossible, for the second pass cannot get started before the first pass has seen the very last NP of the sentence. Further, Heim's approach fails on mutual anaphora. For given that neither of two NPs in a mutual anaphora construction can be interpreted without the other, no level of analysis can have them in the 'right' order. Another departure from P i can be found in Kamp (1981). Here, processing order is highly indeterminisric. Although the numbering of processing steps in the boxes which depict Kamp's Discourse Representations (DRs) may suggest determinism, the numbering constitutes only one of several possible scenarios yielded by the processing rules. In particular, whenever a DR is split into two subordinate boxes—say i, and b2—to represent a universally quantified sentence or a conditional (Kamp 1981), bx and b2 can be processed concurrently, except when an NP from one box has to be used as antecedent for a pronoun in the other. In the idiom of parallel programming, the two processes entertain a producer/consumer relation (see e.g. Ben-Ari 1982): when a pronoun in b, cannot be resolved, control is shifted from b, to b2; as soon as a suitable antecedent has been found there, control is allowed to return to b,. It is due to these departures from P i that kataphoric constructions can be treated appropriately in Kamp (1981).6 From our own point of view, however, P i , being the heart of the 1-t-r hypothesis, deserves to be upheld, of course. On the other hand, principle P2 is too central an assumption of DRT to give up. P3, on the contrary, must be given up anyway in order to account for incompletely perceived discourse: suppose someone overhears (11), hearing everything of it except the first sentence:

286 Forward References in Natural Language

indefinite (as in (10)). However, note that the postponement of resolution is at odds witii the requirements of on-line interpretation. In particular, it rules out that interpretation starts before' DRS construction is finished. Moreover, there are empirical difficulties as well; not all the possible scopes of descriptions are covered. For instance, there is no way, in Asher and Wada (1989)—nor is there in Kamp (1981), see our note 6—to interpret sentences such as (10): (10) Everyfarmer who admired her courts a widow

where the antecedent of a kataphoric pronoun would be introduced in the 'wrong' position.7 For, after the resolution decision y = z, the following box would result:

. X

•y

farmer (x) admire y=z

But this box can only be true relative to an embedding function £ if, for each_g' which differs from g at most in its values for x and y, if g'(x) is a farmer who admires g'{z), then^'(z) is a widow who is courted hy g\x). In other words, if someone is a farmer who admires g\z), then^'(z) is a widow who is courted by him—which is not a viable reading for the sentence. There seems to be no simple way out of this inconvenient situation.8 What is needed is the possibility to introduce the condition widow (z) in b,, rather than b 2, due to its kataphoric link to the RM y which is introduced by the pronoun 'her'. But then this kataphoric link must be known by the rime the widow NP is processed. Consequently, resolution cannot be postponed until the rest of DRS construction is finished. Now we come to our own proposal, in which these lessons are taken to heart. We will not discuss structural constraints of the type proposed in the literature, although some of these are obviously relevant to kataphoric reference, since we have little to add to them. We will think of structural constraints as additional constraints on accessibility. Thus, although DRT on its own would allow coreference in

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

bi

K. van Deemter 287

(12) SHE thinks MARY is pretty the addition of Reinhart's Non Co-reference Rule to DRT forbids the two NPs to relate to the same RM, since 'she' c-commands 'Mary'. We will not choose between different versions of these constraints (e.g. Reinhart 1976, Reinhart 1983, Bosch 1983), nor will we discuss strategies to integrate them with DRT (cf. Asher and Wada 1989 for an interesting proposal). Instead, we will assume a suitably enriched version of DRT and concentrate on the specific mechanisms needed to account for kataphoric reference in an 1-t-r based approach. 4 A TREATMENT OF KATAPHORA IN PATIENT DRT

(13) Whenever she was off duty, Mary spent her time in the swimming pool 'she' may introduce a reference marker x in a box b with property female (x), also written as she (x). Upon encountering the proper name, the condition x = Mary is added to b. In computer science, a similar procedure for dealing with forward references is known as backpatching (Aho, Sethi and Ullman 1986). Backpatching is a way to deal with forward references in programming languages which prevents an entire program from having to be scanned more than once during parsing: a forward reference generates an incomplete translation that is completed later. For instance, forward references in GOTO statements are translated into machine code by first generating a 'skeletal instruction' in which the target address of the GOTO statement is left open until the target instruction is reached, so that its address is known. Thus, no second 'pass' of the program text is necessary. Our treatment of forward references in natural language will mirror this procedure. Assume that an NP arises in a condition occurring in a DR m, that is part of a DRS K. Let u* be M, with the additional information that u is incomplete. The notation [a: = w] stands for (f>, with a everywhere replaced by u. Conm is the set of conditions in the box m. Um is the set of RMs in m. V is the total set of variables available as RMs. Uk Q V is the set of variables used in K. Finally, J O (m) denotes the set of boxes that are accessible to m. The central rules for processing kataphoric NPs in m are the following two principles: PATIENCE PRINCIPLE: A kataphoric pronoun introduces a new discourse entity into m (with appropriate number and gender features) that is marked as incomplete. (Formally: a pronoun a in can be processed as follows: add

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

We propose to allow that pronouns can introduce RMs. Although such RMs will not be complete as they stand, the idea is to be patient and to allow that the process is completed when the antecedent is reached. For instance, in order to arrive at one of the readings for the sentence

288 Forward References in Natural Language

In the sequel, we shall assume that these rules belong to patient DRT, alongside Kamp's DRS construction rules. This time, of course, processing operates from left to right. In particular, by requiring that in a DR of the form m' -* m, DRS construction processes m before m, DRS construction is forced to proceed deterministically from left to right. Before we actually illustrate the operation of the principles of Patience and Completion, we will add provisions for 'deviant' scopes, not only to move proper names into their required wide scope position, but also to allow the scope of other NPs to diverge from their place in surface order. OPTIONAL RISE PRINCIPLE: Pronouns, definites and indefinites can introduce an RM in any existing DR higher up in K's accessibility hierarchy. (Formally: An indefinite or pronoun a can introduce their RM x € V — UK in any member m' of K^ (m). If a is a pronoun, then an incompleteness mark is added. Further, add a = u to Con,,,' and add [a: = u] to Con,,,. This 'Quantifier Raising' principle validates 'wider than surface' scopes for all except quantifying NPs. To illustrate the rules so far, consider (14) Whenever she was offduty, the president spent her time in the swimming pool

where there is an ambiguity in the relative scopes of'the president' and 'whenever'. The rules lead to the representations shown in Figure 1. DRS-i is obtained via Patience and Completion only. DRS-2 results from Optional Rise. Our earlier sentence (10) would be analysed on the same pattern as DRS-2. Yet, it seems that there is a fundamental problem with the treatment sketched: consider a kataphoric pronoun occurring outside the scope of a conditional. This pronoun will only have universal meaning (and thus belong at the left-hand side of a split box) if it will later be completed by a quantifying NP; but the future occurrence of a quantifying NP cannot be anticipated. This seems to imply that introduction of the RM for the kataphoric pronoun must

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

to Um a suitable element u# from V — UK, and add a = u and [a: = w] to Conra.) COMPLETION PRINCIPLE: A non-pronominal NP /? can pick up a reference marker w* from a box m', introduced by an earlier occurring kataphor, the incompleteness mark is deleted; the conditions normally associated with /? are added torn'. Other conditions on u are added as normal. (Formally: When processing ^3, choose a 'suitable' member u# from an element m' from K > (m), add (f>[j8: = H] to Con,,,, and: —if /? is an indefinite of the form 'a rj' or a quantifying NP of the form 'every r)\ then add T^(M) to Con,,,'. —if yS is a proper name, then add u = /? to Conm'. —if /? is definite description, then add the content of the description to Conm'.)

K. van Deemter 289

DRS—1:

swimming (x,t)

.x

she(x) the president(x,f')

.t

offduty^,?)

Explanation: The variable I ranges over time intervals. DRS-1 deals with all those .v and I for which x is female and president at (. DRS-2, where /' is cither utterance time or reference time, is verified if there is a female president x at 1', such that, for all intervals 1 during which x is off duty, x is swimming at (.

Figure 1

be postponed—which would be at odds with the 1-t-r approach to kataphora. Instead, we will assume that an incomplete RM may be introduced wherever the other rules allow it, but add the following constraint on completion: PROPER PLACE PRINCIPLE: (i) An RM subordinate position cannot be completed by an indefinite NP9 or by a proper name; (ii) An RM in the principal DR cannot be completed by a quantifying NP.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

DRS—2:

290 Forward References in Natural Language

Clause (i) will block a reading with narrow scope for 'a president' in (15) Whenever she was offduty, a president spent her time in the swimming pool. As a result, 'she' cannot have universal meaning. Now consider the sentence (16) The widow he loves is courted by each farmer.

(17) A rich widow attracts the attention of each poor bachelor with wide scope for the desirous bachelors:

. x widow(x) x attr attention of y one can generalise over pronouns and quantifying NPs as follows: OPTIONAL LEFTWARD MOVEMENT: Assume that the current sentence has introduced at least one RM into m. The pronoun or quantifying NP a can introduce its RM into a new box m' to the left of the current box, which has to contain only the conditions normally associated with a. (Formally: Replace m by m' —• m. Let y e V — UK be the RM newly introduced into m'. Now if a is a pronoun, then Con,/: = [y = a}. If a has the form 'every rj\ then Conm': = {77(y)}. Further, add [a: = y] to Con,,,.) The kataphoric reading of (16) would be derived by applying leftward movement to the pronoun 'he' in that sentence. S PATIENT DRT PUT TO WORK We have seen how Patient DRT deals with simple sentence-internal kataphors and how it manages to account for some of the difficult scope problems they raise. Now we will briefly show how Patient DRT deals with some of the difficult cases noted in section 2, namely mutual anaphora, full NP kataphora, and extrasentential kataphora.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Clause (ii) forbids that 'each farmer' completes the RM introduced by 'he' if this RM is part of the principal DR. Optional Rise allows a pronoun to be introduced into a 'higher' DR, but only if this DR already exists. If we want to derive the—somewhat problematic—kataphoric reading of (16), we should also allow the introduction of the pronoun into a new box immediately to the left of the current box. This resembles the behaviour that is sometimes noted in quantifying NPs, namely that they have wide scope over a preceding indefinite. Therefore, if one also wants to derive a reading of

K.. van Deemter 291

Mutual Anaphora We claim that, in contrast to the approaches of Heim and Kamp,1 approach can deal with mutual anaphora. It may be instructive to compare this with the situation in computer science when a program contains two statements, x and y, whose bodies contain references to each other. For instance, [*:] I F . . . THEN GOTO y... [y:] I F . . . THEN GOTO JC. ..

Full NP Kataphora Until now we have only dealt with cases where the kataphor is a pronoun. To account for non-pronominal anaphora, assume that a full NP introduces a setreference marker X along with a condition NP(X) (in the fashion of van Eijck (1983); if this NP is anaphoric to another NP with reference marker V, the relativised condition NP y (X) is generated. In the case of a relational NP (such as 'the parents'), NP y (X) holds if X contains the elements which stand in the required relation to the elements in the antecedent set Y (van Dccmter 1989). Now kataphoric full NPs can be covered if Patience is stretched to cover full NPs. For in that case, an NP can give rise to the condition NPV(X), even though Y was not introduced before. For example, the sentence (18) Whenever THE PARENTS sleep, THE PUPPIES do not eat can, on its kataphoric reading (see section 2), be represented as shown in Figure 3. The relativised condition (the parcnts)v(X) can be glossed as 'X contains the parents of all the elements in the antecedent set Y (and nothing more)'. Along these lines, an adequate kataphoric reading for the sentence

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Due to backpatching, parsing is not troubled by the mutuality involved in this situation. Using primes for translations, translation of x will contain the 'skeletal instruction' GOTO' y as a part. At this stage an address, say x', is allocated for x in memory. As a result, the relevant part of y can be translated: GOTO' .v'. This enables the program to substitute the address y' for y in the skeletal instruction which translated x: GOTO' y', which completes the translation. The same holds for our analogue of backpatching: after one 'pass' of a sentence with mutual anaphora, all the necessary information is collected. Our 'mutual anaphora' sentence (5) leads to the representation depicted in Figure 2. The embedding conditions of DRT arc satisfied if there is a man and a woman, x and y, where x works in y's department, y is the manager who hired x, and y interviewed x. The desired reading is obtained without difficulties.

292 Forward References in Natural Language

A woman who works in his department was interviewed by the manager who hired her. . x woman who works in his department (x) woman (x) x works in his department

x works in y's department the manager who hired her (y) the manager who hired x (y) x was interviewed by y Figure 2

.X

.Y

(theparents)y(X) (the puppies) (Y)

sleep (X)

not i

Figure 3

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

. y male (y)

K. van Deemter 293

results. All the rules for pronominal kataphora remain valid. For instance, the Proper Place Principle predicts, quite adequately, that a kataphoric reading of 'all customers' in (19) Whenever

ALL CUSTOMERS

were gone,

A TAXIDVLIVER fell

asleep

Extrasentential Kataphora Considerations of computational complexity can also be brought to bear on extrasentential kataphora. As we have seen, there is some doubt about the acceptability of kataphors such as the one in (20) First HE lost his wallet. Then HIS car got stolen, FRED was having a bad day

(Asher and Wada 1989). In order to account for the facts, we would propose to rule out extrasentential kataphora explicitly: NON-COORDINATION CONSTRAINT: Kataphors cannot cross sentence boundaries (more generally: coordination boundaries). Instead, we would explain (20) as a case of backwards anaphora. It has been argued (cf. Weijters 1989, van Deemter 1989) that proper names are no exception to the rule that definite NPs can be anaphoric. An example in Maes (1990) runs as follows: (21) The inventor of dynamite had a profound influence on the nature ofwarfare. Alfred Nobel. . . where the description 'the inventor of dynamite' is assumed to be an antecedent to the proper name 'Alfred Nobel'. Now if proper names can be anaphoric, the proper name 'Fred' in (20) can be analysed as anaphoric to the RM introduced

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(in which 'all customers' means 'all customers of the taxidriver') is only possible if'a taxidriver' gets wide scope over the tense-operator in 'whenever'. In other words, we claim that (19) is about one particular taxidriver, rather than about taxidrivers in general. Extra constraints on full NP kataphora may be needed. For instance, the facts suggest that definite NPs cannot so easily be equated to future antecedents.1' If true, this observation can be explained along the following lines: suppose we would allow identity kataphora by means of a definite description. Given that the distance between a definite NP and its antecedent can be very great (Grosz and Sidner 1986), a kataphoric reading can never reliably be inferred from the absence of backwards-anaphoric antecedents. Therefore, the possibility of a kataphoric reading would complicate resolution considerably. This argument does not apply to pronouns, as their antecedents can much more often be found in the current or previous sentence (Pinkal 1986, Ariel 1985).

294 Forward References in Natural Language

by 'he'. Note that, due to the Non-Coordination Constraint, (20) is not a case of kataphora and, consequently, the RM introduced by 'he' remains incomplete. This explains why (20) may be less than felicitous (lacking an antecedent for 'he' in an earlier sentence), but nevertheless understandable, in the same way as incompletely perceived discourse (cf. (n)). Note that if'Fred' is replaced by 'someone' the two NPs cannot co-refer, which is explained by the assumption that indefinites are never anaphoric (familiarity hypothesis). Thus, the NonCoordination Constraint precludes that the parser needs infinite patience: given any bound on sentence length, this constraint induces a bound on the maximal distance between kataphor and antecedent.

The Patient DRT treatment of kataphoric references shows that the 1-t-r approach to anaphora—which, we have seen in the introduction, has much to commend itself on independent grounds—can provide adequate descriptions for most kataphoric constructions. In dealing with these constructions, we have only accounted for grammatically possible readings, disregarding the further question how to decide which of them is most likely to be intended. However, Patient DRT raises a number of questions we cannot avoid saying a few words about. We will briefly discuss three of these, dwelling somewhat longer on the last one than on the other two.

When and why is Patient DRT's backpatching method an appropriate strategyfor dealing with indeterminism? In this paper, a number of phenomena are described that seemed to resist 1-t-r DRS construction and we have dealt with them by means of backpatching. But similar phenomena exist at other levels of parsing. For instance, at the level of speech recognition, phoneme pairs such as 'w' and 'u:' can only be told apart with the help of future phonetic material. Similarly, in syntactic analysis, only new syntactic material can decide whether, for example, 'flying' is a present participle (in 'Flying planes are dangerous') or an NP (in 'Flying is dangerous'). The same thing occurs at the level of semantic interpretation, since semantic ambiguities are often resolved by future context. We do not mean to imply that all these different-level kataphoric phenomena are to be treated by means of backpatching. A prudent general rule seems to be the following: when the processing of the kataphoric element faces finitely many 'resolution' candidates, then it is preferable to proceed by trial and error, backtracking over the different candidates; the above-mentioned

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

6 CONCLUSIONS

K. van Dccmtcr 295

examples belong to this category. When, to the contrary, there arc infinitely—or otherwise inconveniently—many candidates, then backtracking has to give way to a backpatching strategy such as outlined in the body of this paper. The kataphoric phenomena in the realm of DRS construction clearly belong to this category, since the number of possible referents for a kataphoric pronoun can, before the antecedent is processed, at best be limited to the universe of discourse as a whole.

What has Patient DRTgained us in terms of the prospectsfor incremental semantic interpretation?

(22) First he lost his wallet. (Then . ..) is interpreted as 'at least one (male) person lost his wallet'. Interpretation of unfinished sentences is problematic, however. To illustrate, suppose the language fragment in Kamp (1981) is enlarged with conditional sentences of the form 'S, if S,', then straightforward truth conditional interpretation of the first sentential part of (23) (John will succeed]s if he is lucky will, too optimistically, say that John will succeed. At this stage, it is unclear how serious these problems must be taken. Either they may be regarded as harmless semantic gardenpath phenomena—with sentence intonation, if and when it is available, coming in to provide extra information. Or, alternatively, they may be taken as arguments for the psychological reality of the level of Discourse Representations. For if a human interpreter of (23) has, after parsing the first clause, some degree of understanding of what is said, and if his understanding is not captured by truth conditions, then it might be hard to improve upon the DRS level as a reflection of this understanding.

Can kataphors be dealt with in DPL? From our point of view, one of the most promising rivals to DRT as a semantic theory of anaphora is J. Groenendijk and M. Stokhof s theory of Dynamic Predicate Logic (DPL) (cf. Groenendijk and Stokhof 1987, 1988). Their main

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

It is clear that, on the premises of DRT, 1-t-r DRS construction is a prerequisite for on-line interpretation. But it is still a long way from 1-t-r DRS construction to on-line interpretation. What we do have—due to the Non-Coordination Constraint—is interpretation per completed sentence. Incompletely perceived discourse (11) and sentences that purportedly contain extrasential kataphors are attributed interpretations in which unresolved pronouns arc existentially quantified:

296 Forward References in Natural Language

motivation is to design an alternative to DRT that has the same descriptive power, while operating strictly compositional (cf. e.g. Janssen 1983) in the construction of representations. Moreover, DPL adheres to the 1-t-r principle, as we will shortly see. Instead of the box-representations of DRT, DPL employs the syntax of normal predicate logic as a representation language. In order to explain the relevant phenomena (donkey sentences, discourse anaphora, etc.), the semantics of the logical language is changed in such a way that, most notably, existential quantifiers bind variables beyond their scopes. To illustrate, (24)(a) is translated as (24)(b):

Given DPL's semantics, the existential quantifier in (24)(b) binds all the occurrences of x. Translation proceeds in two steps. First, 'Someone walks' is translated as 3x: Walk(x), then 'He talks' is translated as Talk(x). (24)(b) is obtained, as it were, by simple concatenation of these formulas.12 In DRT, by comparison, the addition of a new sentence to an existing DR takes place without a separate representation for the newly parsed sentence; instead, the existing representation is modified in one of several ways.13 DPL avoids such inherently procedural doings: the semantics of the existential quantifier suffices to get the bindings right.14 In order to provide formulas such as (24)(b) with the appropriate meanings, DPL has them denote 'state-changers': technically, formulas denote pairs of assignments, where g is an input assignment (input state) and h an output assignment. In other words, /; may result if the formula is processed in g. Assume, for instance, that£ is the input state; then the processing of (25) 3x: Walk{x) changes £ into a state h that assigns an individual h(x) to x such that h(x) walks. Formally,

(26) \\3x: (x)\\ = { I 3k:g[x)k & t \\\\} whereg[x]k holds ifg(y) = k(y) for each variable y such that y^ x. The fact that processing operates from left to right can be seen from the definition of sentence conjunction: (27) \\(f> and ifj\\ = [ I 3 € \\(f>\\ & e I

It is due to this non-commutative conjunction that variables to the left of a quantifier cannot be bound by a quantifier. Now, given the attractiveness of the DPL perspective, one may try to accommodate kataphors in DPL. Processing

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(24)(a) Somebody walks. He talks. (24)(b) 3x: Walk(x) & Talk(x).

K. van Deemter 297

order could be reversed from 1-t-r into r-t-1 (right-to-left) by the following move:

(28) \\(f> and \3k:

e \\tfj\\ &

But if this is done, only kataphors are allowed, disallowing (backwards) anaphors. In order to make DPL suitable for both kataphors and anaphors, the following combination of (27) and (28) might be proposed, in which conjunction is stipulated to be commutative again:

(29) \\<j> and0|| = [\3k: (e\\48c4\\)}.

(< ?1 k> e \\\\ &

(3o)(a) He walks; Someone talks, (3°)(b) Someone talks; he walks, become equivalent: given an assignment g, both arc true if either somebody walks and talks, or g(x) walks and somebody (possibly somebody else) talks. Overgeneration would, as ever, have to be prohibited by a set of constraining rules. For instance, the kataphoric reading of (3o)(a) could be ruled out by the Non-Coordination Constraint from section 5. However, it will be clear that, from our perspective, the proposal contained in (29) is unattractive, since it would bereave DPL of its 1-t-r orientation.1'' Of course, DPL's 1-t-r perspective can be maintained if the 'patient' approach we have described for DRT is adopted in DPL: an assignment g that is undefined for a variable x may process x 'incompletely', to be completed by a subsequent quantifier under certain conditions... Although this is, technically speaking, possible it would be at odds with the philosophy of DPL to introduce such blatantly procedural elements into a neatly compositional framework. Thus, it seems that although kataphors can be reconciled with the principle of 1-t-r processing, this can only happen at the expense of compositionality. In other words, it might be that non-composirionality of translation is an asset, rather than a disadvantage of DRT. KEES VAN DEEMTER Institutefor Percqjtion Research (I. P. O.) NL - $600 MB Eindhoven The Netherlands

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

As a result, a quantifier 3x can either bind or not bind a given occurrence of x, no matter whether x precedes or follows the quantifier. Consequently, (3o)(a) and (b)

298 Forward References in Natural Language

NOTES 1 The first statement of an 1-t-r principle I have been able to find is Hintikka and Carlson's Progression Principle (Hintikka

2 3

5

threatened the position of the military, THE

KING was toppled. Here it is impossible, on the account of Asher and Wada, to get the reading where the tense operator in 'whenever' has wide scope over the one in 'the king'. 8 In particular, exactly the same result is obtained if the verification clause for boxes of the form />, -+ b2 is redefined in such a way that all the new variables in b, are quantified, rather than only those which have been introduced in b, (i.e. introduced in Ubl). 9 Note, however, that NPs of the form 'a(n) ...' have to be excluded from the first clause of the PROPER PLACE PRINCIPLE

when they are used in a generic sense—to be distinguished from a strictly universal sense. (An example would be a presenttense version of (15): Whenever she is off-

(a) Joan—no doubt the person he admired most—kept Peterfrom going insane. (b) Joan—no doubt the person Peter admired duty, a president spends her time in the swimming pool) While earlier versions of most—kept himfrom going insane. DRT have, I think, not dealt with generic NPs, we will simplify matters somewhat 6 In principle, processing order within a and assume that genericity is not a struc'monolithic' DR is constrained by scope: if tural phenomenon and that any presentNP, and NP, do not belong to &,, b2, tense subject NP of the form 'a(n). ..' can respectively, where i, and b2 stand in the be used generically. Such generic NPs will relation i, -• b2, and if NP, has wider be considered as quantifying, rather than scope than NP,, then NP, is processed indefinite NPs, and consequently they will before NP,. However, in Kamp (1981), not be affected by the restriction in (i). scope order is simplified to coincide with surface order. As a result, as Kamp him10 In Kamp (1981), the desired reading of self points out, Kamp's mechanism does two mutual anaphors that do not belong not account for kataphors such as (10), to different boxes (as in (5)) cannot be where there is a difference between the obtained, since processing within such a surface position of the antecedent and its monolithic DRS does not allow any scope. An even more radical departure departures from 1-t-r processing order from P i than described in Kamp (1981) (see note 6), while DRT also, of course, would be needed in order to get the right does not have provisions for backpatchpredictions in cases like this. See also our ing.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

4

and Carlson 1979). However, as in Heim (1982) (cf. section 3), this principle applies to a level of logical form rather than surface structure. cf. section 3, where Kamp's treatment of kataphors is briefly discussed. For instance, DRS construction in (Kamp 1981) requires syntactic information which can only be available if syntactic analysis can 'look' quite far 'ahead'. Note that the unacceptability of (2) is predicted by Reinhart's Non Coreference Rule, for in (2) the pronoun c-commands the proper name 'Dorit'. An interesting situation in which kataphoric constructions (a) are preferable over their backwards-anaphoric counterparts (b) obtains when a discourse entity (here: Peter) is introduced in a part of the text that is marked as optional reading:

later remarks about Asher and Wada's proposals. 7 A possibly more natural example, where the 'disposed' kataphoric antecedent is a definite description with an underlying temporal quantifier, runs Whenever he

K. van Deemter 299 Updating existing DRs is done in Kamp's rules CR1-CR5. In Patient DRT, rules such as the Completion Principle can change an existing DR in even more ways. In this respect, DPL is in line with the claim in Chierchia and Rooth (1984) to the effect that DRT's embeddability definitions make a definition of accessibility redundant. Actually, as Stokhof pointed out to me, the new definition of conjunction (29) would make DPL virtually indistinguishable from H. Zeevat's system of Static Semantics (cf. Zeevat 1990), where directionality is abandoned completely. From our perspective, of course, Zeevat's proposal has the same drawback as the bidirectional version of DPL (cf (29)), namely that it fails to observe the 1-t-r principle.

REFERENCES Aho, A. V., Sethi, R. and Ullman.J. D. (1986), Chierchia, G. and M. Rooth (1984), 'ConfiguCompilers, Principles, Techniques and Tools,rational notions in discourse representaAddison-Wesley, Reading, Mass. tion theory', NELS 14, Graduate Linguistic Ariel, M. (1985), 'The discourse functions of Student Association, UMASS, Amherst. given information', Theoretical Linguistics, van Decmtcr, C. J. (1989), 'Anaphora Across No. 12. the Board', IPO Ms. 684. Ashcr, N. and Wada, H. (1989), 'A computa- van Eijck, J. (1983), 'Discourse representation tional account of syntactic, semantic and theory and plurality', in Ter Meulcn (ed.), discourse principles for anaphor resoluStudies in Modeltheoretic Semantics, tion', Journal of Semantics, vol. 6. Dordrecht, Foris. Bach, E. (1970), 'Problominalization', Linguis- Grosz, B. and Sidner, C. (1986), 'Attention, tic Inquiry, No. 1. intentions, and the structure of discourse', Bcn-Ari, Y. (1982), Principles ofConcurrent Pro- Computational Linguistics, vol. 12, No. 3, gramming, London, Prentice Hall Inter175-204. national. Groenendijk, J. and Stokhof, M. (1987), Bosch, P. (1983), Agreement and Anaphora: A 'Dynamic Predicate Logic', Faculteit dcr Study of the Role of Pronouns in Syntax and Wijsbcgeertc, University of Amsterdam, Discourse, Academic Press, London The Netherlands. Buehlcr, K. (1934), Sprachtheorie,Jena, Fischer Groenendijk, J. and Stokhof, M. (1988), 'Con(reprint, Stuttgart, Fischer, 1965). text and information in dynamic semanChafe, W. (1976), 'Givenness, contrastiveness, tics', in B. Elsendoorn and H. Bouma (eds), definiteness, subjects, topics and point of Working Models of Human Perception, New view', in C. N. Li (cd.), Subject and Topic, York, Academic Press. New York, Academic Press. Hcim, i. R. (1982), 'The semantics of definite

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

11 Some cases were brought to my attention 13 where a full NP definite appears to be used as a kataphor. A typical example is Even if the bastard ends up bringing in FIFTY million, I don't think Berkeley should hire14 Ronald Reagan to teach political science. What counts is whether this sentence is felicitous at a stage of a discourse where Reagan is not in focus. (Compare our reasoning in relation to sentences (3) and 15 (4)). The situation is not entirely clear bur counter-examples against the suggested constraint may exist. 12 What really makes the treatment compositional is the fact that the calculation of (24)(b) can be viewed as a strictly semantical operation, performed on the meanings—rather than the logical translations—of the constituent sentences (cf. Janssen 1983).

300 Forward References in Natural Language

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

expository texts', Journal of Semantics, vol. and indefinite noun phrases', Ph.D. dissertation, University of Massachusetts, 7.2, 143-174 Amherst, Mass. Pinkal, M. (1986), 'Definite noun phrases and Hinrikka, J. and L. Carlson, (1979), 'Condithe semantics of discourse', in Proceedings of tionals, generic quantifiers, and other COLING-1986 (Bonn), 368-73. applications of subgames', in E. Saarinen Reinhart, T. (1976), 'The syntactic domain of (ed.): Game-theoretical Semantics, Reidel, anaphora', Ph.D. dissertation, MassachuDordrecht. setts Institute of Technology, Cambridge, Mass. Janssen, T. (1983), 'Foundations and applications of Montague grammar', Ph.D. thesis, Reinhart, T. (1983), 'Co-reference and bound University of Amsterdam, The Netheranaphora: a restatement of the anaphora lands. questions', Linguistics and Philosophy, vol. 6. Kamp, H. (1981), 'A theory of truth and Sanford, A.J. and Garrod, S. C. (1981), Undersemantic representation', in J. Groencnstanding Written Language, Chichester/ dijk, T. Janssen and M. Stokhof (eds), New York, John Wiley and Sons. Formal Methods in the Study of Language,Sidner, C. (1983), 'Focusing in the compreMathematical Centre Tracts no. 135, 277hension of definite anaphora', in M. Brady 322. and R. Berwick (eds), Computational Models ofDiscourse, Cambridge, Mass., MIT Press. Karttunen, L. (1971), 'Definite descriptions with crossing co-reference', Foundations of Weijtcrs, A. (1989), 'Denotation in discourse', Language, No. 7. Ph.D. thesis, Katholicke Universitcit Lea, W. (1980), Trends in Speech Recognition, Nijmcgen, The Netherlands. New York. Zecvat, H. (1990), 'Static semantics', inj. van Bcnthem (ed.), Partial and Dynamic SemanMaes, A. 1990, 'The interpretation and repretics 1, DYANA report for ESPRIT BR3175, sentation of co-referential lexical NPs in Del. R2.1.A, January 1990.

Journal ofSemantics 7: 301-319

© N.I.S. Foundation (1990)

A Psycho-Linguistic Crossroads: Frequency of Use ERICA C. GARCIA University ofLeiden

Abstract

1 INTRODUCTION* No argument is presumably necessary for the mutual relevance of linguistics and psychology, but the self-evidence of the connection is commensurate only with the difficulty encountered in actually establishing it. In this paper we attempt to show the relevance to the problem of a certain type of data—the frequency distribution of linguistic units, something largely ignored in linguistic analysis. Even the most cursory examination of actual language use reveals obvious differences in the frequency with which different forms occur in comparable (but none the less distinguishable) environments. The skewings in distribution carry through to finer and finer levels of (syntactic) analysis; it follows that the skewed distribution is not arbitrary but governed by an inner logic. The distribution of a form should follow, ideally, from its own value and place within the linguistic system: we will argue that in order to establish a connection between linguistic structure and frequency ofuse, (general) cognitive considerations must be appealed to, though—as a mere linguist—we are unfortunately unable to indicate what specific principles must come into play. The general analytical problem will be illustrated with data on the use of deictic forms. We will argue that the (linguistic) difference between competing devices can be compared to and matched with the greater or lesser attention involved in the identification of the referent; the referent's salience consequently becomes a critical parameter for the choice of form. However, different strategies can be followed in order to establish what counts as 'salient'.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The units postulated in linguistic analysis differ not only as to their structural properties, but also in their relative frequency of use in different contexts: the former should, ideally, shed light on the latter. Such a connection between linguistic analysis of the system and the use to which it is put requires an understanding of what constitutes 'appropriate' use of a form, and an appeal to psychologically plausible cognitive processes. The need to make psychological sense of linguistic frequency data is discussed on the basis of distributional skewings in the exploitation of stronger vs. weaker deictic devices in Spanish, Portuguese and Dutch, with special attention for the problem posed by competing strategies.

302 A Psycho-Linguistic Crossroads: Frequency of Use

It is finally argued that in order to relate linguistic structure to actually observed (frequency of) use, it is necessary to (i) identify the cognitive categories underlying the choice of form; (ii) operationalize those cognitive categories in linguistic terms, other than the forms at issue. In particular, it is necessary to identify what (contextual) traits instantiate specific cognitive values, in order to ascertain whether the expected use is in fact observed.

In their important paper on the establishment and maintenance of reference, Marslen-Wilson, Levy & Tyler correctly point out that (1982: 340): 'if language production . . . is tuned to the properties of the comprehension system, and given a theory of the properties of this system, then this theory should be consistent with the patterning of different referential devices.' Different referential devices must, of course, differ in some way (op. cit. 345, 359) presumably relevant to the psychologist's goals. But for the linguist—who is responsible for the grammatical analysis of the language—nothing is less given than the 'different referential devices' that constitute the psychologist's starting point. It is the grammarian's task, precisely, both to identify the units and to characterize them in such a way that their use makes sense. And indeed, the first problem faced by the linguist is to decide what are, in fact, different occurrences of the same form. The indubitable existence of homonyms forces the question whether all instances of what 'sounds the same' are, in fact, uses of the same linguistic unit. For example: is the it of It rains the 'same' it of I saw it; and is the that in That is a pencil the 'same' that as in / think that you are mistaken, I want the pencil that he gave you, Speak clearly that he may hear

you, and so forth? A decision as to what range of uses can reasonably be regarded as exploitation of the 'same' form obviously requires a prior understanding of what in principle makes a use possible, namely, a plausible and motivated connection between the value postulated for the form and the cognitive operation exemplified by its use. Such a principled connection becomes even more desirable if one asks of a grammatical analysis that it shed light not only on where forms show up, but also on how often they do so.1 For instance: it is well known that (i) both that and which can connect a relative clause to its antecedent, as in: I want the pencil that he gave you,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

2 THE PROBLEM

E. C. Garcia 303

I want the pencil which he gave you, and that, in fact, no explicit connection is necessary, as shown by I want the pencil 0 he gave you; (ii) the frequency with which, that, which and 0 are resorted to as 'relative pronoun' varies, depending on such factors as the animacy of the antecedent, its role in the relative clause, the degree of juxtaposition of antecedent to clause, etc.

from the point of view of linguistic theory, it would be much more interesting if patterned behavior in linguistic variation could be shown to be a result of linguistic competence, rather thanjust a. part of it, in the same way as the regular behavior of a coin uponflippingis a resultoi its physical structure. The ideal of linguistic analysis is, undoubtedly, to achieve a qualitative description of language structure that will shed light on the quantitative properties of language use—especially since diachronically the latter determine the former (c£ DuBois's (1987: 851) dictum 'Grammars code best what speakers do most'). Now it might be supposed that frequency of occurrence—and indeed occurrence tout court—is merely a reflection of speakers' communicative needs, and hence of no real interest to the study of language (or of the mind), cf. Chomsky (1957: 16-17; 1966: 36). If, however, the (relative) frequency of occurrence of different forms reveals a regularity that recurs regardless of the subject matter, this regularity may be due not so much to what the speakers want to say—in the sense of what they refer to—as to how they find it convenient to say it. Quite naturally, then, functional analyses (e.g. Givon 1983) explicitly aim at explaining the qualitative/quantitative continuum reflected in language use. But there are two critical points where they generally fall short of their goal: (i) the grammatical devices whose use is to be accounted for are not characterized in such a way that the devices can be 'matched up' with their use other than by (arbitrary) stipulation; (ii) the (cognitive?) process responsible for the match is not specified. For instance, Givon (1983: 17) discusses a series of different referential devices, and ranks them according to a scale that is claimed to 'code' the functional

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The important point, now, is that the distribution is non-symmetrical, i.e. one option (say 0 ) is favoured under different circumstances from those where another (e.g. which) is most frequent. This is the subject matter of syntactic variation. One would imagine—and hope—that the different nature of 0 , that, and which ('nothing' vs. explicit connector, a 'demonstrative' vs. an 'interrogative' pronoun) is not irrelevant to their greater or lesser frequency in particular types of contexts. As Naro (1980: 165-6) puts it:

304 A Psycho-Linguistic Crossroads: Frequency of Use

3 METHODOLOGICAL DESIDERATA The psychological interpretability of linguistic analyses is not merely desirable in itself: it is required by considerations strictly internal to linguistic analysis. In this section we shall attempt to sketch why (relative) frequency of use bears on linguistic analysis, and how cognitive processes are relevant to the connection. We begin by listing some obvious criteria to be filled by a linguistic analysis. Psycholinguistic research quite naturally investigates either speaker or hearer strategies (and chiefly the latter,Brown & Yule 1983: 2i5;Tomlin 1987: 457), but every language user obviously alternates between the two roles. There is thus at least the possibility that the same type of knowledge about linguistic forms is used differently, to solve different problems, rather than two different (speaker vs. hearer) types of knowledge being associated with each linguistic form. Neutrality between speaker and hearer is therefore the first requirement for a linguistic analysis. The very postulation of a single type of knowledge that is put to different uses means that we must distinguish between the knowledge that a language user has of a linguistic form and his actual use of it: this, in turn, implies abstraction, and the subsumption of different uses under one denominator. It is here that, as pointed out in section 2, linguistsfiercelydisagree (Fox 1986:26-7): what are realizations of the same form, why they count as realizations of the 'same' form, and how the form is to be defined is a central problem in linguistics.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

domain of'topic identification in discourse'. He does not, however, state exactly what in the nature of different devices results in dieir being rankable along that particular scale, nor—and this is the really criticalflaw—doeshe make clear why the one end of the linguistic scale should correspond to 'most continuous/ accessible topic' rather than to the other extreme of the functional domain. Iconicity is claimed to underlie the projection (Givon 1983: 18), but the nature of that iconicity is never stated. What is lacking, then, is a characterization of the devices, independent of their use, from which this use can plausibly be expected to follow. A linguistic analysis that attempts to take quantitative facts of usage into account cannot but invoke, then, cognitive considerations relevant to the production and comprehension systems—just as the psychologist relies on independent information about the different referential devices. At which point one begins to suspect that two unknowns may have to be solved simultaneously: the linguistic system, and the production/comprehension strategies applied in the exploitation of the system.

E. C. Garcia 305

(i) The communicative problem for which the different forms in question are resorted to must be identified. For instance, in the case of referential devices, the problem at issue appears to be the identification of different kinds of referents (or of the same type of referent under different circumstances). (ii) The different forms in which the general problem poses itself must be mapped on to a single cognitive dimension along which they can be ranked from most-to-least. In the case of pronominal/deictic reference it is generally assumed that the relevant dimension involves attention? an example of a motivated ranking is found in Marslen-Wilson et a/.'s discussion of reference within narrative structure (1982: 35 5-6, passim). (iii) The various linguistic devices resorted to for the solution of the communicative problems listed under (ii) must be identified and described, both as to form and content. In particular, a linguistic description must specify the information contributed by a linguistic form. An example of a (partial) characterization is, for instance: she: 3rd person fern. s g(iv) The various linguistic devices identified under (iii) should be mapped on the same cognitive dimension as in (ii), and ranked along this dimension; the rationale of the ranking must be made explicit. (v) The double mapping/ranking carried out under (ii) and (iv) yields, in fact, a matching of linguistic devices with particular cognitive tasks. This

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

It sometimes appears that the solution to this problem is sought in description, via the listing of different uses: cf. Fox's (1986:28, 37,41) identification and discussion of allegedly conventionalized patterns of pronoun use.2 The attempted escape from the dilemma is, however, illusory: some identification among different uses must take place if any pattern is to emerge at all. It is clear that no linguistic analysis can avoid positing unobserved entities if it is ever to make sense—via the language user's finite knowledge—of the infinite use to which the language is put. If the question, then, is not whether to abstract, but only how, or how much to abstract, a control must be sought for the units hypothesized, and the only reasonable place where such a control can be found is in their actual usage. Usage, however, is the output of actual production, and the input to actual comprehension. If usage data—such as the relative frequency of different options—is to be the touchstone and testing-ground of our analysis, our analytical categories themselves must be capable of interpretation in terms of a mental substance or cognitive currency that relates to production/comprehension processes. What would such an interpretation look like? In our opinion the following requirements must be met:

306 A Psycho-Linguistic Crossroads: Frequency of Use

matching may perhaps be seen as a kind of 'x is a y' conceptualization, which we take to underlie the process of symbolization.

4 P R O N O M I N A L REFERENCE AND A T T E N T I O N We will now try to illustrate the 'matching' of linguistic and cognitive categories just discussed with data bearing on 'reference' via pronominal and deictic expressions in Spanish and Portuguese: the attempt exphcitly to match forms to tasks will acquaint us with the problems that arise in the operation. We start from a basic—and presumably uncontroversial—assumption, namely, that the very fact of reference involves an appeal to the hearer's

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

From this matching of forms with tasks there follows an expectation of 'preferred' (i.e. normal) use: if human beings function at all economically, the device providing 'most' of whatever it is that the cognitive dimension involves should naturally be more often resorted to for the tasks calling for 'the most' of that same dimension. In shore once we know the relative position of two communicative tasks on the cognitive dimension (for instance, their relative demands on attention) we should be able to predict which of two (likewise rankable) linguistic devices should be the one preferred for a particular task. The testing of such a prediction is, however, not unproblematic: the communicative tasks are essentially cognitive in nature, and do not occur as such in a text. It is accordingly necessary to 'interpret' them linguistically (or to interpret the text 'cognitively'), i.e. determine what properties of a text, specifically of the contexts within which the forms at issue appear, are natural reflexes of a specific communicative problem or cognitive task. Only then will we be able to tell whether a form 'matches up' with the task to which it is most appropriate, by seeing whether it indeed occurs in the 'appropriate' linguistic contexts, i.e. those instantiating the right task. This means that the identification of the communicative problems referred to under (ii) above must be accompanied by a motivated characterization of the actual shape they take in recognizable and definable grammatical, semantic, or pragmatic categories that are identifiable in a text. For instance, the greater or lesser demand on, or appeal to, attention that is made by different referents should be traceable as syntactic or pragmatic properties of the pronoun's 'antecedent'. In summary: a prediction is tested by establishing whether or not the occurrence of particular linguistic devices correlates positively with specific properties of the text which can plausibly be viewed as a reflection of the communicative tasks to which the said linguistic devices are (respectively) more appropriate.

E. C. Garcia 307

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

attention. As Ehlich points out (1982: 324): 'To communicate effectively, the hearer's attentiveness needs to be brought into concord with that of the speaker . . . By means of deictic expressions, the speaker brings the hearer to focus on some specific, directly accessible element.' Tomlin states (1987:458): 'the syntax of reference in discourse production is tied directly to psychological processes of attention . . . the alternative use of a noun or pronoun in discourse production is a function of attention allocation by the speaker.' But if this is the case, our characterization of referential linguistic devices should be such that—given plausible assumptions about how humans husband their attention—we can show how the properties of, say, pronouns vs. deictics vs. nouns—as linguistic forms —make these different resources more or less appropriate for die various attention-calling and attention-allocating manoeuvres that are called for in coherent discourse. The view most frequently taken of pronouns is that they are indeed pronouns,4 i.e. expressions referring back or forward to an 'antecedent' nounphrase lexically identifiable in the surrounding discourse. Though this is not the whole story,D it is undeniable that adequate tracking of reference—whether it be maintained or switched—plays a crucial role in the comprehension of discourses of any complexity. It is equally evident that different entities referred to in a discourse are not equally present in short-term memory: some are more salient than others, and hence come to mind more readily. In terms of attentiveness, an entity that already enjoys the hearer's attention needs to have attention drawn to it less than one that must be looked for, or sorted out from among other possible candidates. Since linguistic devices differ in the amount of information they convey with respect to the nature of the envisaged referent, it is not at all surprising that (Marslen-Wilson et al. 1982: 345-6) the degree of specificity of a referring device should be inversely proportional to the extent to which the discourse context allows the referent to be identified. After all, an imprecise, uninformative, non-salient referential device forces the hearer to rely on contextual clues for the identification of the relevant referent, and a contextually obvious entity is precisely the one that is given by the context. The identifying force of the context, then, both allows and is evoked by the uninformativeness of the referential device: the 'matching' of form to task to context is as natural as it is inevitable. The mechanism is particularly clear in the prototypical case of weak reference—zero mention—which is characteristically resorted to under conditions of maximal contextual givenness. And indeed, when the context, by itself, forces an entity on the hearer's attention, any extra pointing would be not merely superfluous but potentially misleading. Conversely, since the more informative device is die cognitively more salient one, and thus lays a higher claim to the interlocutor's attention, it is more appropriate for the task making the higher demand on attention. In Figure 1 we

308 A Psycho-Linguistic Crossroads: Frequency of Use

Cognitive Task

Linguistic resource most efficient matching MORE

Refer to maximally given referent

A

t\

A Z 0

x

Minimally informative device

y

Maximally informative device

H

Z Refer to minimally given referent

7£-1

H H

Figure I Matching of resource to task

attempt to sketch the matching of form to task just discussed. The distributional consequences of Figure i are self-evident: in principle, form x should be used only for task A, while cases of A should reveal only use of form x (and, analogously, Z and y). Such an ideal one-to-one match is, as we shall see below, departed from in practice. But if Givon's maxim (1983: 18), 'Expend only as much energy on a task as is required for its performance', indeed governs human beings' behaviour, and thus also their choice of linguistic form, we should in any case expect most x's and A's matching up as predicted. A particularly clear example of the kind of distribution discussed is the use of the Spanish 3rd person personal pronouns si vs. el, ella, ello, ellos, ellas 'he, she, it, they (masc), they (fern.)'. Both alternatives refer exclusively to the third person; si (traditionally but incorrectly analysed as a 'reflexive' pronoun) is invariable, while el etc. provide number and gender information about their referents. (Hereafter we use // as a cover term for the entire paradigm opposed to the invariable si). If means are suited to ends, we should expect the maximally imprecise si to be preferentially used for contextually very obvious referents. This is indeed the case: we find (Garcia 1983,1986, in press) that the overwhelming majority of sfe antecedents are contextually more given than e/'s. This contextual givenness manifests itself in the fact that they (i) are intra-sentential—hence are presumably processually more accessible than extra-sentential antecedents; (ii) have a central, or prominent, syntactic role, such as Subject and Accusative Object—hence can presumably be expected to be in the forefront of the interlocutor's attention;

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

LESS

E. C. Garcia 309

(iii) have no competition from other entities in the immediate context—hence are 'uniquely' given. Conversely, reference to extra -sentential antecedents, to antecedents in noncentral roles, or to an antecedent that must be distinguished from among other, similar, entities in the immediate context strongly motivates recourse to el. The same basic pattern recurs in other languages and grammatical domains. One example is semi-literate Rio de Janeiro Portuguese, where explicit (extra) reference from within the relative clause to the antecedent of the relative pronoun alternates with the absence of such an extra mention, as illustrated in the examples below:

O //'wo [que eu comprei (ele)] ebom

'The book [that I bought (it)] is good' Oftlme [queeufalei (dele)] ebom

'The film [that I talked (about it)] is good' We begin by considering the inferential problem posed by the grammatical nature of relative clauses. A relative clause is defined by the lack of one argument, which must be located outside the boundaries of the clause, but before the introductory 'relative pronoun'—in this case, the invariable que. It is in fact this search for and identification of the 'antecedent' that relates the information provided in the relative clause to the antecedent, making of the combination a coherent cognitive whole. Successful integration of the antecedent into the relative clause clearly depends on two conditions being mev. (i) the antecedent is recognized as belonging to the clause; (ii) the clause is perceived as requiring the antecedent. The first condition is satisfied by word order: the antecedent closely precedes the introductory que, so diat the 'connection' between the two is patent. The second condition is met by the existence, within the relative clause, of an informational gap: the antecedent is indispensable if a coherent whole is to be made of the contents of the clause. We have seen, however, that in semi-literate spoken Rio de Janeiro Portuguese the functional gap within the relative clause is sometimes filled: but with a pronoun, a form that provides only partial information (gender/ number) as to the nature of the antecedent. By being there, the pronoun gets noticed (and obviously more than mere 'nothing' would be); but being merely a

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

O menino [que (ele) estuda] aprende

'The boy [who (he) studies] learns'

3io A Psycho-Linguistic Crossroads: Frequency of Use

pronoun, however, the gap-filler fails to provide full information. It consequendy triggers (i.e. reinforces) the search for an antecedent—which is, of course, the antecedent to the relative clause as such. We find, now, that insertion of the pronoun is preferentially resorted to (Mollica 1981) under conditions of contextual opacity of the connection between clause and antecedent, namely:

In shore both in the Spanish as in the Portuguese case we find skewed frequency distributions where the stronger device is preferentially used when the intended referent cannot be immediately and effortlessly identified from the immediate context. If the contexts are characterized in cognitive terms relevant to the problem, then, an appropriate and economic matching of linguistic resource to communicative function is observed.

5 ICONIC REFERENCE The question now is whether this matching between the informativeness of the linguistic device and the context-givenness of the referent can fully explain the use of forms, i.e. how often and where they are used. Because, as Yule correctly points out (1982: 318-19), identification of reference is not always necessary for successful communication: from which it follows that the contextual obviousness of a referent need not be the only factor that influences the choice of stronger/weaker referring devices. The intrinsic nature of the referent may play a role as well in the choice of linguistic expression. It is actually very understandable that this should be the case, since one points things out not merely because they are hard to identify but also—in fact primarily!—because they call attention to diemselves. Cologne Cathedral is regularly pointed out by and to awestruck tourists not because it is likely to be missed (a fundamental reason for pointing out the minutiae of the carving on the choir stalls) but because Cologne Cathedral is on no account to be missed. An alternative rationale for the use of referential devices, then, is to resort to the stronger device for the intrinsically more attention-deserving referent. This would match form to task according to the principle that it is appropriate to pay

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(i) when the antecedent does not immediately precede the relative pronoun: in this situation the connection between relative clause and antecedent is harder to make, since the greater distance between the two elements detracts from the salience of the antecedent; (ii) when the referent's role within the relative clause is a more peripheral one, so that the relevance of the antecedent to the clause is less readily perceived.6

E. C. Garcia 311

Table 1 Percentage of si (total si + // references) for human/inanimate referents (different syntactic conditions) Author

Referent

Condition

Cortazar Pp. grouped with entity to which reference is made

Human Inanimate

Reference to subject: no competition from similar entity in immed. context

Human Inanimate

Reference to subject: competition in immed. context from similar entity

Human Inanimate

86 100

32 9i

Mallea

Sabato

94

88

100

100

M. Estrada 100 02

87

88

95

100

(100)

100

low figures 16

75

20 80

60

80

75

95

6 THE LOCUS OF STRATEGIES If our characterization of the two sets of skewings is correct, we (seem to) have a problem on our hands: the one strategy appears to reverse the other, since the strong referential device is used both for inherently prominent and for contextually non-salient referents. This is shown in Figure 2, where we summarize the data. In both cases the use of the referentially stronger device

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

more attention to inherently important entities, which could roughly be formulated as 'Refer iconically: use the referential device which by its own nature most effectively suggests the intended referent'. The referential devices discussed in the preceding section are, in fact, also used in this new fashion. We find that in Spanish—under the same syntactic conditions—the percentage of si (the weaker, less informative form) is higher in the case of inanimate than of animate referents, as shown in Table 1, where data are presented from four corpora of Spanish prose written by distinguished Argentine writers. Similarly, in Portuguese (Mollica 1981), presence of the extra reference in a relative clause is favoured by the animacy (especially humanness) of the referent.7 Once more, then, we have a plausible matching between linguistic device and a cognitive property—in this case, the extent to which the referent draws attention to itself—with the most 'natural' couplings being preferred in actual use. But this new plausible matching is obviously very different from the one discussed in section 4 above!

Braz. Port.

c o.

n

SALIENCE

Device favoured by Intrinsic

Spanish

oaded from jos.oxfordjournals.org by guest on January 1, 2011

Language

p OP

o

Distant antecedent: less salient

stronger: el (gender/number specification) weaker: si (invariable)

Animacy: more salient

stronger: extra pronominal reference

Animate, Singular: more

Non immediate antecedent, Peripheral role in clause

weaker: no extra reference

Inanimate, Non Singular: less

Immediate antecedent, Central role

Inanimacy: less salient

n

Configurational

is

Close antecedent: more salient

Figure 2 Recourse to stronger/weaker referential device, depending on intrinsic/configurational salience

:less

c o

E. C. Garcia 313

ability.9

These two ways of engaging attention may be related to the two cardinal tasks performed by any speaker, who must, in the first place, articulate (i.e. categorize) his experience: this would explain why intrinsic prominence should call his attention and hence affect his choice of referential device. But categorization is only half the story, the reference must be integrated with other references and other forms, i.e. it must be embedded in a context. It is this second, syntactic, operation that would make configurational considerations , relev;ant.10

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

correlates with intrinsic noteworthiness, but with contextual non-salience of the referent. We cannot avoid the problem by viewing the two skewings as the reflex of distinct strategies that 'conventionally' apply under different conditions (Fox 1986:40-1,42-3,passim): the contradiction we face is embodied in one and the same set of data, since the parameters in question co-occur, and every instance of pronoun use is both contextually and intrinsically motivated. Note, moreover, that the skewing for intrinsic prominence can in no case be viewed as an effect of discourse structure: humans are likely to be 'topics', and as such likely to be repeatedly referred to. This should make them contextually more given than inanimates: yet, at least in Spanish, it is precisely humans that score a lower percentage of the weaker si. Finally, we observe that the same double-use rationale also holds for direct appeals to attention via extra-linguistic devices such as gestures (MarslenWilson etal. 1982: 346-51) and pitch accents. In Dutch, for instance, presence vs. absence of a pitch accent has been shown (Nooteboom etal. 1980) to depend on whether the accented element refers to the single most probable choice among potential referents in discourse: presence of accent is thus favoured by non-predictability, i.e. contextual non-salience of the referent. But in Dutch (as in many other languages, cf. Marslen-Wilson etal. 1982: 376 fn. 22) prosodic accent can also serve to emphasize its referent, for any discourse relevant reason, independently of context-givenness. Both/either contextual and/or intrinsic 'salience' can thus trigger the use of accent. Why does a single expressive formal dimension correspond to two distinct strategies? What is it that allows intrinsic prominence and contextual nonsalience to be expressed by the same linguistic form?8 And, more generally, how many, and what sort of cognitive considerations may conspire or conflict in determining the choice of form? The question cannot be ignored, and must be answered if frequency of use is to be explained, or serve as a clue to cognitive processes. We should like to suggest that in the cases discussed above the apparent 'contradiction' observed in the use of referential devices actually reflects different cognitive operations: attention may relate both to inherent noteworthiness and to contextual notice-

A Psycho-Linguistic Crossroads: Frequency of Use

7 C O N F L I C T I N G MOTIVATIONS If the preceding line of reasoning is accepted, we must conclude that what is needed in order to explain the frequency-distribution of referring expressions in discourse is not merely (i) an independent analysis of linguistic devices that makes it possible to rank them along (ii) a cognitive parameter that serves as yardstick for (iii) the characterization of different types of reference (the communicative tasks)

(iv) a way of reconciling the various 'real world' or text conditions that ultimately determine, or reflect, the attendability of a referent, i.e. what specific communicative task we face. The question as to what form this evaluation takes is far from academic: languages do not always pattern alike with respect to the 'same' categories. For instance, in the Portuguese data, singular resembles animate in favouring the stronger referential device (Mollica 1981)," and an excellent case for this skewing can be made on grounds of prominence: a single entity is more noteworthy (easier to focus on) than an aggregate. But in Spanish we observe exactly the reverse situation: there it is plural that favours the use of el, the stronger device (Garcia in press)—the reason being, it could be argued, that a plurality is 'more', and hence inherently more prominent, than a single entity. How then is number to be viewed in terms of intrinsic prominence—if in fact it is relevant to prominence at all? Or consider the specificity of the antecedent: in Portuguese, 'specific' referents favour die weaker device. This makes sense in terms of topicality, discourse-givenness of the referent, etc.; distributionally, then, specificity of the antecedent parallels (intrinsic) non-animacy. But in Spanish non-specificity patterns with non-animacy in favouring the weaker sf— 'because', perhaps, it fails to deserve (or permit) explicit pointing out. Either we have erred in our identification of the relevant cognitive parameter (perhaps, with number, it is not inherent salience that is relevant, but heterogeneity vs. homogeneity of reference), or we have failed to identify the relevant textual reflexes of the cognitive categories (the skewing for number may be epiphenomenal to some other skewing), or languages (cultures?) differ as to the relevant cognitive parameters, and/or as to what makes something noticeable or noteworthy. The problem is further complicated by the fact that, under all circumstances,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

but also

E. C.Garcia 315

T a b l e 2 Resolution of conflicting claims on si: contextual givenness, vs. inherent nonprominence (% of si over si + // references) Author Cortazar

Mallea

Sabato

M. Estrada

Human referent, contextually given (subject, no competition from object of same type)

32

87

88

95

Inanimate referent, contextually less given (subject, with competition from object of same type)

75

80

75

95

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

there is but one reference made. What do speakers (or languages) do, for instance, when one and the same referent qualifies for strong reference because of its intrinsic prominence but—because it is contextually salient—also qualifies on contextual grounds for weak reference? There is no reason why the problem should have a single solution: different speakers or languages may strike a different balance when confronted with the same conflict of motivations.12 A diversity of solutions is in fact suggested by the data in Table 2, where we contrast (for the same Spanish corpus as Table 1) two conflicting combinations, namely contextual givenness of a noteworthy referent, vs. contextual nonsalience of an inherently non-prominent referent, under identical syntactic conditions (coreference with the subject of the verb). We find that for two authors (Mallea and Sabato) contextual notkeability overrides in determining recourse to si, while for Cortazar it is inherent noteworthiness which primarily steers the choice between si and el. One fourth author—Martinez Estradashows identical percentages of si for both combinations. There is no reason why different solutions to a conflict of motivations should not be reconcilable with the same unitary analysis of the linguistic devices, or with the same, single mapping of the devices on to the cognitive continuum. The variety in use is, in fact, more easily accounted for if we do assume that speakers learn the same thing in connection with the competing expressive alternatives, since diversity in use is easily relatable to the fact that different individuals inevitably have different values and priorities.13 The question that critically remains, however, is exacdy how such divergence in values surfaces as different ways of gauging the appropriateness of the same two expressive options. It is here that linguists critically need the help of cognitive psychologists, and that psychologists, in turn, may benefit from a different sort of data than the one usually obtained from psycholinguistic experimentation.

316 A Psycho-Linguistic Crossroads: Frequency of Use

8 CONCLUSIONS

is 'yes', what makes it an adequate explanation ? And if the answer is 'no', whatform would an adequate explanation take ?

By trying to make the connection between linguistic categories and cognitive processes we promptly discover that our understanding of the communicative problems, of their interaction, and of their reflection in specific text-properties is woefully inadequate, even in an area as well studied as pronominal reference. We at least have practically no insight into what factors weigh, how, and—especially—why, on the speaker's ultimate assessment of the 'attendability' of a referent. But not till questions of this sort have been answered, and linguistic analysis rests on firmer psychological foundations, will it be possible to solve the— purely analytic—issue that has mainly exercised linguists so far (cf. Fox's (1986) discussion of 'abstractionist' vs. 'conventionalist' analyses), namely, the identification of the different uses of a same form. It is to be hoped that the kind of data linguists can easily come by—more or less detailed qualitative/quantitative descriptions of language use—may prove useful to psychologists interested in the processes underlying speech production and that, in turn, by specifying 'what counts' for language users, psychologists may help linguists explain the striking skewings that are observed as soon as one begins to count. ERICA C. GARCIA Faculteit derLetleren Van Wijkplaats 3 NL - 2300 RA Leiden The Netherlands

* I am grateful to P. Bosch, T. Hoekstra, R. de Jonge, R. S. Kirsncr, F. v. Puttc and two anonymous referees for this journal for helpful criticism of earlier versions of this paper.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

We have tried to argue that linguistic analysis has little value or interest unless it helps us to understand, in a psychologically plausible way, the exploitation of a particular linguistic device for a particular communicative end. We must know, for instance, what makes a less specific form (like si) more appropriate than el for referring to contextually salient but inherently not so noteworthy referents. Linguists often say 'it makes sense that... [there follows the distributional fact to be explained]': what exactly is meant by this 'making sense'? Is the difference in informational load (gender/number marking in el, invariableness of si) an adequate explanation of the different use of these forms? If the answer

E. C.Garcia 317

NOTES they coincides with its referent. There is no previous mention of, or allusion to, 'that kind of snake' in the preceding context. This notion arises at the use of they, not earlier: the use of the pronoun thus creates its own antecedent or referent. 'Normal' pronoun use is just as 'referent creating' as the more infrequent cases illustrated by the sentence above. The fact that in 'normal' use the referent invoked by the pronoun happens to be some entity already present in speaker's or hearer's consciousness is immaterial.

2 What is left unspecified in her study, however, as well as in others of this type, are the (independent) criteria that motivate the subsumption of specific individual use-tokens under a particular type. Not surprisingly, the reader does not know, at the end of such a study, whether the use of forms has been described, or whether the value of the form has been stated. 3 This is understandable, since

5 It is not the purpose of this paper to discuss the relation of pronominal reference to discourse (and especially topic and focus) structure, though this is obviously relevant to how much 'attention' a referent enjoys in a given context. We are concerned with the question of how linguistic devices are matched to cognitive/communicative tasks, and how this is reflected in their frequency distribution. 6 The absence of an argument is more likely to be noticed in the case of central syntactic roles (subject, direct object) than for more peripheral functions: the antecedent is consequently more easily 'read into' the relative clause directly via the (invariable) relative pronoun in the case of a central role than of a peripheral one. It is for this reason that peripheral roles have greater need of the extra 'boosting' of the antecedent. 7 That humans should rate as important is not surprising, when we consider that the speakers who make the choice between si and el are themselves human beings. The texts we are confronted with (so far) are the output of (notoriously anthropo/ egocentric, perhaps even selfish) human beings: they are consequently likely to betray a human-centric view of the world, and of what matters in it.

(a) identification of the intended referent (in the context) may involve different degrees of attention; (b) different sorts of referent may draw attention to greater/lesser degree; (c) different entities in a given context differ as to the degree to which/ whether they are in the focus of attention. What we mean more specifically by 'attention' will become (intuitively) clear, we trust, by our discussion of specific data/phenomena in Sections 4-7 below. 4 Pronouns obviously refer in their own right, and the fact that the referent they evoke often happens to coincide with those evoked by noun-phrases in earlier discourse is natural, but not essential. In 'Don't fool with that snake: I think they're poisonous", the 'antecedent' of

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

1 The distribution of pronouns, for example, has generally been seen by most linguists—especially those aiming at formalized and explicit grammars—as a problem of rules, the goal being to state the conditions under which different referential devices get into the right places and stay out of the wrong ones (cf. Reinhart 1983; the literature is enormous). Within such an approach the rule is the characterization of the device, but—in our opinion—it gives little insight into the facts of usage, especially into the quantitative ones.

318 A Psycho-Linguistic Crossroads: Frequency of Use hearer's perspective), especially when the interdependence of the two functions (ibid.: 369-70) is taken into account. 11 Since Mollica does not cross-classify her data, it is impossible to tell whether this is a genuine skewing, or merely the result of an imbalance in her corpus. But the example is valid enough for the general point we are trying to make. 12 The possibility of arriving at different solutions is, naturally, of crucial importance for language change. 13 Moreover, the position that speakers agree on the nature of the linguistic devices, but differ as to the uses to which they may be put, is the only one really compatible with new applications of linguistic forms—which is, in essence, what all language use consists of. Note, moreover, that it only makes sense to speak of idiosyncratic differences in the evaluation of something when that something is, in fact, the same. You may prefer chocolate to vanilla ice-cream, while my taste may be just the opposite: but for this disagreement to be possible at all we must agree as to what we are talking about, namely, our identification of both chocolate and vanilla as specific flavours.

REFERENCES Bock, J. K. (1982), 'Toward a cognitive same, similar, or different', in R. J.Jarvella & W. Klein (eds), Speech, Place and Action, J. psychology of syntax: information proWiley, Chichester. cessing contributions to sentence formulation', Psychological Review, 89: Fox, B. A. (1986), 'Local patterns and general principles in cognitive processes: anaphora I-+7in written and conversational English', Brown, G. & G. Yule (1983), Discourse Text, 6: 25-51. Analysis, Cambridge University Press, Cambridge. Garcia, E. C. (1983), 'Context-dependence of language and linguistic analysis', in F. Chomsky, N. (1957), Syntactic Structures, Klein-Andreu (ed.), Discourse Perspectives on Mouton, The Hague. Chomsky, N. (1966), Cartesian Linguistics, Syntax, Academic Press, New York. Harper & Row, New York. Garcia, E. C. (1986), 'Reflexivity turned back on itself, in O. A. Jaeggli & C. SilvaDu Bois, J. W. (1987), 'The discourse basis of Corvalan (eds), Studies in Romance Linguisergativity', Language, 63: 805-55. tics, Foris, Dordrecht. EbJich, K. (1982), 'Anaphora and Deixis:

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

8 The problem under discussion is reminiscent of the conflicting accounts (Bock 1982: 36) given for certain experimental data, which are open to interpretation under either contextual or perceptual salience rationales. Bock (1982: 37) summarizes the paradox by saying 'given-new ordering thus appears to directly contradict the basis of a focus-of-attention explanation'. 9 What may be involved here is the degree of (in)dependence of a figure and the ground against which it appears. For example: there are some people who are intrinsically important to us (family, friends) whom we recognize whenever we see them, regardless of place, rime, or occasion. Other individuals for whom we care less (perhaps our butcher, or the bus driver) we recognize and identify only under familiar circumstances, i.e. in the context where we expect to meet with them: these do not exist independently of a particular setting. 10 The two 'speaker's tasks' we have described seem to correspond to Marslen-Wilson, Levy & Tyler's (1982: 367) distinction between 'location' and 'predication' functions (viewed from the

E.C.Garcia 319

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Garcia, E. C. (in press), 'Grasping the nettle: Naro, A. (1980), 'Review of D. Sankoff (ed.), 'Linguistic Variation', Language, 56: variation as the proof of invariance', in L. 158-70. Waugh & S. Rudy (eds), New Vistas in Grammar: Invariance and Variation, John Nooteboom, S., T. Kruyt & J. Terken (1980), Benjamins, Amsterdam. 'What speakers and listeners do with pitch accents. Some explorations', Manuscript Givon, T. (1983), 'Introduction', in T. Givon no. 397, IPO, Eindhoven. (ed.), Topic Continuity in Discourse, John Benjamins, Amsterdam. Reinhart, T. (1983), Anaphora and Semantic Marslen-Wilson, W., E. Levy & L. K. Tyler Interpretation, Croom Helm, London. (1982), 'Producing interpretable dis- Tomlin, R. S. (1987), 'Linguistic reflections of course: the establishment and maintencognitive events', in R. S. Tomlin (ed.), ance of reference', in R.J. Jarvella & W. Coherence and Grounding in Discourse, John Klein (eds), Speech, Place and Action, J. Benjamins, Amsterdam. Wiley, Chichester. Yule, G. (1982), 'Interpreting anaphora without identifying reference', Journal of Mollica, Ma. C. (1981), 'La "copie" dans les Semantics, 1: 315-22. constructions relatives en portugais', in D. Sankoff & H. Cedergren (eds), Variation Omnibus, Linguistic Research, Edmonton.