Sei sulla pagina 1di 8

"Compare and contrast the Garden Path Model and the Constraint Based Model of sentence processing.

What are their relative strengths and weaknesses?"

Dionne Angela Donnelly Module Code: PSYC214 Word Count: 1786

Sentence processing (or parsing) is thought to be best examined by assessing our performance on understanding ambiguous sentences. This is usually tested by comparing eye movements on ambiguous and non-ambiguous sentences. Two opposing models of parsing have emerged from such analyses. These are the garden-path model (GPM) which is an autonomous, two-stage model of parsing, and the constraint based model (CBM) which is an interactive one-stage model. There is an abundance of research on these two models, which both support or contest each one depending on their individual strengths and weaknesses. This research points to both theories being at least partially correct, but the GPM being the better theory of the two.

The GPM was proposed by Frazier (1979, 1987; Frazier & Rayner, 1982, cited in Jay, 2003), and it dictates that sentence processing is an autonomous two-stage process. It is autonomous because the first stage (or pass) involves making initial attachments based on syntactic processing only, and only one (the simplest) syntactical structure is considered at a time (Eysenck & Keane, 2005). This stage involves attachments being made based on two of Kimballs (1973, cited in Garman, 1990) principles: Minimal attachment and late closure. The minimal attachment principle involves choosing the structure which when adding new lexical items to the sentence leads to the fewest nodes (Jay, 2003). Late closure involves new words being attached to the current phrase/clause if it is grammatically sound (Eysenck & Keane, 2005). Of these two principles, minimal attachment takes precedence (Harley, 2001). The GPM is deemed a two-stage process because if the sentence to be parsed is ambiguous and requires a second pass, pragmatic, semantic and thematic information is used to revise the parse tree (Harley, 2001). In contrast to the above model, the CBM (MacDonald, Pearlmutter & Seidenberg, 1994, cited in Eysenck & Keane, 2005) states that sentence processing is a one-stage process involving an interaction of semantic, syntactic, 2

discourse, and frequency-based information (also referred to as constraints) (Harley, 2001). All of these are available to the parser immediately, with all possible interpretations being available at the same time, and these are ordered depending upon their strength of activation (Eysenck & Keane, 2005, pg. 364). This strength is determined by the number of constraints which support each structure. Garden-pathing is thought to arise when the correct structure receives less activation that an incorrect structure (Eysenck & Keane, 2005; Harley, 2001).

Research which provided support for the principles of the GPM (and which therefore opposes the CBM) was conducted by Rayner and Frazier (1987), who found (using eye movement data) that processing was faster for nonminimal attachment sentences that contained complementisers such as that in order to reduce ambiguity, than for sentences that did not contain complementisers. For example in the sentences: (1) The criminal confessed that his sins harmed many people (2) The criminal confessed his sins harmed many people Subjects, using the principle of minimal attachment, found (1) to be unambiguous as the parser sees his sins as the subject of the complement harmed many people (Pine, 2010), whereas (2) resulted in a garden path when the reader reached harmed (the disambiguating region) as the reader treats his sins as the direct object of the verb confessed (Pine, 2010). This study shows that without such complementisers we read past the clause boundary and fail to close the clause correctly (Jay, 2003). An example of a similar study which instead provides support for the CBM was conducted by Trueswell, Tanenhaus & Kello (1993). They compared reading times and eye movement data for parsing of garden path sentences that made either direct object reading more or 3

less likely (Pine, 2010). The results showed that the complementiser effect seen above was due to how strongly a verb is associated with the that complementiser, and was not because of syntactic misanalysis as proposed by the GPM.

The principle of late closure has also been supported by research, which again contradicts the predictions of the CBM. Frazier and Rayner (1982, cited in Pine, 2010), studied the reading times for sentences which were either consistent or inconsistent with late closure, for example: (3) Since Jay always jogs a mile and a half this seems a short distance to him. (4) Since Jay always jogs a mile and a half seems a short distance to him. Sentence (3) requires reader to keep VP jogs open and attach NP a mile and a half, which according to the GPM model should be easy for the parser. Sentence (4) requires reader to close VP after jogs which should have caused a garden path effect at seems (Pine, 2010). Reading times were longer for (4), which indicates that the parser spent longer attempting to process the sentence. This contradicts the predictions of the CBM as if all the possible constraints were considered simultaneously, reading times would have been the same for both the ambiguous and the unambiguous sentences.

One of the major contrasts between the CBM and the GPM which has received a considerable amount of interest from researchers is whether semantic information is used by the parser in the first pass. The GPM argues that no semantic information is used. This has received support from studies such as the one conducted by Ferreira & Clifton (1986). They compared eye movement for 4

sentences in which minimal attachment would lead to an ambiguous or unambiguous interpretation. They found that there was a bias for syntactic processing even when such sentences did not make thematic sense and the reader had to reanalyse the whether the verb should have been a main verb or a verb of reduced relative. One again, similar research by Truesman, Tanenhaus and Garnsey (1994) yielded a different conclusion, in favour of the CBM. They used the sentences: (5) The author read by the student was very difficult to understand (6) The book read by the student was very difficult to understand. Here, the GPM would state that a garden-path effect should occur regardless of the meaning of the author or book which would change whether the verb read would become a main verb or a reduced relative. However, Truesman, Tanenhaus and Garnsey found that when a semantic constraint is strongly activated, reduced relatives are just as easy to process as any other structure. This contradicts the idea that the parser would have to backtrack and reanalyse read due to its ambiguous syntactical structure. Furthermore, the CBM predicts that eye movement studies will show how the ambiguity of part of a sentence affects processing, until there is enough contextual information to allow perceptual suppression of the other possible interpretations (Mackay, 1970, cited in Garman, 1990, pg. 359). Support for the CBM comes from a study by Pickering and Traxler (1998, cited in Eysenck & Keane, 2005), which the following sentences were used: (7) As the woman edited the magazine amused the reporters. (8) As the woman sailed the magazine amused the reporters. As these sentences are syntactically the same, if interpreted using the GPM they would both cause the parser the same difficulty. But, if interpreted by the CBM this would not be the case due to the 5

semantic bias toward the wrong structure in sentence (1). Eye movement analysis showed that more regressions and longer fixations in the disambiguating region whilst reading sentence (1). This could be interpreted as showing that semantic information was available to the parser in order for them to parse (8) more easily than they could parse (7), and this clearly contradicts the predictions of the GPM. However, Ferreira and Clifton (n.d., cited in Harey, 2001) argued that semantic information does not help to either prevent or cause the garden-path effect, but instead aids recovery from it. Nonetheless, CBM proponents argue that contexts are never semantically neutral, we always have prior knowledge and expectations of what is likely to appear next in a sentence or what words are likely to appear together (Harley, 2001).

Despite the above research showing a similar amount of support for each model, there are differences between the two models regarding their theoretical basis. The CBM has a number of theoretical problems which mean that it is almost impossible to test scientifically. Unlike the GPM, the CBM is very vague and offers so many potential constraints that it is almost impossible for them all to be wrong. Therefore, the CBM should make clearer predictions about which constraints are most salient to the parser and if there are any exceptions to such a rule. The GPM makes clear predictions that syntactic structure is most important to the parser, and such predictions can be clearly supported or refuted by the evidence. In contrast to this, researchers aiming to support the CBM can claim almost any results fit the predictions of the model due to their vagueness. Also, the CBM should be more specific (as the GPM is) about when interaction effects occur (Pine, 2010) as it is probable that all constraints are not needed at all times as this would increase working memory demands (Pine, 2010). The GPM accounts for this problem as the parser minimises demand by using the principles of minimal attachment and late closure. 6

These generalisations usually work because we only rarely encounter ambiguous structures that are not the correct syntactic structure. If the CBM were more specific, it would allow it to be tested in its own right, rather than by using procedures designed to test the GPM (Pine, 2010). Furthermore, the CBM does not make clear whether constraint information is present immediately or if it only appears soon after reading begins. In relation to this, the semantic effects seen in the above research by Truesman, Tanenhaus and Garnsey (1994) and Pickering and Traxler (1998, cited in Eysenck & Keane, 2005) may just show that the first stage of the GPM is extremely short or difficult to identify using the above procedures (Pine, 2010).

In conclusion, the above empirical research does not indicate whether one model is better at explaining parsing of ambiguous structures than the other. It only shows that they both receive similar amounts of support depending on the types of sentences/syntactical structures used. Garrett (1970, cited in Garman, 1990) reviewed the evidence and concluded that there was enough empirical support for the validity of both models. He also observed that the research which supports the CBM predominantly involves tasks that facilitate immediate processing, and the studies which provide support for the GPM require judgements subsequent to the sentence being processed and stored in memory. Despite this, the consideration of the theoretical difficulties of the CBM leads one to believe that whilst both models may be are partially correct, the GPM is the more theoretically sound and scientifically testable theory of the two.

References Eysenck, M. W. & Keane, M. T. (2005). Cognitive Psychology A Students Handbook. Hove: Psychology Press. Garman, M. (1990). Psycholinguistics. Cambridge: Cambridge University Press. Harley, T. (2001). Psychology of Language From Data To Theory (2nd Ed.). Hove: Psychology Press. Jay, T. B. (2003). The Psychology of Language. New Jersey: Prentice Hall. Pine, J. (2010). Autonomous and interactive models of sentence processing. PSYC214 Language & Thought Lecture Slides. Rayner, K. & Frazier, L. (1987). Parsing temporarily ambiguous complements. The Quarterly Journal of Experimental Psychology, 39(4), 657-673. Trueswell, J. C., Tanenhaus, M. K. & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19(3), 528-553.

Potrebbero piacerti anche