patterns of sentence processing deficit: processing simple sentences can be a complex matter

J. Neurolinguistics, Volume 6, Number 2, pp. 79-101, 1991 0911-6044/91 $3.00+.00 Printed in Great Britain Pergamon Press plc

Patterns of Sentence Processing Deficit: Processing Simple Sentences can be a

Complex Matter

Maria Black

University College London

Lyndsey Nickels

Birkbeck College, London

Sally Byng

The National Hospitals College of Speech Sciences, London

ABSTRACT

The performance of normal and aphasic subjects on a sentence-picture matching task shows that many different factors can affect accuracy. The data presented in this paper suggest that, for normal subjects, reversal errors are not errors of sentence parsing or comprehension, but

reflect ‘late’ stages of processing which mediate the translation between language and pictures. We argue that normal performance is significantly affected by the semantic or conceptual properties of the verbs used and by the interaction of verb meaning and sentence meaning.

Aphasic errors, on the other hand, can arise from different sources and have multiple causes. We discuss data from three aphasic subjects who appear to have similar linguistic symptoms, and we argue that each patient fails on this task for different reasons.

INTRODUCTION

Sentence-picture matching tasks have been used in both psycholinguistic experi- mentation and in the assessment of aphasic patients. The stimuli used in this task are usually reversible active sentences, with or without passive counterparts, which must be fully processed to achieve accurate performance; that is, the thematic roles of the noun phrases in a sentence cannot be identified correctly on the basis of semantic or pragmatic plausibility alone. Failure to perform correctly on this task has often been interpreted as an inability to form an adequate linguistic representation, particularly

80 Journal of Neurolinguistics, Volume 6, Number 2 (1991)

at the syntactic level - hence the tendency to consider the task a test of “syntactic comprehension.“’

However, there is increasing awareness in both the psycholinguistic and neurolinguistic literature that the task does not exclusively tap syntactic parsing and that different levels of linguistic processing are involved (Tyler 1987; Carrithers 1989). Indeed, since performance is usually measured in terms of accuracy (error rates), what we may be measuring is the outcome of sentence comprehension processes and not necessarily the comprehension processes per se. Furthermore, we are not only measuring the end product of linguistic processing, but also the result of non-linguistic

processes since picture recognition and interpretation, and their interaction with language, are also involved. Failure on this task, therefore, can reflect breakdown or impairment of many different types of processes.

In this paper, we will discuss some theoretical reasons and empirical evidence to show that the processing of simple reversible sentences in a sentence-picture matching task is a complex matter involving factors which have not usually been taken into account and are still poorly understood. We will then present evidence from three aphasic patients to show that aphasics who appear to have similar linguistic symptoms can fail on this task for very different reasons.

WHAT IS INVOLVED IN THE TASK?

What does a subject have to do in order to select the appropriate picture in this type of task? The answer to this question may partly depend on how the task is presented. If a spoken (or written) sentence is presented first, the subject has to form a representation of the input sentence and must hold that representation long enough to guide and constrain picture selection. Once the pictures are revealed, each picture has to be interpreted, and a representation of the depicted event or state must be constructed. The representation of the pictures then has to be compared against that of the sentence and a decision about ‘the best fit’ must be made.

In comprehending simple active sentences with two noun phrases and a verb, a listener has to go through at least the following processing ‘stages’:2

(i) the incoming string has to be processed acoustically and phonologically so that the input string can be segmented into its component elements - phonological words and phrases. As each phonological word is identified, lexical access and recognition take place; (ii) the initial process of segmentation and recognition serves as the basis for the construction of a syntactic representation which specifies the linear order of elements and the major structural relations between them. This syntactic representation serves as the input to a set of ‘mapping’ functions or procedures which link

Patterns of Sentence Processing Deficit 81

particular structural positions, such as the Subject or Object positions, with par-

ticular thematic roles. But at this stage the representation is purely structural and does not yet contain any information about the thematic roles of the noun phrases - one does not know ‘who is doing what to whom;’ (iii) the particular thematic roles to be mapped onto structural positions must be selected by reference to the lexico-semantic representation of the verb. If, for instance, the verb expresses an action (e.g. kiss), the thematic roles to be mapped will be Agent and Theme/Patient, while if a verb of perception or psychological state is involved (e.g. see), the relevant thematic roles will be Experiencer and Stimulus.3 Without accessing the lexico-semantic representation of the verb, the listener does not know which thematic roles are available for assignment (see Chomsky 1985; Pinker 1990); (iv) each thematic role has to be mapped onto a particular structural position, e.g. the Agent might be mapped onto the Subject position, and the Theme/Patient onto the Object position. The mapping might be determined by item-specific information, or by general procedures which apply to particular pairs or sets of thematic roles, irrespective of the particular verb or lexical item involved (see Williams 1981); (v) the semantic ‘content’ of the phrases must be combined and integrated to form a full semantic representation of the sentence, specifying the precise nature of the event or state expressed by the sentence, and the precise nature of identity of the participants that play certain roles in that event or state; (vi) this semantic representation serves as the basis for inferential processes, judgements of semantic or pragmatic plausibility, and the identification of particular features or properties that might guide or aid the process of picture recognition and interpretation. Although very little is known about how picture perception after hearing a sentence might differ from picture perception without linguistic input, it is reasonable to assume that hearing a linguistic description of an event or state affects our perception of a picture of that event or state; for instance, one of the pictures in the array might be immediately disregarded and not fully processed simply because it does not contain some feature or property associated with the particular kind of action or state expressed by the sentence. Or we might attend to different aspects of the picture depending on whether we hear a verb that ‘focuses’ on the effect of an action (e.g. $fill> as opposed to a verb that ‘focuses’ on the action itself (e.g. pour).

This is obviously a very simplified sketch of the main ‘stages’ involved in performing this task, and it begs a number of important theoretical questions about the precise nature of the representations and procedures involved, and the temporal relations between different levels of processing. But it should at least make clear that performance could


be affected by factors at any one of these levels. Selection of the wrong picture could

mean that the sentence has been incorrectly interpreted and the picture to match that

incorrect representation has been selected; or that the sentence has been correctly

interpreted but a picture that does not match it has been selected because something

has gone wrong in the translation from language to picture, or in the process of picture

recognition itself.

In the case of an aphasic patient, identification of a deficit means showing that the

breakdown occurs at a particular stage, or in the ‘translation’ between two stages. For

example, an aphasic subject can be said to have a ‘syntactic deficit’ only if we can

show that s/he cannot construct the syntactic representation required for stage (ii).

Similarly, a patient can be said to have a ‘mapping deficit’ (see Schwartz ef al. 1985;

Byng 1988) if the lexical information or procedures, and/or the general mapping

procedures required for stages (iii) and (iv) can be shown to be impaired.

Most psycholinguistic and neurolinguistic studies which have used this task have

focused on ‘early’ stages of processing [e.g. stages (ii) and (iv)] and we still know very

little about stage (iii) or ‘later’ levels of processing. In the next section, we will discuss

some reasons why processing factors at ‘later’ levels might also affect non-aphasic

subjects’ performance.

NORMAL PERFORMANCE

Method

Procedure and materials Six sets of 10 sentences were constructed:

(i) active action: the verbs in this set express actions and assign the thematic role

of Agent to the Subject and Theme/Patient to the Object (splash, scold, expel,

photograph, protect, grab, help, jZnd, wash, weigh); (ii) active non-action, the verbs in this set express perceptual or psychological

stages and assign the thematic roles of (a) Experiencer to the Subject and Stimulus

to the Object (dread, admire, notice, hear, see); (b) Stimulus to the Subject and

Experiencer to the Object (surprise, astonish, delight, shock, impress),4 (iii) passive action: these sentences were derived from those in set (i) by passivizing

the verb and maintaining the order of the NPs (e.g. the judge weighs the pilot => the judge was weighed by the pilot); or by passivizing the verb and changing

the order of the NPs (e.g. the pilot was weighed by the judge); (iv) passive non-action: these sentences were derived from those in (ii) in the way

outlined above;

(v) adjectives: these sentences contained adjectives expressing psychological


states or dispositions (shy, fond, sad, sorry, friendly, nasty, rude, polite, cruel,

pleasant);4

(vi) locative: these sentences contained spatial prepositions (in, on, under (twice), inside, over, after, above, below, behind).

The verbs in (i) and (ii) were matched for frequency, using “total verb frequency” ratings from Francis and Kucera (1982). The adjectives in (v) were frequency matched with the verbs as far as possible, but no frequency match was possible for the prepositions in (vi) because of the limited range of items. Frequency rating for verbs and adjectives were as follows:

Action verbs: scold, 2; splash, 7; expel, 8; photograph, 12; weigh, 33; grab, 37;

protect, 71; wash, 83; help, 352;$nd, 1033.

Non-action verbs: dread, 2; astonish, 6; delight, 20; shock, 23; admire, 32;

impress, 36; surprise, 76; notice, 84; hear, 433; see, 1513.

Adjectives: nasty, 5; rude, 6; polite, 8; shy, 11; fond, 13; cruel,l5; sad, 35;

pleasant, 35; sorry, 49; friendly, 61.

Except for the locative sentences, the Subject and Object noun phrases were both human. The characters were chosen so that they were equally likely, or unlikely, to perform an action or be in a certain state, and were clearly visually distinct. The pairings of characters were: clown/astronaut; boxer/cowboy; judge/pilot; cook/dancer; vicar/sailor; nun/queen; swimmer/workman.

The order of presentation of the sentences was randomized both within and between categories, and this order was maintained for each subject.

For each sentence, three pictures appeared on a page, in a vertical array (see Fig. 1). These pictures consisted of a target picture, a reverse role picture and a lexical distractor picture. Each picture portrayed the same characters; the target and reverse role pictures involved the characters in the same action/state, but the roles of the characters were reversed (e.g. Target: the nun splashes the queen; Reverse Role: the

queen splashes the nun). However, the action/state was portrayed in different ways in each picture to prevent the possible use of cues from picture similarity (e.g. the nun splashes the queen by jumping into a puddle, but the queen splashes the nun with water from a cup). The lexical distractor pictures involve the same character but in a different action/state (e.g. the queen touches the nun). The active and passive forms of a verb used the same picture sets. The left/right order of the characters in the pictures was varied; the vertical order of the target, reverse role and lexical distractor pictures was also varied.

The test was administered in two equivalent versions. Version B was derived from Version A by keeping the pictures constant but changing the order of NPs in half of

NE‘ 6,2--8


the sentences (e.g. Version A: the judge weighs the pilot; Version B: the pilot weighs

the judge), with appropriate adjustments to maintain all the design features outlined above. Eleven subjects were given Version A of the task, and ten subjects Version B.

Each subject heard a sentence or read one on a card. S/he was then shown the array of three pictures. Subjects were instructed to point to the picture that ‘went best’ with the sentence. Eleven subjects were tested on the spoken version of the test, and ten on the written version.

Figure 1.


Subjects

Twenty-one non-aphasic subjects were tested on this task. All subjects were native British English speakers, whose ages ranged from 15 to 80. Three of the subjects were under 20, three were in their 5Os, nine in their 6Os, five in their 7Os, and one was 80. Nineteen of the subjects were female and two were male. None of the subjects had any formal education beyond the age of 16.

Results

As there were no significant differences between subjects’ performances on the auditory and written versions of the task (Mann-Whitney, z = 0. 32 1, p < 0. l), the data from the two versions were collapsed and analysed together. The data from Versions A and B were also collapsed and analysed together as there were no significant differences between these two versions. There were significantly more correct responses than error responses and significantly more reversal errors than lexical errors (Wilcoxon, z = 3.8, p<O.OOl). Figure 2 shows the proportion of correct

responses for each task condition. The subjects’ accuracy was affected by the semantic or conceptual properties of the

verb in the sentence. There were significantly more errors on non-action sentences than on action sentences, and this was true of both active and passive sentences (Mann-Whitney, p<O.O5 and 0.001; Wilcoxon, p<O.O5 and 0.001 respectively). There were also significantly more errors on the adjective sentences, which also express a perceptual or psychological state, than on the action active sentences (Mann-Whitney, p < 0.05; Wilcoxon, p < O.OS), while there was no significant difference between the adjective sentences and the non-action active sentences. These differences, however, cannot be reduced to a difference between actions and states because error rates on the locative sentences, which express a state, were not significantly different from those on the action active sentences, while significantly less errors were made on the locative sentences than on the non-actional active sentences (Mann-Whitney, p<O.Ol; Wilcoxon, p< 0.05). So, there is something specific to perceptual and psychological predicates and/or their pictorial representation that affects accuracy on this task.

The type of verb, and its associated thematic roles, also seems to interact with sentence type. Overall, passive sentences were not significantly more difficult than their active counterparts. There was a significant active/passive difference only for the subset of the non-action verbs which, in the passive, assign the thematic role of Stimulus to the Subject noun phrase, and Experiencer to the noun phrase in the by- phrase (e.g. see, admire) (Mann-Whitney, p < 0.01; Wilcoxon, p < 0.01). There was no significant active/passive difference for the other subset of non-action verbs, nor


10 r

Actwe actvan Actwe no”-acton

Passive actlon Passive non-actm

Pdjectlves Locat Ives

Figure 2.

Patterns of Sentence Prwessing Deficit 87

was there any significant difference between the two non-action subsets in the active

version. For action verbs, there was no significant difference between active and passive

sentences (Mann-Whitney, p > 0.05; Wilcoxon, p > 0.05). But the pattern of errors also suggests that some passives might be more difficult depending on the verb involved, although the data are too limited to draw any firm conclusions in this respect. Subjects made a total of six errors on the active version of splash, photograph, and protect, and no errors on the passive, while for scold, expel, grab, and help a total of five errors were made on the passive, and no errors on the active. No errors were made on either active or passive versions of wash, while there was one error on both versions of weigh and Jind.

It is unlikely that these results could be described in syntactic terms. For instance, there are no systematic syntactic differences between the active action sentences and active non-action sentences we have used, even if we take into account some recent syntactic analyses of psychological verbs (see Belletti and Rizzi 1988). Similarly, there appear to be no syntactic properties that are common to the verbs on which more errors are made, nor any syntactic explanation of why a subset of non-action verbs should be particularly difficult in the passive. On the other hand, the results can be quite naturally described in semantic and/or conceptual terms, which strongly suggests that semantic/conceptual factors play a role in this task, as we will discuss in the next section of the paper.

Visual/conceptual factors might also be involved: the difference between action verbs and verbs and adjectives expressing perceptual and psychological states might be partly due to a difference in ‘picturability’ as psychological states are undoubtedly more difficult to depict successfully. But ‘picturability’ cannot be the whole explanation because similar differences between action and non-action verbs have been found in experiments with children where pictures were not involved and in children’s spontaneous speech (see Sudhalter and Braine 1985; Pinker 1990). And ‘picturability’ alone cannot explain the verb-specific effects in relation to the passive.

What these results highlight is that measuring accuracy of performance on this task might tell us more about ‘late’ stages of sentence interpretation than about parsing and ‘on-line’ comprehension (see Carrithers 1989). To clarify this point, we will now discuss the verb-specific effects in more detail.

VERB-SPECIFIC EFFECTS

Many theorists have claimed that passive sentences differ in meaning from their active counterparts (see Pinker 1990 and references therein), and it is generally agreed that passives have additional pragmatic properties. According to Pinker (1990), the ‘core meaning’ of the passive is roughly equivalent to “X is in the state/circumstance


defined by Y having acted on it.” The ease with which a verb can be passivized, or

the ease with which a particular passive sentence can be interpreted, will therefore

depend on how well the meaning of a verb lends itself to such an interpretation. Some

verbs will not tit into this interpretation at all and cannot be passivized (e.g. * X5 is

costed by the book), while others may undergo passivization but the resulting passive

may be more or less acceptable or easy to interpret depending on the semantic and

pragmatic properties of the noun phrases involved, and/or on the pragmatic context

(e.g. The thief was seen by the police vs? the flower was seen by John). Pinker dis-

tinguishes two main aspects which have to be taken into account in assessing the

compatibility of verb meanings with the interpretation of the passive - what he calls

“the agency effect” and “the predication effect.”

What is meant by “agency effect” is that “ . . . the passive forces an interpretation

whereby the existence of an agent-like argument or party responsible for the circum-

stance predicated of the Theme is asserted. ” “ . . . in the passive, the by-object should

be more agentive (or at least not less agentive), than the surface subject.” (Pinker

1990, pp. 91,93.) It should be stressed that “agentive” or “agent-like” does not mean

that the by-object must have the thematic role Agent, since by-objects with a variety

of thematic roles can appear in the passive. The crucial notion here is that of relatively

greater responsibility or causal weight.

Several researchers have provided empirical evidence that verbs expressing inter-

actions between entities have a sort of inherent causality in the sense that one entity

is seen as more causally responsible or weighted than the other (Brown and Fish 1983;

Caramazza et al. 1977; Au 1986; Corrigan 1988). Although there is little theoretical

consensus on the source of these causal attributions, the empirical findings are quite

strong and reasonably consistent. It is also clear that these relative causal ratings are

not entirely determined by the thematic roles involved. For instance, it is not the case

that Agents are always attributed more causal weight than Themes/Patients, e.g. with

verbs like thank, congratulate, praise, punish, reproach, and scold, Theme/Patients

are attributed more causal weight.5 Consistent causal ratings are also assigned even

if there is no Agent at all. In studies of psychological verbs involving a Stimulus and

an Experiencer, it was consistently found that experimental subjects attributed signi-

ficantly higher causal ratings to the Stimulus, regardless of whether the Stimulus was

the subject or object of a sentence and of the animacy of the noun phrases involved

(Brown and Fish 1983; Au 1986; Corrigan 1988). These studies provide us with causal

ratings for seven out of the ten non-action verbs we used, though it is reasonable to

assume that similar causal ratings would apply to the other three verbs since the same

results have been obtained with a variety of perceptual and psychological verbs.

Given these causal ratings, “the agency effect” can be used to explain why subjects

are less accurate on non-action passive verbs where the Stimulus is in subject position

(e.g. see, admire). If phrases assigned the thematic role of Stimulus were regularly


interpreted as referring to an entity with greater responsibility or causal weight, then verbs that assign Stimulus to the subject position and Experiencer to the by-object would go against “the agency effect” because the by-object would be less “agentive”

than the surface subject. What we are suggesting is that, in the process of understanding a sentence, certain

causal inferences are drawn, for instance, about the relative causality of the participants in the situation expressed by the sentence. Some of these inferences will be based on the meaning of the verb and the relative causality attributed to each of its arguments which are differentiated by their particular thematic role, e.g. verbs of perception or psychological state assign greater causal weighting to the Stimulus argument as opposed to the Experiencer argument. Other causal inferences will be based on the interpretation of the sentence as a whole, e.g. the meaning of a passive sentence forces or licenses the inference that the subject is less causally weighted. If different causal inferences are mutually compatible or reinforce each other, then a causal ‘schema’ or conceptual model of the situation expressed by the sentence can perhaps be constructed more quickly and with less difficulty. Such a ‘schema’ can then be used to guide and facilitate the recognition and selection of the appropriate pictures. However, if the different causal inferences conflict, it might take longer or be more difficult to construct a coherent conceptual model as one or other set of inferences might have to be disregarded or some other way of resolving the conflict might have to be found. In a sentence-picture matching task, this might mean that experimental subjects are less sure about which picture they should select or they might be less accurate in their selection.

Although we have focused on non-action verbs in passive sentences, similar or related factors might affect subjects’ responses on active and passive sentences with action verbs. For instance, an active sentence with a verb expressing a simple action that does not result in a change of state for the object (e.g. kiss) might be easier to interpret and match to a picture than one with a verb expressing an action that results in a change of state for the Object (e.g. splash). On the other hand, verbs like kiss might not be easier in the passive. If, as we have said, the subject of the passive is interpreted as being in a state of circumstance which is brought about by another entity, then verbs that express an action with a specific effect on the Object would fit in best with the meaning of the passive, and indeed many natural languages restrict passivization of actional verbs to these (Keenan 1985). In English, verbs that do not involve a change of state such as verbs of contact (e.g. touch, kiss) and verbs of motion and contact (e.g. kick, hit, slap, grub) can be passivized, but their interpretation and acceptability might involve additional inferences and pragmatic assumptions (e.g. compare Mary was kissed by John with Mary was kissed by Nelson Mandela).

If this analysis is correct, it would imply that the relative difficulty of active and passive sentences with action verbs would also depend on the interaction or compatibility


of the semantic properties of the verb and the interpretation of the sentence as a whole.

It is, therefore, interesting to note that the most commonly quoted studies which have

found a difference between actives and passives in sentence-picture matching tasks

have predominantly used verbs with no specific effect on the Object. Gough (1965,

1966) used only motion and contact verbs (kick, hit). Slobin (1966) found an active/

passive difference only for reversible sentences where, judging from the examples

given, most of the verbs did not involve a specific effect. He found no active/passive

difference for non-reversible sentences and concluded that “. . . non-reversibility

eliminates the problem of passivity, with kernels and passives roughly equal; in fact,

the passive form generally tended to be somewhat easier.” (Slobin 1966, p. 226).

However, the verbs in the non-reversible condition appear to be predominantly verbs

that do imply a change of state for the Object, and it might be the type of verb rather

than non-reversibility that “eliminates the problem of passivity.” The fact that we

failed to find any active/passive difference for action verbs might well be due to the

semantic properties of the verbs we used.

Thus, it would seem that this task can be used to test sentence parsing or ‘early’

comprehension processes only if we carefully control for a number of semantic, prag-

matic or conceptual factors. Alternatively, we might use this task only to investigate

‘later’ stages of processing and the relationship between language and picture recog-

nition. Our interpretation of verb-specific effects suggests that some normal reversal

errors are not errors of sentence interpretation but reflect processes taking place afer

the sentence is interpreted, in the translation between sentence interpretation and

picture recognition. In the next section, we will discuss some data from aphasic

patients which indicate that poor performance on this task can be due to several

different reasons.

APHASIC PERFORMANCE

Method

Procedure and materials Three aphasic patients, AR, EM, and LC, were tested on the task described above.

The same stimuli and procedure were used, and each patient was tested on both the

auditory and the written version of the task. Patients AR and LC were tested twice

to examine the consistency of the results. Patient EM was tested only once as she had

shown consistent performance on a similar sentence comprehension task over seven

consecutive measures. Each patient was re-tested after periods of therapy (see Byng

el al. submitted).


Subjects

Patient AR (see also Nickels et al. in press) suffered a CVA in 1985 at the age

of 65, shortly after retiring from his job as a building surveyor. He had a left hemis-

phere infarction which resulted in a dense right hemiplegia with severe receptive and

expressive dysphasia. At the time of testing, the severity of his dysphasia had

decreased somewhat and he had moderately good functional comprehension if

pragmatic and contextual cues were available. His output was extremely limited and

mainly consisted of single words and stereotyped utterances. On the Boston Diagnostic

Aphasia Examination (B.D.A.E., Goodglass and Kaplan 1972, 1982) Rating Scale

Profile of Speech Characteristics, AR falls in the range for Broca’s type aphasics, with

a severity rating of 1, indicating a profound impairment of communication.

Patient EM suffered a left hemisphere CVA in 1987 at the age of 75. Until retire-

ment she had worked as a clerk for an assurance society. Her stroke resulted in a dense

right hemiplegia, severe receptive and expressive dysphasia with articulatory and oral

dyspraxia leading to a complete loss of spoken output. At the time of testing, EM had

fair functional communication but little spoken output. She falls within the range for

Broca’s aphasics on the B.D.A.E. Rating Scale Profile and has a severity rating of 1.

Patient LC was working as a secretary when she suffered a left hemisphere CVA

in 1971 at the age of 44, resulting in a dense right hemiplegia and severe dysphasia.

In August 1980 there is reported to have been a slight further episode causing her

dysphasia to worsen temporarily. At the time of testing, LC had good functional

comprehension but extremely limited output, although she communicated a great deal

through the use of non-verbal cues. She also falls within the B.D.A.E. range for

Broca’s aphasics and has a severity rating of 1.

The three patients’ performances on a range of language tests are summarized in

Table 1 (see also Byng et al. submitted). All three patients were asked to tell the story

of Cinderella and their samples were analysed according to the procedure of Saffran

et al. (1989). They all showed a pattern corresponding to that of the agrammatic group

in Saffran ef al.‘s study. Patients EM and LC fell at the lower end of the range. A

ctualitative analysis of their samples was also carried out using the method of Byng

and Black (1989). As Table 2 shows, the patients’ ability to produce complete predicate-

argument structures was very limited. Patient AR relied primarily on single phrases,

though he also produced a lot of incomplete utterances such as ‘Cinderella is . . ‘.

Patients EM and LC mainly produced single words and showed almost no ability to

combine phrases of any type in their output.

The patients’ ability to produce verbs was tested using a picture-naming task in

which half of the pictures represented a single object, and half represented an event.

The nouns and verbs involved were phonologically and orthographically identical,

e.g. hummer (N) and hummer (V). All three patients could retrieve nouns more easily


TABLE 1 Patients’ Language Impairments

Tests

Lexical decision*

Proportion correct Patients

AR EM LC

Auditory presentation Visual presentation

Picture naming

Spoken Written

Comprehension of single words - word-picture matching (with semantic and visual distracters)*

Auditory presentation 0.95 0.88 0.90 Written presentation 0.63 0.92 0.82

Comprehension of sentences

T.R.O.G. (Bishop 1982) Auditory presentation

Comprehension of pictures (Pyramids and Palm Trees Test, Howard and Patterson)

0.72 0.47 0.58

0.81 0.92 0.85

Oral reading

High imag., high freq. High imag., low freq. Low imag., high freq. Low imag., low freq. Non-words

0.9 0.25 0.45

0.65 0.15 0.3 0.45 0.15 0.05 0.55 0.05 0.05

05 0 0

0.94 0.65 0.62 Repetition

Auditorv short-term memorv

Repetition Span - digits 2 2 2 Matching Span - digits -5 4 5

0.93 0.91 0.81 0.84 0.86 0.85

Unable Unable Unable

* Stimuli taken from PALPA (Kay ef al. in press).


TABLE 2

Measures of Structural Quality of Sentence Production

Proportion correct

Patients

Type of structure AR EM LC

Single phrase 0.53 0.94 1.00

Noun phrase + ‘be’ 0.32 0 0

Verb plus 1 argument 0.03 0 0

Verb plus 2 arguments 0.09 0 0

Verb plus 3 arguments 0 0 0

Arguments minus functions 0.01 0.05 0

Embedded sentences 0 0 0

Total utterances analysed 34 18 12

Verb/noun naming

Nouns

Verbs

0.40 0.29 0.42

0.27 0.08 0.11

than verbs, though only EM and LC were significantly more impaired on verb retrieval

(see Byng et al. submitted).

Results

The patients’ performance on the sentence-picture matching task is shown in Table

3. The performances of LC and EM are never above chance either overall or in any

one category. The performance of AR is never above chance overall and is not con-

sistently above chance over the two presentations in any category. All three patients

make significantly more reverse role errors than lexical distractor errors (all comparisons

p > 0.000 1, Binomial Test).

None of the patients showed any significant differences between performance on

auditory and visual presentation either in terms of total scores or category by category

comparisons (all comparisons p >0.05, McNemar’s Test).

Similar results were obtained when the patients were re-tested on this task after

periods of therapy. Their overall performance remained the same, though AR’s per-

formance on action active sentences had consistently improved.

Given these results and the similarities in their output, these three patients could be

said to suffer from similar sentence processing deficits. However, further input testing


TABLE 3

Tests

Reversible sentence comerehension test

Proportion correct

Patients

AR EM LC

Presentation

I II I I II

Auditory presentation

Active action

Active non-action

Passive action

Passive non-action

Adjectives

Locatives

0.5 0.7 0.4 0.4 0.6

0.8 0.7 0.6 0.5 0.5

0.6 0.5 0.4 0.7 0.5

0.5 0.2 0.4 0.5 0.6

0.6 0.7 0.6 0.6 0.7

0.4 0.8 0.7 0.6 0.4

strongly suggests that they have quite different problems which could not be clearly

distinguished on the basis of the sentence-picture matching task.

DIFFERENTIATING THE DEFICITS

As we have discussed in Section 1, different processing ‘stages’ are involved in

performing the sentence-picture matching task. Correct interpretation of the input

sentence depends, in the first place, on encoding and retaining the order of the noun

phrases around the verb (stages (i) and/or (ii)), so we can, first of all, ask whether

patients like AR, EM and LC, who fail on this task, can do this. To investigate this

issue, we tested the three patients on the following tasks.

Task 1: Picture Selection of Two Characters

The patients were asked to point to two characters in the order in which they were

presented, after hearing two names, e.g. ‘cowboy, nun’. The pictures were visible

during presentation of the names. The characters, taken from those used in the

sentence-picture matching task, were presented in a vertical array of three and were

not involved in any action.

Task 2: Picture Selection of Two Characters, with Filler Noun

In this task, the length of the incoming stimulus was increased by adding an unrelated

noun, e.g. ‘nun, cup, astronaut’. The filler noun was not depicted in the three pictures.


Task 3: Picture Selection of Two Characters, with Reversible Sentence

The patients were asked to point to two of the three characters depicted in the same arrangement as in the previous tasks, but this time the two names were presented as part of a reversible sentence, e.g. ‘The cowboy washes the boxer’. The pictures provided for matching did not depict this event, but comprised three individual characters, one of which had to be rejected and the other two pointed to in the order of presentation, as in Task 1.

Task 4: Reversible Sentence Matching

The patients were asked to judge whether two spoken sentences were the same or not. The sentences used were those from the sentence-picture matching task and were made different by simply changing the order of the two noun phrases, e.g. the nun splashes the queen, the queen splashes the nun; the vicar washes the sailor, the vicar

washes the sailor.

Patients were tested twice on Tasks l-3, and once on Task 4. Their performance is shown on Table 4. Patient AR’s performance indicates that he can encode the order

TABLE 4

AR EM LC

Task 1 Selection of two characters

Task 2 Selection of two characters - filler noun inserted

Task 3 Selection of two characters - given a sentence

Task 4 Reversible sentence matching

Task 5 Sentence anomaly judgements* e.g. the baby drops the toy; the toy drops the baby

0.4 0.83 0.83

N/A 0.14 0.57

1.00 0.74 0.72

0.94 0.36 0.82

0.76 0.6 0.24

NOTE: In the interests of ease of comparability all scores in this table have been corrected for chance. *Derived from Saffran er al.


of two noun phrases when these are in a sentence (Task 3), and that he can encode and maintain order information even when two sentences have to be compared (Task 4). But he cannot retain the order of phrases when the phrases are part of an unstructured list (Tasks 1 and 2). Patient EM can encode the order of two noun phrases in either a list form or in a sentence (Tasks 1 and 3) but her performance is significantly affected by additional material which cannot be integrated into a structured representation (Task 2) @ > 0.001, Fisher Exact Test). Although she can encode order information, she does not appear to be able to maintain that information when a comparison between sentences is required (Task 4). No strong conclusions can be drawn about LC’s performance since it was erratic over testing sessions; she appears to be able to encode the order of two noun phrases but it is not clear that she can encode and maintain this information when list length is increased and/or when the two noun phrases occur in a sentence (Tasks 2 and 3).

Given these results, it is unlikely that AR’s and EM’s failure on the sentence- picture matching task was due to an inability to encode linear order for the purposes of sentence comprehension. In LC’s case, however, we cannot exclude that such an impairment might be one of the factors contributing to her failure on that task.

To confirm these conclusions and further examine the patients’ sentence comprehension abilities, we tested them on an anomaly judgement task.

Task 5: Sentence Anomaly Judgement#

The patients had to decide whether a sentence they heard made sense or not. The sentences were made anomalous by reversing irreversible sentences, e.g. the baby

drops the toy, the toy drops the baby. Seventeen out of 20 of the verbs used expressed actions. Twenty well-formed and 20 anomalous sentences were presented in random order. Anomalous sentences were never presented on the same occasion as their correct counterpart.

In principle this task involves the same processing ‘stages’ as the sentence-picture matching task, at least in terms of sentence comprehension. Subjects will not achieve adequate performance if the order of noun phrases is not encoded correctly, if information about thematic roles is not retrieved and mapped onto the correct positions, and if the semantic content of the phrases in particular positions is not processed and evaluated correctly. However, it might be argued that, in this task, a complete semantic representation for the whole sentence need not be constructed since the anomaly could be detected by simply combining and evaluating the meanings of the first noun phrase and the verb, or the meanings of the verb and the second noun phrase. In other words, it might be possible to short-cut processing in ways that are not possible in the sentence-picture matching task. Although this is a possibility for


some sentences, several of the sentences we used cannot be judged correctly on that basis since neither NP +V nor V +NP are in themselves anomalous or implausible, e.g. the ray drops rhe baby. Relatively good performance on these sentences should, therefore, indicate that subjects are not using such a short-cut strategy.

We found that both AR and EM were relatively unimpaired on this task as a whole, and were 100% correct on the sentences where partial evaluation would have led to an incorrect judgement (see Table 4). For AR, this result is in line with his performance on the active action sentences of the sentence-picture matching task, when he was retested after a period of therapy. Patient EM’s better performance on this task, on the other hand, implies that her failure on the sentence-picture matching task might be primarily due to an impairment at those processing levels that are not involved in the judgement task. In this respect, the sentence-picture matching task is more akin to the sentence matching task we have discussed above (Task 4), since in both cases two or more structured representations have to be maintained and compared. This is what EM does not seem to be able to do. These results confirm that neither AR nor EM have a significant problem in encoding order information as performance on this task would also be significantly affected by such a problem.

Patient LC’s performance is very poor on this task too. As we have already pointed out, she might not be able to encode order reliably, particularly when it is the order of elements in a structured representation, as opposed to the order of elements in a list. This inability would also affect her performance on the judgement task.

From the testing we have carried out, there is no evidence that AR and EM have any major deficit in constructing an ordered and structured representation of input sentences [see the first section, stages (i-ii)]. However, both patients might have difficulty in accessing information about the argument structure of verbs, including thematic role information, and might therefore not be able to map thematic roles onto a structural representation of the input. This might well be the only reason why AR fails on the sentence-picture matching task, particularly on those sections of the task where lexical access has not been facilitated by therapy (see Nickels et al. in press; Byng et al. submitted). Patient EM’s main problem on this task, on the other hand, might stem from a general difficulty in holding and/or comparing representations. Finally, LC might have a number of problems at different processing levels, including one at ‘early’ stages of processing.

Additional testing of the three patients also shows that there are differences in their ability to recognise pictures representing an event or state, although their ability to recognise pictures depicting single objects does not appear to differ significantly (see Table 1). The patients were presented with a set of 40 pictures. In half of the pictures an event was happening, and in the other half nothing was happening. For example, in an ‘event’ picture a car might be seen driving down a road, whereas a ‘non-event’ picture might comprise a street scene with nothing happening in it. Half of the event


pictures had an animate agent and half an inanimate one. The patients were asked to divide the pictures into those where something was happening and those where nothing was happening. Patients AR and EM had no difficulty with this task, but LC scored 53% correct (not significantly above chance, Binomial Test). So, LC’s poor performance on the sentence-picture matching task might also arise from a problem in picture recognition and/or interpretation,

CONCLUSION

The data from normal and aphasic performances on a sentence-picture matching task show that a number of different factors can affect accuracy. In the case of normal subjects, some reversal errors do not seem to be errors of sentence comprehension: they appear to reflect ‘later’ stages of processing which mediate the translation between language and pictures. The semantic or conceptual properties of the verbs used seem to affect performance, particularly when they interact with the meaning of the whole sentence. We have suggested that, in the process of understanding a sentence and selecting a picture, a number of casual inferences are drawn which are based on the meaning of the verb and the relative causality attributed to each of its arguments. These inferences then have to be combined or reconciled with other causal inferences based on the interpretation of the sentence as a whole.

In the case of aphasic subjects, errors might arise from different sources and have multiple causes. The three patients we have studied all show similar linguistic symptoms and can all be classified as ‘Broca’s aphasics’. Yet their poor performance on the sentence-picture matching task does not seem to stem from the same underlying deficit, since each of them shows a different pattern of performance on other tests which tap related aspects of processing. Therefore, great caution should be exercised in drawing conclusions from performance on this task. Even if the sentences used are simple, the processing involved is complex.

ACKNOWLEDGEMENTS

We thank AR, EM and LC and their relatives for their willingness to participate in this study. We are grateful to Chris Heron and Elaine Giles for putting us in touch with the patients. Georgia Vaux drew the pictures for the sentence-picture matching task. We also thank Max Coltheart for his assistance in planning this study, which was sup- ported by MRC Project Grant No. G8514094N to Prof. M. Coltheart and Dr S. Byng.

NOTES

1. For a review of studies of sentence comprehension in aphasia, see Berndt (in press).

2.

3.

4.

5.

6.


Depending on one’s theoretical assumptions, the processing of passive sentences could involve more complex processing at the syntactic level and/or in terms of the mapping of thematic roles. However, we will not deal with these issues as they are not relevant to the findings reported in this paper. The label ‘Theme/Patient’ has been used in this paper simply because both terms have been employed in the literature to refer to the thematic role assigned to the Object of an action verb. Neither our interpretation of the task nor our data analysis is crucially dependent on a specific definition of these terms or on a particular theory of thematic roles. These verbs, and some of the adjectives in (v), can also have an action interpretation. However, the possibility of such an interpretation does not seem to have significantly affected subjects’ performance (see Results section). Many of the verbs with causal attributions to the Theme/Patient seem to fall within the category that Pinker (1990) labels “cause-focus” verbs, i.e. verbs that encode information “that the event was caused by some antecedent agent-patient relation.” So, perhaps the “focus” of the verb also enters into these causal attributions. We are grateful to Dr E. Saffran for making these materials available to us.

REFERENCES

Au, T. K. F. 1986 “A Verb is Worth a Thousand Words: the Causes and Consequences of

Interpersonal Events Implicit in Language,” Journul of Memory and

Language 25. 104-22. Belletti, A. and L. Rizzi

1988 “Psych-Verbs and &Theory,” Natural Language and Linguistic Theory

6. 291-352.

Berndt, R. S. n.d. “Sentence Processing in Aphasia,” in Acquired Aphasia, M. T. Sarno

(ed.), 2nd edn. New York: Academic Press. Brown, R. and D. Fish

1983 “The Psychological Causality Implicit in Language,” Cognition 14.

231-73.

Wng, S. 1988 “Sentence Processing Deficits: Theory and Therapy,” Cognitive Neuro-

psychology 5. 629-76.

Byng, S. and M. Black 1989 “Some Aspects of Sentence Production in Aphasia,” Aphasiology 3.

241-63.

loo Journal of Neurolinguistics, Volume 6, Number 2 (1991)

Byng, S., L. Nickels and M. Black n.d. “Replicating Therapy for Mapping Deficits: Remapping the Deficit.”

Caramazza, A., E. Grober, K. Garvey and J. Yates 1977 “Comprehension of Anaphoric Pronouns,” Journal of Verbal Learning

and Verbal Behavior 16. 601-9. Carrithers, C.

1989 “Syntactic Complexity Does Not Necessarily Make Sentences Harder to Understand,” Journal of Psycholinguistic Research 18(l). 75-87.

Chomsky, N. 1985 Knowledge of Language, Its Nature, Origin and Use. New York: Praeger.

Corrigan, R. 1988 “Who Dun It? The Influence of Actor-Patient Animacy and Type of Verb

in the Making of Causal Attributions,” Journal of Memory and Language 27. 447-65.

Francis, W. N. and H. Kucera 1982 Computational Analysis of Present-day American English. Providence,

RI: Brown University Press. Goodglass, H. and E. Kaplan

1972 77re Assessment of Aphasia and Related Disorders. Philadelphia: Lea and Febiger.

1982 The Assessment of Aphasia and Related Disorders, 2nd Edn. Philadelphia: Lea and Febiger.

Gough, P. B. 1965 “The Verification of Sentences, ” Journal of Verbal Learning and Verbal

Behaviour 4. 107-l 1. 1966 “The Verification of Sentences: the Effects of Delay of Evidence and

Sentence Length,” Journal of Verbal Learning and Verbal Behavior 5. 492-96.

Howard, D. and K. Patterson n.d. Pyramids and Palm Trees (unpublished).

Kay, J., R. Lesser and M. Coltheart n.d. Psycholinguistic Assessment of Language Processing Abilities (PALPA).

London: Lawrence Erlbaum Associates (in press).

Keenan, E. 0. 1985 “Passive in the World’s Languages,” in Language Typology and

Syntactic Description I: Clause Structure, pp. 243-81, T. Shopen (ed.). New York: Cambridge University Press.

Nickels, L., S. Byng and M. Black n.d. “Sentence Processing Deficits: a Replication of Therapy,” British Journal

of Disorders of Communication (in press).


Pinker, S.

1990 Learnability and Cognition: 7he Acquisition of Argument Structure,

Cambridge, MA: The MIT Press.

Saffran, E., R. S. Bemdt and M. Schwartz

1989 “The Quantitative Analysis of Agrammatic Production: Procedure and

Data,” Brain and Language 37. 440-80. Schwartz, M. F., M. C. Linebarger and E. M. Saffran

1985 “The Status of the Syntactic Theory of Agrammatism,” in Agrummatism,

pp. 83-124, M.-L. Kean (ed.). London: Academic Press.

Slobin, D. A.

1966 “Grammatical Transformations and Sentence Comprehension in Child-

hood and Adulthood,” Journul of Verbal Learning and Verbal Behavior 5. 219-27.

Sudhalter, V. and M. D. S. Braine

1985 “How Does Comprehension of Passives Develop? A Comparison of

Actional and Experiential Verbs,” Journal of Child Language 12.455-70.

Tyler, L. K.

1987 “Spoken Language Comprehension in Aphasia: a Real-time Processing

Perspective,” in The Cognitive Neuropsychology of Language, pp. 145-

62, M. Coltheart, G. Sartori and R. Job (eds). London: Lawrence Erlbaum

Associates.

Williams, E. S.

1981 “Argument Structure and Morphology,” i%e Linguistic Review 1. 81-114.

patterns of sentence processing deficit: processing simple sentences can be a complex matter

Documents