Institut des Sciences Cognitives, CNRS UPR 9075
67, Boulevard Pinel, 69675 BRON Cedex, France

A central function in syntactic analysis or syntactic comprehension is the assignment of thematic roles to noun phrases in sentences. In a simplified manner, we can consider that in languages like French and English, there is a default or canonical order in which thematic roles are assigned (e.g. "Agent Object Recipient" in English for the canonical sentence "John gave the ball to Mary.") However, in non-canonical sentences (e.g. "The ball was given to Mary by John."), this ordering is transformed, and thematic role assignment is guided, in part, by function items (e.g. closed class words, including prepositions "to" and "by", or grammatical morphemes).

The objective of this paper is to describe the possible evolutionary development of a capacity for thematic role assignment in syntactic analysis, based on dissociable processing of function and content words. The method will be to first identify a quantifiable behavioral measure of thematic role assignment, and to then propose a plausible model that yields this identified target behavior. The model will be constrained by the requirement that to the largest extent possible its functional components should correspond to behavioral or computational capabilities whose underlying neurophysiology is known. The system’s development will thus represent a minimal cost, largely attributable to the recombination of existing computational capabilities. The model will then be tested against the established behavioral criteria, and will be further validated to predict specific analysis impairments in agrammatic adults, and also the presence of primitive analysis capabilities that should be present in infants. The ensemble of behavioral and simulation results will be discussed in the context of economy in evolution of syntactic analysis capabilities.

The ability to assign thematic roles has been quantified in different clinical tests used to asses agrammatism in aphasia. A well know version developed by Caplan et al (Cognition, 1985) consists of 9 sentence types of varying syntactic complexity, five canonical and four non-canonical, and will serve as our target problem. Five sentences of each type are used for a total of 45 sentences. The sentences are constructed so that no semantic interpretation can contribute to thematic role assignment, which must proceed entirely as guided by function items. Sentences are real aloud to subjects in a pseudo-random order, and after each sentence, the subject should indicate by pointing at photographs "who did what to whom". Interestingly, a rather significant subgroup of aphasics with left-hemisphere lesions that include the peri-sylvian cortex in and around Broca’s area demonstrate a thematic role assignment deficit that is highly selective for non-canonical sentences. The specific vulnerability of this comprehension factor indicates that it is a specialized capability, not directly related to other aspects of syntactic analysis.

In order to realize this thematic role assignment task, a system should first be capable of distinguishing function words (or morphemes) from content words. Numerous behavioral and event-related brain potential studies indicate that indeed, adults process function words and content words in a dissociated manner. The system should also be able to store the content words in a form of working memory, and then to access the contents of this memory in a non-standard order (i.e. different from the input order) guided by the function items. This capability to transform the order of content words, guided by the function words, provides the basis for the assignment of thematic roles.

We have recently developed such a model for the processing of serial surface structure, and abstract transformational rules for non-linguistic sensorimotor sequences (Dominey et al. 1998). The model was developed to simulate human behavior in artificial grammar learning tasks. In these tasks, two isomorphic sequences ABCBAC and DEFEDF have different surface structure but share the same abstract grammatical structure 123213 in which the second triplet is obtained via a systematic transformation of the first. We can thus consider the abstract structure 123213 to be non-canonical, whereas 123123 would be canonical. We demonstrated that while one relatively simple "surface" system (a recurrent network (Dominey et al. 1995)) can learn and reproduce individual sequences, the "abstract" system requires the addition of the working memory processes described above in order to acquire and transfer knowledge of the abstract structure to isomorphic sequences. This resulted in the development of a dual-system model with separate sub-systems for surface and abstract structure respectively. We argue that the functional components – recurrent connections and working memory – exist in the non-human primate and are thus evolutionarily plausible building blocks, and that the important evolutionary step may involve a novel recombination of these capacities.

We have recently demonstrated (Dominey & Ramus 2000) that in its untrained initial state, the "surface" system of the dual stream model is sensitive to the temporal or rhythmic structure of language as demonstrated in newborn infants (Nazzi et al. 1998), as well as to the serial structure and statistical regularities of speechlike sound sequences as demonstrated in 8 month olds (Saffran et al. 1996). Similarly the "abstract" system is capable of learning within 2 minutes to discriminate between sentence with the abstract structures AAB vs ABB vs ABA as demonstrated in 8 month olds (Marcus al. 1999).

We now examine the behavior of the resulting dual-system model in syntactic comprehension of canonical and non-canonical sentences (Dominey 1999). The crucial "evolutionary" modification is that function and content items are treated in dissociated processing streams. Function items are represented by the "surface" system and guide the application of transformational rules that are represented by the "abstract" system. We can demonstrate that after training on a "supervised" version of the Caplan task in which the correct thematic role assignment is provided, the model can then perform the standard task correctly, including the generalization to new sentences. Interestingly, when rendered agrammatic by disruption of the representation of the function items within the recurrent network, the model fails in the processing of transformational structure both for lingustic and non-linguistic sequences. With respect to this prediction, we have recently demonstrated that agrammatic patients fail to process non-canonically ordered sentences both in natural and artificial grammars (Dominey & Lelekov 2000).

Together these observations allow us to propose a neural network model that simulates the combination of simple functions that allow the realization of a reduced form of thematic role assignment in syntactic analysis. The model explains data on normal and aphasic performance, and also predicts performance of infants and aphasics in a novel artificial grammar learning task. While clearly a long way from a complete model of syntactic analysis, this simplified model provides a point of departure for discussions of how syntactic analysis might and might not be implemented. Recurrent cortico-cortical connections provide the basis for context encoding necessary for the representation of syntactic organization of function words. Likewise cortico-subcortical working memory circuits provdie the necessary storage for on-line processing of open class items. Both of these functions have been demonstrated in non-human primates. What they lack is the coordinated cooperation of these two functions so that sequential context can guide the application of transformations on items stored in the working memory. This suggests that while monkeys may posses capabilities for sequence learning and working memory, they are separated from man in part by an evolutionary step providing the innate specification of these systems’ interaction required for this aspect of syntactic analysis.

 Conference site: http://www.infres.enst.fr/confs/evolang/