| The present thesis is part of a larger effort to locate the production and perception of language within the broader context of brain mechanisms for action and perception more generally. As the first step, we use the task of describing visual scenes to explore the suitability of the currently proposed framework of a schema-based linguistics. We developed a new kind of semantic representation, SemRep, which is an abstract form of visual information with an emphasis on the spatial linkage of entities, attributes and actions. SemRep provides a compact graph-like structure with enough formal semantics for verbal description of a scene, reducing the relatively complex task of semantic processing to a graph matching task. The present thesis reports results on implementing the production of sentences using Template Construction Grammar (TCG), a new form of Construction Grammar distinguished by its use of SemRep to express semantics. Constructions, represented as schema instances in our approach, compete and cooperate to cover the SemRep to produce a description of the visual scene at hand. In our approach, the vision system interprets a part of the scene under attention by creating or updating the corresponding SemRep while the language system applies constructions on that part of SemRep by the principles of TCG. The current work proposes specific mechanisms on how a representation (i.e. SemRep) is built from the perceived visual scene and what influences the choice of constructions for the produced utterances. More specifically, the complexity of a perceived event and the constraints on available computational resources are hypothesized to be the main driving force of the resultant sentential structure being produced. The former affects the coverage of the perceived subscene, which represents a particular view on the scene at a certain moment, and the resultant formulation of SemRep. The latter, which is parameterized as the threshold of utterance, limits the amount of time and constructions used for formulating descriptions, resulting in different degrees of "well-formedness" of produced sentences. To test hypotheses, we conducted a series of eye-tracking experiments with experimental settings to induce various levels of event complexity (e.g. showing scenes of different event structures) and threshold (e.g. imposing time pressure). Based on the examination on the time-locked eye movements and recorded speech, the present thesis presents supporting evidence for the proposed mechanisms. We conclude by demonstrating how the combinations of various levels of threshold and event complexity in the framework of SemRep and TCG can address both of the apparently opposing strategies in sentence production: the "structural view" which asserts the preparation of sentential structure and the preparation of each constituent are interleaved, and the "incremental view" that claims that those are separated in an orderly fashion. |
