The Capacity Constrained Parsing Model

Macdonald et al. (1992) propose a model (the Capacity Constrained Parsing Model) in which they take up an argument often raised in criticising multiple analysis models, i.e. the argument that multiple analyses create too great a short-term memory burden to be viable. They propose that, when faced with a TASS, every individual initially represents two analyses, and that the individual’s working memory capacity determines the duration for which both analyses are maintained. Individuals with a high capacity working memory are assumed to maintain both analyses for a longer time than

individuals with a low capacity working memory, who will abandon one analysis, the less activated one, more quickly. Thus, subjects with high working memory capacity should perform worse in processing sentences with syntactic ambiguities than subjects with low working memory capacity, because they maintain multiple analyses for longer. They suggest that there is a double cost for maintaining multiple analyses for a long time: firstly, the authors simply assume that maintaining multiple analyses is more costly than maintaining just one analysis; secondly, they propose that maintaining multiple analyses uses up resources required to execute other processes of sentence

comprehension, such as constructing a referential representation of the sentence. Such higher-level processes have to be postponed and time-consuming catching up occurs towards the end of the sentence.

An individual’s working memory capacity was defined operationally as her score on the "reading span task" in which subjects read aloud a set of unrelated sentences and had to recall the last word of each sentence of the set. The number of items in a set ranged from 2 up to 6 or more sentences, depending on the subject’s performance. Low-span subjects recalled an average of 2.5 or less words and high-span subjects an average of 3.5 or more words. It is not clear what the reading span task measures. The

"rationale behind the test is that the comprehension processes used in reading the sentences should consume less of the working memory resources of high span readers. These readers would thus have more capacity left to hold the final words of the sentences" (Just & Carpenter, 1992, p. 125).

Macdonald et al. (1992) assume that the reading span measure reflects quite different processes than the digit span measure or the ability to recall a list of unrelated words. It seems somewhat surprising, however, that reading span and digit span should be completely unrelated: if a subject had a very low digit span, she would be unlikely to remember many of the sentence-final words in the reading span task, however good her sentence comprehension processes might be. The authors report high positive correlations with other measures of reading comprehension. It is unclear why Reading Span correlates with sentence comprehension measures but Digit Span does not, given that both probably involve the use of the articulatory loop. Such a correlation is also surprising given neuropsychological evidence that short term memory deficits as assessed by the Digit Span measure do not interfere with sentence comprehension (Shallice, 1988).

Macdonald et al. (1992) tested their model with an experiment involving temporarily ambiguous sentences containing a main verb whose past tense form and past participle form are identical thus allowing both a main verb analysis (example a) and a reduced- relative clause analysis (example b):

a) The experienced soldiers / warned about the dangers / before the midnight / raid.

b) The experienced soldiers / warned about the dangers / conducted the midnight / raid.

(Slashes indicate regions used in data analysis). Unambiguous control sentences for a) contained a verb which allowed only the main-clause analysis either because its past tense form is different from its past participle form or because it is strictly intransitive (eg. ’giggled’). Control sentences for b) were created by adding a relative pronoun and an auxiliary. Examples a’) and b’) serve as illustrations:

a’) The experienced soldiers spoke about the dangers before the midnight raid. b’) The experienced soldiers who were told about the dangers conducted the midnight raid.

Subjects read sentences presented on a computer screen using a word-by-word non- cumulative self-paced reading paradigm (moving window paradigm).

Let us first focus on sentences such as a) above, in which the main clause analysis turns out to be correct. There were no effects in the first analysed region (eg. ’warned about the dangers’). In the next region, there was an ambiguity effect for all subjects, which the authors interpreted as indicating that both subject groups built up multiple analyses in the ambiguous sentences. On the last word (eg. ’raid’) high-span subjects took longer than low span subjects in the ambiguous as compared to the unambiguous sentences. The authors conclude from this that low span subjects had given up one of the multiple analyses, whereas high span subjects still maintained both. Furthermore, high-span subjects’ reading times for ambiguous (but not uambiguous) sentences increased throughout the regions and were highest at the last word, which the authors attribute to the subject having to catch up with higher-level processes which were delayed because multiple analyses had been maintained.

correct (such as b) above). The authors report an interaction of Ambiguity and Region (ambiguous sentences produced longer reading times for regions 2 and 3, but not for region 1). Interestingly, however, there was no interaction of Ambiguity with Reading Span.

Thus, the effect of reading span seems to have been limited to the main clause

sentences where a high reading span led to longer reading times on the sentence-final word.

It is somewhat surprising that there is never an increase in reading times in the region which contains the ambiguous verb. If multiple analyses are maintained, then surely they should be maintained immediately after the ambiguous verb occurs, and the processing cost, which the authors claim is caused by maintaining multiple analyses, should be measurable as soon as the ambiguous verb occurs (or at least by the time three more words are encountered, i.e. by the end of the region in which the

ambiguous verb occurred). The authors do not address why this burden is not evident earlier, i.e. before the fourth word following the ambiguous verb.

The ambiguity effect in region 2 reported for the ambiguous main verb sentences for both subject groups seems to me to be the piece of evidence potentially most

damaging to a single analysis model. A single analysis model predicts that only the (preferred) main clause analysis is maintained, and since there is no disambiguating information in this region, no extra processing should be required. The multiple- analysis model, however, assumes that maintaining two analyses (even if one is activated to a smaller degree than the other) creates a "burden" which should show up as increased reading times.

Macdonald et al. (1992) report a follow up experiment (in which only high span subjects participated) in which only ambiguous main verb sentences (such as a) above) were presented, but no ambiguous relative clause sentences (such as b) above). This experiment was conducted to rule out the possibility that subjects had been induced to generate multiple analyses upon reading the first verb because of the presence of the

ambiguous relative clause sentences in which the unpreferred analysis turns out to be correct. Again, reading times in region 2 (eg. ’before the midnight’, see example a) above) were longer in the ambiguous than the corresponding unambiguous sentences.

However, there seems to be still another alternative interpretation of the increase in reading times in region 2 of the ambiguous as compared to nonambiguous main verb sentences. The increase might reflect comprehension processes which were induced by the questions asked after each sentence. Note that these yes/no questions were designed "to assess which interpretation was assigned to the first verb" (Macdonald et al., 1992, p. 64). The sentence ’The experienced soldier warned about the dangers before the midnight raid’, for example, was followed by the question; ’Did someone tell the soldiers about the dangers?’ for which (presumably) the correct answer is ’no’, since ’warned’ was here used as a transitive verb, i.e. the soldiers warned somebody else rather than being warned themselves. Such questions might have alerted the subjects (especially the high span readers who can be assumed to have good verbal skills!) to the possibility that ’warned’ could be analysed as a past participle introducing a reduced relative clause. Other studies typically ask simple yes/no questions on about 25 % of the trials mainly to ensure that subjects pay attention to the sentences. These questions do not normally focus on the critical part of the sentence. Thus, it is possible that the multiple analyses in region 2 might have been maintained because of the potentially suggestive comprehension questions after each trial.

One has to note, too, that the effect in region 2 was small (17 ms on average) and that only an analysis by- subjects but not by-items was reported.

The authors dismiss earlier evidence in favour of single analysis models by claiming that these studies must have used a sample of subjects who happened to be medium- or low- span readers: "such readers are more capacity limited and, consequently, will conform more closely to the predictions of a single representation model" (Macdonald et al., 1992, p. 89). The authors do not discuss, however, why their own low span readers’ performance in region 2 of the main-clause sentences supported the multiple analysis model, whereas this was not the case in other researchers’ samples.

In conclusion, it seems safe to accept the authors’ summary that "there is a cost incurred in processing ambiguous sentences" (Macdonald et al., 1992, p. 76). There is no clear evidence, however, that reading span determines whether or not an ambiguity effect occurs. Neither can one safely conclude that the increase in reading times for ambiguous sentences indicates extra processing costs ensuing from multiple analyses; these increased reading times might merely reflect experimentally induced strategies of generating multiple analyses in order better to be able to answer the comprehension questions following each trial.

The capacity constrained parsing model has also been tested in a study focussing on the interaction of syntactic and nonsyntactic information which was modelled on Ferreira & Clifton’s (1986) experiment. This study will be reported in section 7.4.