4
Cog&ion, I3 (1983) 223-261
Thematic processingin sentencesand texts* DAVID
J. TOWNSEND
Mop tclar’r StateCollege
Since many of the semantic relations that exist between clauses in complex sentences also exist between sentences in texts, a description of complex sentence processing should generalize to sentence processing irl texts. This paper explores the usefulness of such a generalization fer liescribing the online processing of sentences in narratives. Four experiments on the processing and retention of isoiated complex sentences showed that connectives that signal a disruption in the causal and temporal order of propositions influence on-line processing and recall. Four additional experiments on the processing and retention of sentences in narratives showed that causal/ temporal relations within complex sentences and between sentences affect sentence reading time, and the immediate and long term memory organization of propositions. The results sugge ’ partially independent processing systems for both isolated complex sentences and sentences in texts. One set of processes operates on superficial information to obtain a literal propositional representation; another operates on propositional information to obtain a thematic representation of the sentence or text. In general, fuctors that obscure the thematic relevance of a clause or sentence also indirectly affect the processing o.f propositional meaning.
Introduction Most sentences appear in a linguistic context consisting of two or more sentences that are related in some way. In the case of texts, these relations partly constitute the ‘theme’ of the text. This paper shows how we can incorporate a partial description of thematic information into a processing
*This research was supported by National Science Foundation Grant BNS-7817509 and by a Faculty Research Grant from Montclair State College. The author is grateful to M. Cloitre, C. Cummings, N. Flies&man, N. Kelley, J. Kutch, V. Larsen, F. Testa, and A. Trauth for their assistance in data collection, and to T. G. Bever and an anonymous reviewer for critical comments on an early dratt of this paper. Reprint requests should be sent to D. 3. Townsend, Department of Psychology, Montclair State College, Upper Montclair, NJ. 07043, U.S.A.
0010-0277/83~$12.20
0 Elsevier Sequoia/Printed
in The Netherlands
224
D. J. Townsend
model, how the text processor obtains thematic information on-line, and how thematic processes interact with individual sentence-processing mechanisms. Sentences in a text are often organized along some pattern which is related to the theme of the text. In the narrative pattern causal and previous events occur earlier in text than do events which are effects or which occurred later in time, as in the passage in Table 1. This pattern is implicit in the structure of story grammars (e.g., Mandler and Johnson, 1977); it is related to narrative themes such as a moral, conclusion, or a new state of awareness in the protagonist, that grow out of the events that had been presented in the narrative (e.g., Brooks and Warren, 1970). This suggests that apprehension of causal and temporal relations between sentences is part of the thematic processing of narratives. Although causal and temporal relations are obviously not present in all texts, many texts contain other intersentence relations and patterns of sentence ordering that are related to their predominant theme (cf., Brooks and Warren, 1970). Table 1.
Example of a narrativetext (adapted from Schank and Abelson, I9 77) 1. 2. 3. 4. 5. 6. 7. 8. 9.
John was thirsty. He opened a can of beer and went into the den. There he saw a new chair. He sat down in it. Suddenly the chair tilted over and John fell on the floor. His beer spilled all over the chair. His wife heard the noise and ran into the den. She saw that her new chair was ruined. She became very angry. _.-
Semantic relations between sentences may be derived from expectations based on -[he pattern of sentence-ordering that exists earlier in the text or in similar texts, and from connectives. Similar inter-sentence relations may occur for both clauses and sentences that are adjacent in text. Table 2 shows some clausal and sentential connectives that signal similar causal and temporal relations (see Brooks and Warren, 1970; Carpenter and Just, 1977; Dik, 19168;van Dijk, 1977; Jespersen, 1940, 1964; Quirk ef al., 1972). The similarities in semantic relations within complex sentences and between sentences in texts suggest similar processes of integrating the clauses of complex sentences and the sentences of texts. If those integration processes are similar, a theory about the effect of inter-clause semantic relations on clausal processing will also describe the effect of inter-sentence semantic relations on sentence processing in texts.
Thematic processing in sentences and texts
Table 2.
225
Some meanings qf connectives 1, Cause: because, if, for, since Because her new chair had been ruined, she was angry. 2. + Prior: since, after, Previously, Beforehand After hearing a loud noise, she ran into the den. 3. Simultaneous: as, when, while, and, Meanwhile, Simultaneously When the chair tilted over, beer spilled on it. 4. -Prior: bezfore,until, and, Afterward, Subwquently John opened a can of leer before he went ,nto the den. 5. Effect: so, and, Consequently, Therefore John was thirsty so he opened a can of beer. 6. Denial of Expected Causal Role: although, while Although her new cha!r was ruined, she wasn’t angry. 7. Unexpected Effect: but, whereas, However, Nevertheless, Yet Her new chair was ruined, but she wasn’t angry.
However, sentences and texts differ in one crucial respect: the sentence possesses an internal structure that the text lacks. Rules that govern the arrangement of words within a sentence do not apply to the arrangement of sentences within a text. Thus, there cannot be complete overlap in the knowledge that is used for processing sentences and texts. These properties of sentences and texts suggest that text processing consists of two partially separate processes: thematic processing and propositional processing. Thematic processing determines the relation of incoming propositions to previously processed propositions by applying the meanings of connectives, by applying patterns that had been established earlier in the text and in similar texts, etc. Thematic processing integrates just-apprehended propositions into a thematic representation of the text; for example, in the case of narratives, it determines the causal/temporal relation5 of incoming propositions to previously processed propositions. Propositional processing obtains the literal meaning of clauses by applying lexical 2nd syntactic knowledge to the words in the clause. Propositional processing retains hypotheses about syntactic structure and literal meaning until the propo:Gtion has been integrated with the preceding text. Since sentences 1-5 of Table 2 preserve the causal/temporal patterns typical of narratives, they could all be substituted for correspondmg sentences in the narrative in Table 1 without complicating thematic processing. For example, if sentence 1 in Table 2 replaced sentences 8 and 9’ of the narrative in Table 1, thematic processing determines that the event in the initial clause is the cause for the event in the final clause, just as it would if the original sentences 8 and 9 were in the story.
226
D. J. Townsend
However, thematic processing in narratives is often complicated by descriptive portions that do not follow a causal or temporal pattern, by flashbacks, by the omission of an event which is needed in order to complete the underlying ‘causal chain’ (see Schank and Abelson, 1977), or by a denial that some expected event occurred. Such disruptions of the predominant pattern of narratives require time-consuming inferences in order to ap prehenei their relation to the underlying theme. Sentence 6 in Table 2 does not preserve the causal pattern. The use of although denies an expected cause-effect relation (see Dakin, 1970); in sentence 6 of Table 2 althotcgh denies the expectation that “her new chair was ruined” caused “she was angry”. On the basis of the disruption in causal ordering alone we would expect a disruption of thematic processing. But full apprehension of the thematic meaning of sentence 6 also requires a determination of what the story-teller had expected to be the resuit of the presumed cause in the although clause, for example, whether the story-teller had expected “her new chair was ruined” to cause John’s wife to be angry, to entail that the carpet was also ruined, or to cause them to buy a new chair. Thematic processing cannot determine which of these the story-teller expected without knowing what unexpected event actually did occur; only when it is known that “she wasn’t angry” can thematic processing determine that the story-teller expected “her new chair was ruined” to cause John’s wife to be angry. Since thematic processing cannot determine the thematic relevance of the although clause on the basis of the propositional information contained in it, its superficial information is retained until further information is obtained that allows it to be thematically integrated. As propositional processing provides information from the final clause about what unexpected effect . .dd occurred, thematic processing can determine the specific causal relation that the story-teller had expected. In this analysis of text structure thematic integration is an organizing principle which governs text processing, according to the following hypotheses: A. If thematic integration of a proposition requires further information that is not explicitly stated in the clause, its on-line processing will be less complete, propositional hypotheses based on it will be relatively less accessible, and superficial information from it will be relatively more ;acce&ble. B. If thematic integration of a proposition is not complete at the clause boundary, the availability of superficial information from the clause and the thematic demand of obtaining further information will influence propositional processing of the next clause.
Thematic processing in sentences and texts
227
thematic pattern is utilized as an organizing principle ng of a text, a proposition which is the basis for that pattern will be more accessible after a sentence containing it has been comprehended. (3. If the underlyin
B. If clauses or sentences are ordered in a way that disrupts on-line thematic integration, integration time for the clauses or sentences will be increased and recall will be poorer.
E. If a just-comprehended clause or sentence provides information that is relevant for thematic integration of the next clause, comprehension time for the next clause will be reduced. Townsend and Bever (1978) provide evidence for Hypothesis A. In that study, subjects heard a fragment of an initial subordinate clause introduceri by one of five different connectives. Immediately after hearing the fragment, one group of subjects heard a probe word, which they classified as being in the fragment or not; this word recognition task measured the listener’s online accessibility to the superficial form of the clausal fragment. Another group of subjects read a verb-object phrase, which they classified as quickly as possible as being either consistent or inconsistent with the meaning of the fragment; this synonomy judgemen t task measured the listener’s on-line accessibility to the meaning of the interrupted clause. Figure 1 shows that as the connective signals more explicitly that the clause contains a causal event (cf. Table 2), access to meaning !s better but access to superficial form is poorer. The current experiments provide additional support for Hypothesis A and examine the validity of Hypotheses B-E for processing sentences in isolation and in texts. The experiments demonstrate that the thematic completeness of an isolated initial clause affects on-line accessibility to alternate representations of it (Hypothesis A). A synonomy judgment task shows that for a thematically incomplete clause, there is greater accessibility to its meaning after the entire sentence has been heard than there is during listening to the clause ; for a thematically complete clause, there is no such difference (Experiment I). Second, a word-naming task (cf., Tyler and Marslen-Wilson, 1977) shows that for a thematically complete initial clause fragment, a continuation word is read more quickly if it is consistent with syntactic information that occurs early in the clause; for a thematically incomplete initial clause fragment, there is no effect of the consistency of the continuation word with syntactic information on time to read the continuation word (Experiment III). These studies indicate that the thematic role of an initial clause affects its on-line representation and the on-line utilization of its syntactic information.
228
D, J. Townsend
Figure 1.
Relative accessibility to words versus meaning in the on-line processing of initial &uses (Adapted from Townsend and Bever, I9 78).
If
Sk
&EN
UHiLE
THOkH
The experiments show that a thematically incomplete context clause influences the propositional processing of the next clause, but that a thematically complete clause does not (Hypothesis B). A question-answering task shows that a thematically incomplete context clause disrupts the typical effect of a within clause syntactic variable (active versus passive form) in a final main clause, but a thematically complete context clause does not (Experiment II). Second, the word-naming taslk shows that unambiguous syntactic information in a final main clause has a greater effect on reading times for a continuation word when the clause is preceded by a themati?:L 7 incomplete clause than when it is preceded by a thematically com$:te clause (Experiment III). Third, the word-naming task shows that readkg times for a word that continues an ambiguous phrase are more strongly affected by specific lexicaf information from the preceding clause when the preceding clause is thematically incomplete than when it is thematically complete (Experiment III). These studies indicate that propositional processingis independent of context only when the context can be thematically integrated independently of what follows; when the meaning of the context has not been thematically integrated, the retention of its superficial
l’ttematic processing in sentences and texts
229
representation and the search for further information in order to them& tally integrate it increases the amount of attention allocated to the processing of the next clause. A second synonomy judgment task shows that the underlying thematic pattern in a narrative is the basis for organizing the propositions of a sentence that has just been heard (Hypothesis C) by showing that the meaning of causal events is more accessible than the meaning of noncausal events (Experiment IV). The exl)eriments show that the meanings of connectives and the ordering of events in sentences and texts influence sentence processing time and recall (Hypothesis D). It is shown that the recall of both isolated sentences containing a causal or temporal connective (Expenment V) and of nzm;;;tives without connectives (Experiment VI) is better when the order of presentation of the events matches a causal or temporal order. Second, it is shown that a match between the order of presentation of events in a narrative and the underlying temporal pattern reduces sentence reading time (Experiment VI). Third, it is shown that less time is needed for constructing a sentence that continues a narrative when the connective introducing the most recentlyread sentence signals more explicitly that the sentence follows causally from the preceding sentence in the text (Experiment VII), and, for narratives ending in a complex sentence containing a causal connective, when the causal event appears earlier in the sentence (Experiment VIII). The:se studies on sentence reading time, continuation time, and recall indicate that thematic integration occurs during the processing of sentences in text and that the ease of on-line thematic integration influences Ilater recall. The continuation time task also shows that the meaning of a context clause or sentence influences thematic integration (Hypothesis E). Continuation time following a sentence containing a&tzough is faster when the unexpected effect appears earlier in the sentence (Experiment VIII); this indicates that knowing the actual unexpected effect as the presumed cause is being read facilitates thematic integration of a denial of an expected causeeffect relation, This effect of context on thematic iintegration is similar whether the relevant context is the initial clause of a sentence or a preceding sentence: a context sentence that paraphrases the unexpected effect in the although sentence also reduces continuation time following an although sentence. Even stronger evidence for Hypothesis E is tlhat the meaning of a context sentence interacts with expectations about thematic order to inrluence the inter-clause relations that are assigned to a complex sentence that has two or more potential thematic roles. By showing that similar effects occur whether the materials consist of isolated sentences or texts, these experiments indicate that thematic integration
230
D. J. Townsend
is an organizing principle that influences the processing of both isolated sentences and sentences in text. Nevertheless, propositional processing and thematic integration are partially distinct processes: propositional processing utilizes surface information that has not been thematically integrated whereas thematic integration utilizes surface cues, expectations of thematic ordering, and propositions that have already been thematically integrated. Propositional and thematic processing are not distinct in that they rompete for the same processing resources: the demands of thematic processing can influence the allocation of attention for propositional processing. Experiment I
The frrst experiment examines performance on a small set of items in two tasks. It obtains two types of data to determine the nature of the on-line representations of a clause that deviates from the causal/temporal pattern. The frost type of data concerns the accessibility to literal meaning near the end of the clause boundary and after the sentence boundary. Subjects heard either a fragment of a sentence ending before the last word of an initial subordinate clause, or an entire sentence that completed the fragment. After hearing a fragment or sentence, the subject determined whether a visually presented verb-object phrase was consistent in meaning with what s/he had heard. In the critical cases, the verb-object phrase paraphrased the verbobject phrase of the initial clause of the fragment or sentence. If the on-line representation of an initial noncausal (whi/e) clause is not fully semantic, response times for synonomy judgment will be faster after the whole sentence has been heard than they are while hearing the clause; if the online representation of an initial causal (since) clause is fully semantic, there will be no difference in response times near the end of the clause and after the whole sentence has been heard. The second type of data concerns the search rate for a target word in a clausal fragment. Subjects heard the same fragments as in the previous task, but now determined whether a probe word had occurred in the fragment. I assume that subjects have retained more superficial representations of the clausal fragment, including information about word order, to the extent that their response times are faster for targets that had occurred earlier in the fragment. If the on-line representation of an initial while clause is more likely to contain syntactic information, word recognition times will be faster for targets occurring earlier in the fragment. If the on-line representation of an initial while clause is more abstract than an unanalyzed, ordered sequence of the words in the fragment, the rate of searching the mental representation
Thematic processing in sentewes and texts
231
t will be faster than the rate of subvocal speech. The two types ther suggest that the on-line representation of thematically incomplete clauses is a mixture of superficial and semantic information, but that of a thematically complete clause is primarily semantic,,
Method Sixty-four right-handed undergraduates at Montclair State College and Columbia University served individually as subjects. Thirty-two participated in the post-sentence meaning task, sixteen in the fragment meaning task, and sixteen in the fragment word task. The sex of subjects was balanced across various conditions of these experiments. Subjects listened to tapes recorded by a male speaker with normal intonation through the right channel of a set of headi?hones. Subjects in the two fragment tasks (word versus meaning) heard all but the last word of the critical initial subordinate clauses, and subjects in the post-sentence meaning task heard complete two-clause sentences that continued these critical initial clause fragments. At the end of the fragment (or sentence) all subjects heard a 50 msec, 500 Hz tone. One-third second later, subjects in tne word task heard a single word which they were to verbally classify as having occurred in the fragment or as having not occurred in the fragment. One-third set after hearing the tone, subjects in the meaning tasks saw a slide containing a. 2-4 word predicate phrase which they were to verbally classify as being either consistent or inconsistent with the meaning of any part of the fragment or sentence they had heard. Pretesting showed that test phrases for positive trials were judged similar in meaning to the initial clause of the sentence, whereas phrases for negative trials were judged dissimilar in meaning to either clause. There were 7 critical clauses which were identical across the three tasks; 4 clauses contained while and 3 contained since (see Table 3). In the two fragment tasks, these were embedded in lis s of 34 fragments, and in the postsentence task, the lists contained 41 complete sentences. The remaining fragments in each task were used to prevent detection of which clause would be tested (in half the positive trials, the final clause was interrupted before the last word, and was synonomous with the test phrase or contained the target), to prevent bias toward ‘yes’ responses (in approximately half the trials the test phrase was not synonomous with either clause or the sentence did not contain the target), and to control for the grammatical class of target words. The critical clauses contained either 10 or 11 one-syllable words, one of which appeared in one of two positions without changing the meaning of
232
D. J. Townsend
Table 3.
Mater@ used in Experiment I Since Tom has poured the red wine and served the iced/tea, the guests are waiting to be seated for dinner. b. Since Tom has served the red wine and poured the iced.. .
la.
2a. Since Bill swept, cleaned, and mopped the floor of the/porch, Linda carriedout the garbageand mowed the lawn. b. Since Bill cleaned, mopped, and swept the floor of the . . . 3a. Since Pete will sum come home from his job at/school, Ann will leave to do the weekly grocery shopping. b. Since Pete will come home ,~n from his job at . . . 4a. While the kizrgecouch or the small chair must be/moved, the rest of the furniture can remain where it is. b. While the small couch or the large chair must be . . . 5a. While teams play there a lot at the end of/fall, the brand new field is not used at all in the spring. b. While teams play a lot there at the end of . . . 6a. While Bob did put down some new tiles in the/hall, no work needed to be done in the apartment. b. While Bob did put some new tiles down in the . . . ?a. While the French chef boiled eggs, fried ham, and sliced/bread, the butler had little to do but stand around. b. While the French chef fried ham, boiled eggs, and sliced . . . Note: Target words are italicized. Slashes mark the end of fragments.
the clause. On the average, 3 syllables (the range was 2-5) separated the early and late positions of the moveable word. In the word task, target position varied within subjects and within items; in all three tasks, the conjunction introducing a clause varied *withinsubjects. !fkults and discussion
Errors occurred at a rate of 3.5% in the fragment meaning task, 4.7% in the post-sentence meaning task, and 3.5% in the fragment word task. Response times for the error trials were replaced by the mean response time for correct responses from the appropriate cell. Figure 2 shows the mean response times for correct responses on the two meaning tasks. In reporting statistics that indicate the generality of effects to the larger population of subjects and materials, I follow guidelines provided by Clark (1973). Since a given set of words in Experiment I was presented with only one connective across subjects, words by connectives were not nested within
Thematic processing iIt sentences and texts
233
to propositionalmeaning of initial clauses near the clause boundaryand after the sentence boundary.
Figure 2. Accessibility 1200 I150
2
1100
I
..
.
. -
SINCE
..a....
WHILE
BEFORE CLAUSE BOUNDARY
.
Af-TER SENTENCE BOUNDARY
subjects. Hence, statistics based on subjects as a random effect do not treat materials as a random effect. For experiments with this type of design, I report both Fi, treating subjects as a random effect, and Fa, treating materials as a random effect. In other experiments reported here, however, subjects were tested on a given set of words with different combinations of independent variables, and a single analysis of variance treats both subjects and materials as random effects (see Clark, 1973, p. 348); in these cases, I report only F statistics. Accessibility to the meaning of an initial w/zile clause was 220 msec faster after the whole sentence was heard than it was during listening to the while clause, but the corresponding increase in accessibility to the meaning of an initial since clause was 28 rnsec. The increase in accessibility to meaning after the whole sentence was heard was larger for initial while clauses than for initial since clauses, & (1,46) = 6.15, p < 0.05, F2 (1,5) = 4.29, p < 0.10, indicating that the on-line representation of an initial while clause is not fully semantic. The word recognition results support this conclusion. Table 4 shows that target position had opposite effects on word recognition times in clauses beginning with the two connectives: recognition times were slower for late targets than for early targets in initial while clauses, but the opposite was found in initial since clauses, F1 (f,30) = 7.08, p < 0.05, Fz (1,s) = 5.43, p < 0.10. The fact that recognition times increased with target position in while clauses suggests that subjects searched their representation of the fragment
234
D. J. Townsend
Table 4.
Meun respol;ose times (msec) to targets in initial since and while fragments Target location ~~~~~~~ S&t?
while
--
Early
Late
1038 954
1025
947
one ‘word’ at a time from ‘left to right’. Such a search procedure follows from the assumption that the subjects had retained word order information from these fragments. The difference between response times to early and late targets in while clauses was 71 msec,. Since the early and late targets were separated on the average by three syllables, the serial search rate was about 24 msec per syllable, which is much faster thzn the fastest rate of subvacal speech (about 150 msec per syllable]). This inr”ifs&esthat the on-line representation of an initial while clause is not merely an unanalyzed, subvocal representation of speech. Since the :i;ynonomyjudgment task showed that t.he on-line semantic representation of an initial whi!e clause is not complete, and the word recognition task showed that the on-line representation is not a literal representation. of speech, the on-line representation for initial while is probably a mixture of superficial and semantic information. The text processor retalins the superficial form of the initial while clause and hypetheses about its meaning until the unexpected effect stated in the final clause allows it to determine what effect the speaker had expected to occur based on the event in the while clause. Since the thexnatic meaning of an initial since clause, on the other hand, does not depend on information in the final clause, there is no need for easy access to its superficial representation. Experiment II
Experiment II demonstrates that the thematic integration associated with a noncausal initial clause affects propositional processing in the next clause. It shows that the typical effects of a within-clause syntactic variable-active versus passive form -do not occur for a main clause following a noncausal clause, but do occur for a main clause following a causal clause. Subjects listened to two-clause sentences containing initial clauses introduo& by either since or thoug& which convey causal/temporal and advernative meanings respectively. The final main clauses were either active or
Thematic processif
Table 5.
g
in sentences and texts
235
Materials used in Experiment I1 la. b. 2a. b. 3a. b. 4a. b. Sa. b. 6a. b. 7a. b. 8a. b.
Since the boy moved the furniture, the maid swept the carpet. . . . the carpet was swept by the maid. Since the state budget has been much too large, the governor signed the tax bill. . . . the tax bill was signed by the governor. Since visitors were not welcome, the dog chased the mailman. ..a the mailman was chased bv the dog. Since Matgaret got ptomaine poisoning, mother bought a new dishwasher. . . . a new dishwasher was bought by mother. Though the soldiers were not in uniform, the drill sergeant yelled the commands. . . . the commands were yelled by the drill sergeant. Though the jockey lost his balance, the horse jumped the hurdle. . . . the hurdle was jumped by the horse. Though nothing had been stolen, the cop arrested tL bandit. . . . the bandit was arrested by the cop. Though nobody had spotted any whales, the young sailor dropped the anchor. . . . the anchor was dropped by the young sailor.
passive non-reversible clauses. After hearing the sentence, subjects read and answered an active question. If the thematic meaning of an initial clause has no effect on within-clause processing of thlz final clause, questions about active final main clauses will be answered more quickly than questions about passive final main clauses (cf., Wright, 1972). On the other hand, if a thematically incomplete clause affects processing of the next clause, the typical effect of syntactic form will not be obtained in the next clause.
Twenty-four right-handed Montclair State College undergraduates individually heard one of eight lists of 41 sentences. In each list there were 4 critical sentences kntroduced by since and 4 by though (see Table 5). Half of each type of critical sentence was paired with a question about the logical object, and half was paired with a question about the logical subject. The form of the question was counterbalanced with the length of the final clause, which was either 6-7 or 8-9 syllables long in active form. The active or passive form of each critical final main clause varied across lists. Hence, active versus passive form varied within subjects and within items; conjunction varied within subjects. Each list also contained 9 practice sentences and 24 filler sentences to balance for tests of other combinations of connective, clause position, clause type, length, and question type. The questions were presented by slide and the procedure was otherwise identical to that used in the post-sentence meaning task of Experiment I.
236
D. J. Townsend
Results and discussion
Errors occurred on 2.1% of the critical trials, and were excluded from staltistical analysis of response, time data. Table 6 shows the mean response times for correct responses in critical trials. Response times were faster for active than for passive fmal main clauses preceded by a since clause, but not for final main clauses preceded by a rho;xgh clause, F1 (1 ,I 6) = 7.92, p C 0.05,Fz (1,4) = 12.6, p < 0.05.The effect of within,-clause syntactic Cohnplexity depends on the connective introducing the context clause, i.e., on the semantic relationship between the two clauses. The text processor processes a final main clause that is preceded by a since clause independently of the context clause, but this is not the case for a final main clause that is preceded by a though clause. One interpretation of these results is that propositional processing of a final clause preceded by an adversativr: clause occurs relatively quickly in order to obtain information that can be used to thematically integrate the adversative clause; this interpretation was examined further in Experiment III. Table 6.
Mean response times (msec) to questions about depending on the context clause
active ver iUS passike clauses
.-a Context
Active final maiu
Passivefinal main
SiXICe Though
1417 1615
1561 1427
Experiment III
Experiment III demonstrates that the on-line accessibility to syntactic hypotheses in a clause depends on an interaction of the causal meaning of the introductory connective and the position of the clause in the sentence, It also investigates t.he role of superficial and semantic information from alternative thematic contexts on the formation of hypotheses about the structure of an ambiguous phrase. The paradigm was a word-naming task (Tyler and Marslen-Wilson, 1977), in which subjects heard a sentence fragment consisting of a subordinate clause, followed by an ambiguous phrase of the form ‘verb+ing noun’, which could either be a progressive adjectival (as in [ 1I ) or a gerund (as in [ 21). In Tyler and Marslen-Wilson (1977) the context clause was consistent with either the adjectival (e.g., [3] ) or with the gerund (e.g., [4] ) interpretation.
Thematic processing in sentences and texts
(1)
.,. dying flowers
(2)
watering flowers . . . dead ones are ugly, growing flowers . . . Although looking after plants is easy, growing flowers
(3) (4)
237
...
...
...
After the critical phrase, a slide containing is or are appeared; the subject’s task was to read aloud the word on the slide. Tyler and Marlsen-Wilson found that target reading times were faster for continuations that were appropriate for the meaning of the context clause. They interpreted their results to show that people use contextual semantic information to make online decisions about the structure of an ambiguous phrase. Note, however, that in (3) and (4) there are morphologically marked number cues (e.g., is versus are) that agree with the direction of the semantic bias of the context. In Tyler and Marslen-Wilson’s materials, 90% of all numerically distinct cues in the context agreed with the direction of its semantic bias. This suggests that listeners may have been using superficial cues rather than semantic information to guide syntactic processing. Experiment I indicated that such superficial cues are more available for adversative clauses (like whi and although) than for causal clauses (like since and if). In Experimen II there were three types of critical trials. First, there were initial subordinate clause fragments ending with an unambiguous phrase such as (1) or (2). Second, there were initial subordinate clauses followed by a main clause fragment ending with an unambiguous phrase. Third, there were initial subordinate clauses followed by a main clause fragment ending with an ambiguous phrase; the initial subordinate clauses contained a proposition whose meaning produced a bias toward one interpretation of the ambiguous phrase, and they contained explicit cues, such as the presence of is versus are and number information attached to the verb, that also produced a bias toward one interpretation of the ambiguous phrase. Experiment III differs from Tyler and Marslen-Wilson (1977) by systematically varying the connective introducing the context clause and the semantic, syntactic, and lexical cues in the context clause. The hypotheses were: (a) if clauses that match expectations for thematic organization produce more complete on-line representations, syntactic information in an initial if’ clause will be more fully utilized to predict incoming words within the clause. (b) if a context clause that has not been thematically integrated produces a search for information that can be used to integrate it, syntactic information in a clause following an initial though clause will be more fully utilized to predict incoming words within the second clause.
238
D. J. Townsend
(c) if a context clause that has not been thematically integrated is retained in relatively superficial form, the superficial form of an initial though clause will be more available for resolving an ambiguity in the following clause than will that of an initial if clause. Method Each of 24 ambiguous phrases of the form ‘verb&g noun’ appeared across lists with 8 different contexts, defined by if or ?/tough introducing the context clause, a semantic bias toward the gerund or adjectival reading of the ambiguous phrase, and a singular or plural main verb in the context clause. Half of the contexts contained is or are and half contained the singular or plural form of some other verb. The contexts for 20 of the ambiguous phrases contained no ‘verb&g noun’, two contained an unambiguous terbing noun’ phrase with the gerund interpretation, and two contained one with the adjectival interpretation. None of the contexts contained a pronoun with an unspecified referent. The semantic bias of the contexts was verified by a procedure similar to Tyler and Marslen-Wilson (1977): 48 subjects were asked to continue the context plus ambiguous phrase fragments with an appropriate verb; only those context-ambiguous phrase fragments that consistently produced ‘correct’ continuations were retained for the experiment (4 sets of context-ambiguous phrases were omitted, and the remaining contexts produced continuation verbs consistent with the semanticallybiased number at a rate of 79.9%). Examples of the variations in the materials appear in Table’7. There were 24 final clause unambiguous ‘verb-ingnoun’ phrases, half with the gerund interpretation and half with the adjectival interpretation, as verified by the continuation pretest. Each unambiguous phrase appeared across lists in four different contexts, defined by if or though introducing the context and by the number attached to the main verb in the context claluse. Except for the fact there was no variation in semantic bias, the contexts for final clause unambiguous phrases were identical in form to those for fmai clause ambiguous phrases. There were 12 initial clause :mambiguous ‘verb&g noun’ phrases, half with each interpretation. Each of these phrases was introduced by either if or though across lists. There were 28 initial clause filler fragments and 32 final clause filler fragments, constructed as in Tyler and Marslen-Wilson (1977), except that 8 initial clause fillers were introduced by ifi 8 by though, and 12 by no connective;and that 8 fmal clause ffllers were introduced by ifi 8 by though, and 16 by no connective in the initial cause and 80 in the foal clause. There
Thematic processing in sentences and texts
Table 7.
239
SampIe ,?agmevats used in Experiment 111 1. UNAMBIGUOUS INITIAL CLAUSE FRAGMENTS A. Gerund Structure: If riding subways.. . 8. Adjectival Structure: If diving submarines.. . II. UNAMBIGUOUS FINAL CLAUSE FRAGMENTS A. Gerund Structure: If the boxer wants (boxers want) to avoid unnecessary injuries, dodging punches.. . 13.Adjectival Structure: If the instructor gets (instructors get) zJl the students involved, dancing classes . . . III. AMBIGUOUS FINAL CLAUSE FRAGMENTS A. Context clause does not contain is/are. 1. Singular Semantic Bias: a. Singukr verb number bias: If the pit crew works very efficiently, cars . . b. Plural verb number bias: If the pit crews work very efficiently, racing 2. Plural Semantic Bias: a. Singular verb number bias: If a young boy enjoys intense competition, cars . . . b. Plural verb number bias: If young boys enjoy the intense competition,
racing cars.. . racing racing
CUS...
B. Context clause contains is/are. 1. Singular Semantic Bias: a. Singular verb number bias: If the pilot is required to attend flight school, landing planes . . . b. Plural verb number bias If the pilots are required to attend flight school, landing planes . . . 2. Plural Semantic Bias: a. Singular verb number bias: If the airline’s ground crew is on the runway, landing planes . . . b. Plural verb number bias: If ground crews are very often on the runway, landing planes . . .
were 18 practice fragments that preceded the 108 fragmerlts in each ‘:ist. The fragments were recorded in monotone voice by an experienced male reader on one channel of tape, and on the other channel, a timing tone FNas placed so that it coincided with the onset of the last word of each fragment. Two lists of target words were prepared. The targets for ambiguous fragments were either is or are. The targets for other types of fragments were a mixture of is and are, but predominantely some other verb. The two lists of targets differed in the number attached to target verbs for ambiguous and unambiguous final clause fragments, and unambiguous initial clause fragments.
The designs for ambiguous and unambiguous fragments differed. For ambiguous fragments, whether or not the context contained a form of be varied within subjects but between items; connective, semantic bias, verb number bias, and target varied within subjects and items. For both initial clczuse and final clause unambiguous fragments, the structure of the fragment-ending phrase varied within subjects and between items; connective and target varied within subjects and items. The instructions to the subjects were to listen to sentence fragments, read aloud as quickly as possible the word that appeared by slide when the fragment ended, and then indicate whether or not the word was an appropriate or inapproptiate continuation. The sentence fragments were presented binaurally through headphones and were preceded by a warning tone one second before the onset of the fragment. The timing tone, which was not heard by the subject, triggered a shutter which allowed the projection of the word onto a screen in front of the subject, and simultaneously triggered a millisec timer. Twenty-four right-handed male native speakers of English at Columbia University served as subjects as part of a course requirement. Results and discussion Table 8 shows mean response times as a function of the connective intro-
ducing the initial clause and whether or not the number information on the target was expected based on local structural information. For ambiguous
fmgments, Table 8 shows response times for semantic expectations (whether the number attached to the target matched the semantic bias of the context, as determined by the pretest), verb number expectations (whether the number attached to the target matched the number attached to the verb in the context), and lexical expectations (whether the target was identical to a word appearing in the context). For initial clause fragments introduced by if, response times were faster when the number on the target verb was expected (based on the structure of the phrase ending the fragment) than when it was structurally unexpetted, but for initial clause fragments introduced by though, the structural expectedness of the target verb had no effect on reading times, F( 1,44) = 6.99, p < 0.05. These results agree with those of Experiment I: the allocation of attention to propositional processing of an initial clause depends on how easily the clause can be thematically integrated. However, the opposite effects occurred in unambiguous final clauses: structural expectations had an effect in stroug/z contexts but not in if
Them tic processiflg in sentences and texts
243.
T&k 8. Target reading times fmsee) for ambiguous and unambiguous fragments If
Thoush
Expected
Unexpected
Effect
-. Expected
UNAMBIGUOUS Structural expectations initial clause Final clause
673 756
796 739
123* 17
769 677
765 746
-4 69*
AMBIGUOUS Semantic expectations
681
719
38
641
667
26
Verb number expectations No is/are
in context
is/are
in context
Unex petted
611
695
Effect
85*
771
764
-7
622
658
36
612
647
35
599
739
140*
*p CO.05. contexts,
F( 1,44) = 5.82,, p < 0.05. Thus, predictions based on syntactic analysis of the final clause are more accessible when the clause follows an initial though clause than when it follows an initial if clause. These results suggest that the listener conducts syntactic analyses following a thematically incomplete clause relatively quickly in order to obtain information that can allow a thematic interpretation of the initial clause. Target reading times in final clauses were faster for ambiguous fragments p< 0.01, even when the than for unambiguous ones, F(1,44)=8.45, comparison involves only trials with is/we targets, F (1 ,1104) = 6.81 !J p <: 0.01. This suggests that, in genera!, semantic and superficial biases do not eliminate consideration of one of the alternative interpretations of the ambiguous phrase. There was a slight adv-ntage in target reading times when the target was consistent with the semantic bias produced by the proposition in the context clause. However, the overall semantic effect was not significant, F (1,44) ‘= 2.20, p > 0.10, nor did semantic appropriateness interact with conjunction, F<
1.
Target reading times were significantly faster, however, when the number information attached to the initial clause verb matched that of the target,
242
D. J. Tolwnsend
F (?, ,44) = 5.9 1, p < 0.05. The verb number effect was largely due to though
fragments which contained is/ure in the context clause (a 140 msec effect), F( 1,44) = 4.99, p < 0.05;it was negligible in though fragments without is/are in the context and in either type of if fragment, all Fs < 1. The results suggest that context effects on the syntactic processmg of a final main clause occur only when the context has not been thematically integrated and hence, when lexical and syntactic information from the context are easily accessible. An initial if clause is a complete processing unit in that its propositional meaning is stored in a thematically integrated form at the clause boundary and does not affect the processing of the next clause. An initial though clause, however, is an incomplete processing unit. At a though clause boundary, the processor retains a mixture of superficial and propositional information, awaiting full thematic closure; of these representations, spedfic lexical items, like is/are, exert the strongest effect on structural hypotheses in the next clause. The results indicate that a critical factor in the phenomenon of ‘propositional closure’ is thematic integration of the proposition. Experiment IV Experiment IV shows that the causal pattern of narratives is the basis for the organization of the propositions of a sentence in text just after the sentence has been understood. A post-sentence synonomy judgment task (c.f., Experiment I) shows that just after a complex sentence has been read in a narrative context, accessibility to clausal meaning increases as the connective introducing the clause possesses a more causal meaning. The subjects read two long stories sentence by sentence; Table 9 shows a portion of one of these stories. The experimenter occasionally interrupted the presentation of the stories after a two-clause sentence. The subjects’ task was to indicate whether or not a 2-4 word verb-object phrase was similar in meaning to any part of the sentence most recently read. The critical trials involved complex sentences in which there was a synonomous match with a final clause introduced by because, after, when, and, before, or although. If the reader organizes the propositions of a just-comprehended complex sentence in a causal or temporal pattern, accessibility to propositional meaning will be faster for causal clauses than for noncausal clauses. Method
Two stories, Characteristics of Vampires (Garden, 1972) and The S.S. Watertown(Seymour, 1976), were adapted for a synonomy judgment task.
Thematic processing in sentences and texts
Table 9.
243
A portion of a story from Experiment IV (Adapted from Garden, I9 72) 1. A young man made a bet he would not be afraid to spend a night alone in the house. 2. At first, he was not too uncomfortable. 3. A lamp was put out. 4. The fire in the grate suddenly went out. 5. Ho had expected such events. 6. The shrieks and laughter started to make him very nervous. I. At last, he could stand it no more. 8. He flung open the door and ran outside. 9a. Something was apparently trying to choke him because b. Something really strong grabbed his ankle after c. He knew he was in trouble when d. Something really strong grabbed his ankle and e. Something really strong grabbed his ankle before f. He tried not to be afraid although he felt cold fingers closing around his throat. TEST PHRASE: Squeezing his neck. 10. The young man was found the next morning alive, but in shock.
The adapted version of the Vampire story contained 135 sentences and that of the Watertown story contained 62 sentences. For each story, 16 verbobject phrases were constructed so that 6 were synonomous with tile final verb-object phrase in complex sentences, 6 were nonsynonomous with the final verb-object phrase in complex sentences, 2 were synonomous with the verbobject phrase in simple sentences, and 2 were nonsynonomous with the verb-object phrase in simple sentences. The judgments of synonomy versus nonsynonomy were verified by synonomy ratings of the phrases and sentences by an Experimental Psychology class at Montclair State College. The sentences corresponding to these phrases appeared randomly throughout the stories. Six complex sentences in each story that were paired with synonomous phrases constituted the critical sentences. These six sentences appeared in six different forms such that (a) the final clause was introduced by because, after, when, and, before, or although, (b) the initial clause was modified in order to make the resulting sentence meaningful as a whole, and (c) the final clauses were identical except for the connectives introducing them (see Table 9). These variations in the six complex sentences produced six versions of the stories, each version containing one critical sentence with each of the six connectives. The two test stories were preceded by a short practice story containing two trials. The order of presentation of the test stories was counterbalanced.
244
D. J. Townsen
The stories were presented on slides, one sentence per slide. Subjects advanced to the: next sentence by pressing a button, but they were allowed a maximum of 4 seconds for each sentence. After reading a test sentence a slide with red background containing the verb-object phrase appeared instead of the next sentence. The subjects’ instructions were to say ‘yes’ or ‘no’ if the verb+bject phrase was similar or dissimilar in meaning to any part of the preceding sentence. The vocal response automatically recorded the response time and caused the slide projector to display the next sentence of the story. The subjects were 48 Montclair State College undergraduates who were tested individually and paid for their participation. Results and discussion
Errors occurred on 5.4% of the critical trials. Response times for these trials were replaced by the appropriate cell mean. The mean response times for co-rrect responses appear in Figure 3, which shows that post-sentence accessibility to meaning differed among the various final clauses across subjects and materials, F(5,115) = 10.1, p < 0.01. Accessibility to meaning was slowest for clauses introduced by connectives which explicitly signal a noncausal event (i.e., before and although). The meaning of potentially causal clauses is more accessible. Experiment V
Experiment V shows that there is a causal organizing principle, as well as a temporal organizing Iprinciple (Clark and Clark, 1968), for retaining isolated complex sentences. The paradigm was that of Clark and Clark (1968): after reading a group of six unrelated sentences, subjects recalled the sentences as accurately as possible. Each sentence contained one of six connectives (because, since, after, when, while, or before) and appeared in one of two clause orderings (main first uersus subordinate first). If people organize complex sentences in memory according to causal and temporal principles, the recall advantage of main-first order should increase as the causal explicitness of the connectives decreases (cf., Table 2). Method
Seventy-two pairs of sentences of the form ‘Pronoun-Verb-Article-Noun’ were conjoined with either because, since, after, when. while, or before. In
Thematic processing in sentevrcesand texts
245
blocks of 12 pairs, each connective was used twice, once with initial main clause and once with initial subordinate clause. In each half of the 120pair blocks, each connective was used once. Twelve lists were generated by having each pair appear once across lists in each conne&ive . by clause order combination. A noun, intended to serve as the ‘subject’ of each sentence within a pair, was selected for each of the 72 sentence pairs. The noun and the conjoined sentence pairs appeared on IBM cards in the form: (5)
THE CANARY. IT SUNG THE SONG BECAUSE IT ATE THE SEED.
For each deck of noun-sentence pairs there was an additional deck containing only the noun cues (e.g., THE CANARY). Within blocks of 6 cards, the noun cues were presented in random order. The subjects’ task was to study each noun-sentence pair for 10 sec. After studying 6 pairs, they referred to the deck containing only noun cues and wrote down ds accurately as possible the sentence that had been paired with each noun cue. The procedure was identical in all other respects to Clark and Clark (1968). Twenty-four students in an introductory psychology class at Montclair State College served as subjects in a group. Results and discussion Verbatim recall was somewhat higher with the main-subordinate order (53%) than with the subordinate-main order (47%). The effect of clau,se order on recall, however, interacted with connective, I”t(S,BOO)= 2.36, p < 0.05. Figure 4 shows this interaction in terms of the relative superiority of initial main recall over initial subordinate recall for each connective. Three connectives which are either explicit in signalling a noncausal relationship (while, before) or inr:xplicit in signalling a causal and temporal relationship (when) were much easier to recall with initial main clause. At the other extreme, the connectives which explicitly signal a causal or prior relationship (because, after) show a slight superiority in recall with initial subordinate clauses. Together with Experiment IV, these results indicate a tendency to organize propositions in memory as cause-effect and first event-second event.
Experiment VI Experiment VI indicates that (a) the degree to which the pattern of sentence ordering in a narrative corresponds to the temporal pattern influences the
246
D. J. Townsend
Figure 3.
Accessibility to propositional meaning of j&al clauses after reading sentences.
2500
l
2480*
2460 ' A Y
2440 *
3 2 F G
0. 0
Y, P
2420 '
2400.
2380
l
amount of time that is devoted to reading individual sentences and the long term recall of the text, and (b) part of the processing time for sentences early in a text is devoted to detecting the principle that underlies the pattern of sentence ordering. The critical variable was the order of presentation of the events in texts. In normal stories (e.g., Table 1) each sentence contained an event which logically followed in time the event of the preceding sentence. In pairreversed stories, within each pair of sentences in the normal order, sentences were presented in reversed order. In backward stories, the sentences were presented in exactly the opposite order with respect to normal stories. If one component of sentence processing in the early portion of a text is a determination of the underlying pattern of sentence ordering, sentence reading times will be longer early in the story than late in the story, regardless of the pattern actually used (cf., Cirilo and Foss, 1980). If the organization of the events in stories into a temporal pattern is a component of on-line text processing, departures from the temporal pattern will disrupt sentence-reading times (cf., Baker, 1978; Mandler, 1978).
Thematic processing in sentences and texts
247
Method
Six normal stories containing 10-20 sentences were constructed so that the event in sentence n logically occurred either prior in time or simultaneously with respect to the event in sentence n + I, but no temporal connectives were used. In the pair-reversed version, sentences within a pair in the nornal version were presented in reversed order; for example, sentences a, b, c, d, e, f in the normal version were presented in the order b, a, d, c, f, e, in the pairreversed version. In the backward version, sentences were presented in the order f, e, d, c, b, a. Sentences in the three versions di’lfered only in that the Grst mention of a noun phrase in the story made use of an indefinite article and the second mention of the noun phrase made use of the definite article. Subjects were tested individually with a Radio Shack TRS-80 computer. The subject first read instructions displayed on the video screen. The subject’s instructions were to read a story one sentence at a time, to take as much time on each sentence as needed to recall the story completely. When the subject was ready for the next sentence, s/he pressed a key on the computer keyboard, which recorded the reading time and presented the next sentence. After the last sentence of a story had been presented, the scre++:il displayed the instruction to write down I&hestory on a sheet of paper. After writing the story from memory, the subject read the next story in the same manner. Each subject read six stories, but only one version of each story. Across six groups of subjects, each of the six stories appeared in each of the three versions such that subjects in each group received two normal stories, two backward stories, and two reversed stories. The order of the story versions was counterbalenced across subjects as was the order of the stories themselves. The subjects were 36 volunteer undergraduates at Montclair State College.
Results and discussion Propositional recall, without regard for order, was highest for the normal versions (66.1%), followed by pair-reverseci versions (61.7%) and lowest for backward versions (52.3%), F(2,60) = 7.23, p < 0.01. The degree of conformity in the pattern of sentence ordering to the temporal principle affects long term recall, supporting and extending previous studies (e.g., Mandler, 1978). Figure 5 ihows the average sentence reading times in the first half UNYUS second half of each of the three versions. Overall, sentence reading times
243
D. J. Townsend
Figure 4.
Relative advantage of main-subordinate* order in the recall of sentences, 22 4 20 ’
18 I 16 1 14 ’ 12 ’ 10 ’
8t 6’ 44 24 0 ‘--2 ’ -4 ’
-6 c
. BECAUSE
SINCE
.
1
AFTER
WHEN
. WHILE
BEFOIiE
were fastest for normal versions (1.75 set), followed by pair-reversed versions (2.23 set) and backward versions (2.58 set), F(2,3118) = 11.46, p C 0.001. These results show that conformity to the temporal pattern affects sentence reading time, indicating that determination of temporal relationships occurs at the level of individual sentences. In all three versions, sentence reading times were faster in the second half of the story than in the first half, F( 1,3 118) = 5.76, p < 0.001. Strictly speaking, the reading times for sentences in the first half wws second half of any version are not directly comparable because they are different sentences. However, the fact that the position effect occurred in all versions supports the view that the text processor devotes time during reading the first few sentences of a text to a determination of the thematic pattern of the text. Experiment VII
Keenan (Reference Note 1) has shown that the extent to which two events are causally related in one’s generallknowledge affects sentence reading time
Thematicprocessingin sentences and texts
249
Sentcmceread@ timesin the first half versussecond half of different versions of stories. --
3200
NDRHAL
--
PAIR-RCVERSAL
. . . . . .
BACKWARD
2800
2
z:
2000
x 9 Y
1600
FIRST
HALF
SECOND HALF
when the events appear as successive sentences in text. i?xperiment VII shows that sentential connectives also cue causal integration. Subjects read two-sentence texts one sentence at a time. Thb riecond sentence was introduced by either therefore, afterward, no connective, mean white, previousty, or however. The subjects’ task was to construct ;t continuation sen.tence for the text after each sentence. If the text processor utilizes the meanings of corrnectives (cf., Tab.le 2) to construct a ‘causal chain’ for the events during reading, continuation times will be faster for therefore sentences, which signal an effect of the preceding sentence, than for however sentences, which deny that the: next event is the effect that one would expect from the preceding event. Method Eighteen two-sentence texts were constructed such that they sounded natural with therefore, afterward, no connective, meanwhile, previously, or however introducing the second sentence. A sample two-sentence text is:
250
(6)
D. J. Townsend
Harry began raising snakes on his farm. Therefore, kids visited the farm everyday.
The eighteen texts were arranged into six lists so that each list contained one of each sentence-introducer in each third of the list. Across the six lists the position of each sentence-introducer in the list was counterbalanced. The connective introducing the critical sentences varied within subjects and within items. Twenty-two filler trials also appeared in each list; ten of these were one-sentence texts without a connective and twelve were three-sentence texts in which the second sentence contained one of the six sentenceintroducers. Thirty volunteer subjects from Montclair State College were tested individually on a Radio Shack TRS-80. Their instrutions were to read each sentence of the story that was displayed on the video screen and to construct a sentence that would be a reasonable continuation of the story. When the subject had constructed a continuation sentence, s/he pressed the button on the computer keyboard. This button-press recorded the time spent viewing the sentence on the screen and simultaneously displayed the next sentence on the screen. However, when the text ended the instruction “Write a continua-ion sentence” appeared on the screen instead of another sentence of the text. At that point, the subject wrote in a booklet the continuation sentence that s/he had in mind. When the subject had finished writing a continuation sentence, s/he pressed the button again, which presented the first sentence of the next text. Results and discussion
The critical data were obtained from t!:e two-sentence texts. Analysis of variance of the continuation times for the first sentence of these texts showed no differences as a function of list, connective, and block, all Fs < 1. Figure 6 shows the mean continuation times for the second sentence. (1~ order to take into account the fact that no-connective sentences were one word shorter than sentences with a connective, the raw mean for noconnective sentences was adjusted by multiplying the average response time per word in no-connective sentences by the average number of words in connective sentences). Continuation times varied for second sentences that were introduced in different ways, F(S, 24) = 234, p < 0.01. Integrating the propositions of successive sentences becomes easier as the connective in the second sentence signals more explicitly that the propositions appear in the expected causal/temporal pattern.
Thematic processirrgin sentences and texts
ure 6.
251
Continuation times depending on connective that introduces a sentence. 5500
5300
sroo
4900
4700
4500
4300 4 7
--1-----1
THEREFORE
AFTERWARD
NO CDNHECTIVE
P&ilILE
PREVI;“SLI
HO&A
Experiment VIII Experiment VIII demonstrates that thematic integration of a complex sentence with context depends on an interaction of the meanings of conjunctions, the ordering of clauses in the complex sentence, and which clause of the sentence presents already-known (i.e., ‘given’) information. Subjects read a context sentence that paraphrased either clearly or vaguely one clause of a target sentence containing because, while, or aCdhough. After reading the target, their task was to press a button when they had constructed a sentence that appropriately continued the text. Experiments III, VI, and VII showed that the ordering of events within and be:tween sentences interacts with connectives and the expected pattern of ordering to influence propositional processing and thematic integration. Those experiments suggest that thematic integration will be easier for because sentences when the causal event (i.e., subordinate clause) appears earlier in the sentence. Experiments I, II, III, and VII indicated that thematic integration demands for a clause which denies an expected causal relation disrupt its
252
D. J. Townsend
propositional and thematic processing. Those experiments su thematic integration will be easier for although sentences when the ‘unexpected effect’ (i.e., main clause) appears earlier in the sentence. In these cases, the availability of information about the unexpected event that actually did occur allows the reader to determine what the writer had expected to follow from the adversative clause as it is being read. Experiments VI and VII demonstrated that thematic integration effects cross sentence boundaries. Hence, a context sentence that provides the same information as the main clause of an although sentence will facilitate thematic integration of an initial although sentence just as an initial main clause facilitates thematic integration of a following although clause. This predicts that target (1) in Table 10 will be easier to integrate when it follows context (B) than when it follows context (A). Table 10.
Sample materialsfor Experiment VIII CONTEXTSENTENCES: A. Direct Paraphraseof Subordinate Clause: Harry takes care of reptiles in his house. B. Direct Paraphraseof MainClause: Children are always hanging around Harry’splace, C. Indirect Paraphraseof Subordinate Clause: Harryis fascinated by exotic animals. D. Indirect Paraphraseof Main Clause: Harry is popular with children. TARGET SENTENCES: 1. Although he raises snakes, kids often visit Harry. 2. Kids often visit Harry, although he raises snakes.
Readers attempt to integrate new information with already-known, given information (e.g., Bock, 1977; Haviland and Clark, 1974) and a new event as the effect of a just-comprehended event (e.g., Experiments VI and VII; Katz and Brent, 1968). Reliance on these strategies produces a preferred thematic ordering of information: thematic integration is easier when given information, causes, and earlier events appear before new information, effects, and later events. In simple narratives these integration strategies overlap: the most recently comprehended proposition serves as both old information and a potential causal event for integration with the next sentence. When a narrative contains a clear adversative like although, however, the thematic demand of infbingan expected but unstated etifect masks the general strategies of integrating new information in terms of old and effect in terms of
Thematic processingin sentencesand texts
2 53
cause, SO that context (B) in Table 10 facilitates integration of target (1) more than does context (A), even though context (A) makes the ‘given’ information appear earlier in target (1). Unlike because and although, while signals two distinct thematic meanings. It may signal an adversative relation; for example, (7a) and (/ ~1) are equivalent. It may also signal a simultaneous temporal relation; for example, (8b) and (8~) are similar in meaning. lVhiZe, however, does not signal a simple causal relation; (8b), unlike (8a), does not state that the subordinate clause event caused the main clause event, and (9b) is similar to (9c) but not (9a). (7a) (7b) (8a) (8b) (8~) (9a) (9b) (9c)
Although Harry took the arsenic, he didn’t die. While Harry took the arsenic, he didn’t die. Because Harry took the arsenic, he died. While Harry took the arsenic, he died. As Harry took the arsenic, he died. Harry died because he took the arsenic. While Harry died, he took the arsenic. As Harry died, he took the arsenic.
Consider how the ‘given-new’ relations between the context and a while sentence will interact with the expectation that events are ordered as causeeffect to influence the meaning assigned to while. After a context sentence has been read, it is assigned tile status of a potential causal event. If the context clearly states the same event as the while clause, rhemz:ic processing must re-assign that event the status of a presumed but unfulfilled cause, as in although, for its appearance with while indicates it does not possess a simple causal role (cf., the difference between (8a) and (8b). But if the context states the same event as the main clause, the assignment of any causal meaning for the context must be dropped, for while sentences cannot have an interpretation in which the main clause event is a causal event (cf., (9a) versus (9b)). This reasoning predicts that the interpretation of a whike sentence depends on which of its clauses presents already-known information: If ‘given’ information is in the while &use, it is interpreted as a presumed but unfulfilled cause, but if it is ti the main clause, the events in the while clause and the main clause are interpreted as simultaneous events. On the other hand, when the context sentence does not clearly paraphrase either clause of the target sentence, the context will not influence the interpretation of while. Instead, the preference for causal order within the sentence alone predominates: for contexts that are indirectly related to the target, the thematic processor will more likely interpret while to signal a (presumed) cause when the while clause is in initial position than when it is
254
D. J. Townsend
in final position. That is, an adversative mesning will be more likely in (10) than in (I 1): (10) (11)
While Harry took the arsenic, he didn’t die, Harry didn’t die while he took the arsenic.
Behavioral evidence for the meaning of while POdepend on clause position in the absence of a context comes from Townsend and Bever (1978): the online Eropositional processing of initial while is like that of initial though, but that of finaSwhile is like final when. Method
Twenty-four context-target passages were constructed. Each of the conjuno tions because, while, and although were paired with eight target sentences. There were eight forms of each passage defined by clause order in the target (initial main versus initial subordinate), which clause of the target the context paraphrased (main versus subordinate), and how direct this paraphrase was (direct versus indirect). Half of the forms of each passage contained a context sentence that directly paraphrased one of the clauses of the target sentences; for example, “Harry raises snakes in his house” is more directly paraphrased by “Harry takes care of reptiles in his house” than it is by “Harry is fascinated by exotic animals”. 4. complete set of variati::jnsof an although passage appear in Table 10. Eight lists were constructed based on the twenty-four context sentenceta!mgetsentence passages. In each list there was one instance of each combination of conjunction, clause order, information order (i.e., given first versus ne:w fust), and type of antecedent (direct versus indirect paraphrase). Across lists, each set of lexical items in target sentences appeared in eight forms. Hence, all factors varied within subjects and within items, except for ccnjunction, which varied only within subjects. Twenty-four practice passages similar to the test passages appeared at the beginning of each list. The subject’s instructions were to read each context sentence, press a re!iponse key when s/he was ready for a target sentencs, read the target sentence and construct a sentence that might follow the context-target pair in a story, press the key, and then report the continuation sentence. Sentences were presented by slide projector. The first key press removed the mntext sentence, started a timer,, and displayed the target sentence. The second key press removed the target sentence and stopped the timer. Fortyeight volunteer undergraduates at Montclair State College were tested individually.
Thermtic processingin sentences and texts
25:)
and the ordering of clauses in the target sentence affected response times, Response times were faster for the subordinate-main order in because sentences (7.09 versus 7.74 SW) and in w/rile sentences (7.22 versus 7.49 set), but not in although-sentences (6.94 versus 6.53 set), Fl (2,80) = 4.83, g < 0.01, F2 (2,21) = 2.21, p > 0.05, supporting the prediction that because-sentences are easier to integrate when the causal event occu.rs initially in the sentence, but that althougizsentences are easier when tile ‘unexpected effect’ is in the initial position. Conjunction and clause order interacted with which clause the context paraphrased, F, (2,80) = 7.45, p < 0.001, F2 f2,21) = 3.18, p > 0.05. The nature of this interaction appears in Figure 7. Response times were faster (by 0.52 set) for because-sentences when the context paraphrased the subordinate Claus< than when it paraphrased the main clause. For although and while sentences, however, the opposite relationship was obtained : response
Figure 7.
Continuation times depending on clause order, cormec;ive, and paraphrase relationsbetween the context and the tavet.
.. 7 9 w wo I: F: Y [
6000
Y
BECAUSE
---
*
ALTHOUGH
....... WIlE 5500
1
RAIN-SUCbDRDINATE
SUBORDINATE-HAIN
CONTEXT PROV I DES THE CAUSE ( PRESUHED CAUSE )
NAIN-SUBORDINATE
SllBORDl NATE-HA1 N
CONTEXT PRWIDES EFFECT ( MEXPECTED
THE EFFECT
’
256
D. J. Townsend
times were faster for althotigh sentences (by 0.94 set) and slightly faster for while sentences (by 0.08 set) when the context paraphrased the main clause rather than the subordinate clause. The results for although sentences indicate that thematic processing uses the context to determine the expected cause-effect relation implicit in the althotigh sentence. This can only be done, however, when the context provides the ‘unexpected effect’ that actually occurred, that is, when it provides the proposition which is started by the main clause of the alfhough sentence. In contrast to Experiments II and III, in which a causal clause had little effect on the processing of the following clause, there was evidence in Experiment VIII that the nature of the context, i.e., whether it paraphraksed the main or subordinate clause, did influence continuation times for sentences with a causal conjunction. The presence of a context effect for causal sentences in Experiment VIII and its absence in Experiments II and III suggests that context alffects only thematic processing and not propositional processing. The difference in continuation times for various because sentences depending on the nature of the context is due to how easily the propositions of the because sentence can be integrated with the context rather than to how easily its propositions are apprehended. The mean response times for while sentences showed a different overall pattern from those for either because or a&/lough sentences. When the context paraphrased the subordinate clause, response times for while were faster in the main-subordinate order, just as in although sentences, but when it paraphrased the main clause, response times were faster in the subordinatemain order, just as in because sentences. These results support the view that the text processor assigns the adversative meaning of although to while when the context paraphrases the while clause, but it assigns the temporal meaning of as to whik when the context paraphrases the main clause. To check further on this interpretation, 56 additional subjects rated how likely it was that a speaker intended while to mean although for all of the context-while sentence pairs; 8 groups of subjects were defined by which context-target version they received for each of the while sentences, They used a scale from l-7, with 7 indicating that an although meaning was very likely. Table 11 shows the mean hkelihood ratings. Overall, subjects were more likely to assign while th[e meaning of although when the context paraphrased the subordinate clause than when it paraphrased the main clause, F (I 34) = 7.79, p C 0.01. This accounts for the response time data shown in Figure 7: when readers already tc~ow the information that is stated in the subordinate clause of a while sentence, they interpret while like alrhough, but when they know the main cliause information, they are more likely to interpret the sentence as stating atemporal setting.
Ti”2ema tic processing in sentences and texts
2 57
Table 11. Mean likelihood ratingsthat whilemeans although Event in context
Direct paraphrase
Hndirect paraphrase
MS
SM
Mean
MS
SM
Mean
Subordinate Main
3.57 2.67
3.52 2.61
3.55 2.64
2.89 3.02
3.66 3.56
3.27 3.29
Mean --
3.12
3.07
2.96
3.61
Overall
3.41 2.97
This pattern, however, occurred only when the context directly paraphrased one of the clauses of the target: in the assignment of meanings to while sentences, the clause that the context paraphrased interacted with how direct the paraphrase was and the order in which the clauses were presented, F (134) = 4.75, p < 0.05. For direct paraphrases, subjects were more likely to interprt:t while as although when the context paraphrased the subordinate clause, ar d clause order had little effect. But for indirect paraphrases, subjects were more likely to interpret. while as aZthough when the clauses were presented in subordinate-main order, and which clause the context paraphrased had little effect, Whethelf while is assigned the status of a temporal setting or the denial of a causal relation depends on the use of a thematic integration strategy: use old information as a potentially causal event for organizing propositions in memory. When the context is clearly related to the while clause, it is interpreted as a (presumed) cause, but if it corresponds to the main clause, it is interpreted as temporal. When the context is not clearly related to the while sentence, the ‘cause first’ strategy operates on the target alone and initial while is interpreted as a (presumed) cause. General discussion These studies demonstrate that semantic relations between the clauses of isolated sentences and between sentences in narratives have similar effects on the processing of clauses and sentences; indicating that within-clause processing is similar whether the clause appears in an isolated sentence or in a text. In the processing of an isolated sentence, the presentation of the events of a sentence out of cause-effect (or first event-second event) order disrupts the comprehension of the sentence: it produces relative inaccessibility online to the meaning of the disruptive clause and the retention of a relatively superficial representation of it (Experiment I), relatively rapid propositional
258
LXJ. Townsend
processing of the next clause (Experiments II and III), and poorer long term recall of the entire sentence (Experiment V). In the processing of sentences in narratives, the text processor organizes the propositions of a just-comprehended sentence according to a causal/ temporal pattern (Experiment IV); departures from the predominant causal and temporal patterns of sentence ordering increase sentence reading times and decrease: sentence recall (ExperLqlent VI), and the ease with which the text processor constructs appropriate continuations for a text depends on how useful the most recently processed sentence is for constructing the relevant thematic relations (Experiments WI and VIII). For both isolated complex sentences and sentences in texts, the processor determines propositional meanings and batween-proposition relationships. The presence of a surrounding context does not eliminate within-proposition processing, but instead, adds the further ta:sk of determining the relationship between the sentence and its context (cf., Clark, 1978; Cutler, 1976). The two general processes in text processing apply different sources of knowledge to different types of information. Propositional processing applies lexical and syntactic knowledge to the words of a clause or sentence to formulate a representation of its propositional meaning. Thematic processing applies knowledge of the meanings of connectives and other surface cues, and knowledge of predominant patterns of sentence ordering to formulate a representation of the relationships between clauses and sentences; the text processor ?ypically organizes these representations according to the predominant pattern found in the text. Propositional and thematic processing interact: when expectations or explicit cues indicate that a proposition disrupts the thematic organization that the text processor seeks, the text processor focuses on gathering further information that can allow a determination of the thematic relevance of the disruptive clause. This reallocation of attention occurs at the expense of a delay in the propositional processing of the clause and the retention of its superficial form. It also entails a focus of attention on propositional procAng of the next clause, greater accessibility to superficial information from the thematically incomplete clause for integraiion with information from the following clause, and poorer accessibility to the proposition once it is stored in memory. The text processor uses the meaning of a context clause or sentence for determining the thematic relevance of a proposition, but not for determining the propositional meaning of a clause. Thematic processing occurs in parallel with propositional processing: as propositional information becomes available, it is thematically integrated with the context. When the thematic relevance of successive propositions corresponds to the processor’s expectations, propositional processing is
Thematic processing in sentences and texts
259
relatively unaffected. In these cases, there are clear ‘closure’ effects at clause boundaries (Experiments I--III); the activities of the thematic processor may be less apparent, but the fact that the nature of a context sentence influences integration time for a because sentence indicates that thematic processing occurs even when events are ordered as expected (Experiment VIII). When the thematic role of a clause does not correspond to the expected pattern, the activities of the thematic processor become apparent quite early in the processing of the nonconforming clause and closure does not occur at the clause boundary as the thematic processor searches for, or infers, information that can make clear the thematic relevance of the clause; unintegrated information from the nonconforming clause is then available for propositional integration beyond the clause boundary. These studies indicate that the organizational unit of comprehension is the linguistic clause or, more precisely, the proposition. The fact that there are effects of thematic integration on sentence processing reinforces this view. All of the observed effects on sentence processing occur just because the proposition is the unit of syntactic and thematic processing-when the thematic role of a proposition cannot be determined from t e information available, its syntactic processing is disrupted. Propositional processing utilizes any unintegrated information that is available; thematic processing utilizes the propositional info.rmation that is available. Although there are many aspects of text structure besides those considered here, the general outline of this model remains valid. The text processor’s goal is to apprehend the theme, it does so by determining the relevance of each proposition to the text structure as the proposition is apprehended, and the appearance of a sentence whose textual role is not easily determined disrupts the propositional processing of the sentence.
References Baker, L. (1978) Processing temporal relationships in simple stories: Effects of input sequence. J. verb. Learn. verb. Behav., 17, 559472. Bock, J. K. (1977) The effect of a pragmatic presupposition on syntactic structure in question answering. J. verb. Learn. verb. Behav., 16, 723- 734. Brooks, C., and Warren, R. P. (1970) Modern Rhetoric. New York, Harcourt, Brace and World. Carpenter, P. A., and Just, M. A. (1977) Integrative processes in comprehension. In D. LaBerge and S. J. Samuels (eds.), %)a& plocesses in Reading: Perception and Comprehension. Hiilsdale, NJ, Erlbaum. Cirilo, R. K., and Foss D. J. (1980) Text structure and reading time for sentences. J. verb. Learn. verb. Behav., 19,96- 109. Clark, H. 11. (1973) The language-as-fixedeffect fallacy: A critique of language statistics in psychological research. J. verb. Learn. verb. Behav., 12, 335459.
260
D. J. l”owmend
ch&
H. X. (1978) infening what is meant. In W. J. M. Levelt and G. B. Flares d’Arcah (eds.), &&s hz the Pcneption ofknguage. New York,Wiley. Chuk, H. H., and Clark, E. V. (1968) Semantic distinctions and memoryfor complex sentences.f$ J.
EX~IV.&y&ol., 20, 129-138. Cmik, F. I. M., and Lockhart, R. S. (1972) Levels of processing A framework for memory research. J. r~?&.Leum. verb.Behav.,11,671-684. Cutler, A. (1976) Beyond parsing and lexical look-up: An enriched description of auditory sentence comprehension, in R. J. Walesand E. Walker(eds.), New Approaches to Language Mechanisms. Amsterdam, North-Holland Publishing Co. Dakin, J. (1970) Explanations. J. Ling., JO, 199-214. Dik, S. C. (1968) cbordinotion: Its lmpliuations for the %ory of General Lilmguistlcs.Amsterdam, North-Holland Publishing Co. van Dijk, T. A. (1977) Text and cbntext. London, Longman. Garden, N. (1972) Vampire& New York, Lippincott. Haviland, S. E., and Clark, H. H. (1974) What’snew? Acquiring new information as a process in comprehension. J. verb. Lewn. verb. Behav.,13,s 12-52 1. Jcspersen, 0. (1940) The ‘split inftitive’ and a system of clauses. J: f. f?. T&t, VI, 153- 17 1. Iespersen, 0. (1964) h’ssentidsof English Grammar. University, Ala., University of Alabama Press. (First published 1933) Katz, E. W., and Brent, S. B. (1968) Understandinp connectives. J. ve& Learn. verb. Behav., II, 501-509. Mandter, J. M. (1978) A code in the node: The use of a story schema in retrieval.Discourse Processes,
1.14-35. Mandler, J. M., and Johnson, N. S. (1977) Remembranceof things parsed: Story structure and recall. Cog. Isychd., II, 111-151. Quirk, R., Greenbaum, S., Leech, G., and Svartvik, J. (1972) A Gmmmcrrof Contempomry English. New York, Seminar Press. Schank, R. S., and Abelson, R. P. (1977) Scripts, plans, goals, and understanding. Hillsdale, NJ, Erlbaum. Seymour, S. (1976) Ghosts. New York, Lippincott.
Townsend, D. J., and Bever, T. G. (1978) Inter-clause relations and clausal processing. J. verb. Learn.
verb. B&ax, Z7.509421. Tyler, IL. K., and
Marslen-Wilson,W. D. (1977) The on-line effects of semantic context on syntactic process@. J. verb. Lewn. verb. Behav.,16,683-692. Q&t, P. (1972) Some observations on how people ar&werquestions about sentences. J, verb, Learn.
verb. Behav..,II, 188-195.
Reference Note 1 Keenan, J. M. Inferring causal connections in prose comprehension. Paper presented at American
Psycl~ologicalAssociation meetings, 1978.
Puisque de nombreuses relations gmantiques existant entre des clauses de phrases complexes se retrouvent entre les phrasesdans un texte, une description du traitement de la phrase complexe devrait &re g&t&aBsablepour le traitement des phrws dans un texte. Dans cet article, on 6tudie l’utilit&
7Rematic processingin sentences and texts
26 1
nbralisation pour dkrire le traitement en temps r6el des phrases dune histoire. Quatre expbdcnces fakes sur le traitement et la ktention de phrases complexes isolkes montrent que les connectifs s&r&ant une interruption de l’ordre causal et temporel des propositions influencent le traitcmcnt en temps r&l et le rappel. Quatre autres expkiences portant sur le traitement et la r&tention des phrases en contexte montrent que les relations causales/temporelles dans des phrases complexes et entre des phrases affectent le temps de lecture et I’acquisition des propositions en mdmoire & long terme et en memoire immbdiate. Ces rkultats suggkent des systGmes de traitement particllcmcnt inddpcndants A la fois pour les phrases cc-mplexes isolCes et les phrases de textes. Une suite de processus opke sur I’information superficielle pour donner une representation propositionnolle littirale, unu autre op6re sur l’information propositionnelle pour donner une reprksentation th6matique dc la phrase ou du texte. En gin&al, les ticteurs qui obscurcissent la pertinence semantique d’unc clause ou d’unc phrase affcctent aussi, indirectement, le traitement du sens de la proposition.