Exemplar-based generalization at the interface between syntax and semantics.

摘要

In exemplar models of language processing, the categories underlying linguistic generalization are temporary, context-dependent, and local summarizations of more concrete knowledge created through online abstraction. This study applies the exemplar-based approach to two linguistic problems at the interface of syntax and semantics that involve combinations of multiple words, namely thematic fit and prenominal adjective order.;The exemplar model of thematic fit presented in this study estimates the plausibility of semantic roles on the basis of online abstractions over verb--argument pairs. The model is both accurate and robust to data sparsity. Whereas the correlation with human thematic fit ratings and the number of verb--argument--role triples for which estimates can be generated are at each other's expense in other state-of-the-art models, the exemplar model's estimates score high on both their coverage of and correlation with a critical set of human ratings.;Kemmerer, Tranel, and Zdanczyk (2009) reported patients who failed to discriminate between preferred and dispreferred orders of prenominal adjectives, yet were sensitive to the order of adjectives in relation to other parts of speech, and able to judge which of two adjectives was most similar to a cue adjective. This study shows that the selective impairment of those patients can be explained as overeager abstraction. Oversmoothing a similarity-smoothed bigram language model resulted in the same performance dissociation between tasks as reported for Kemmerer et al.'s selectively impaired patients. Additionally, the strength with which the patients preferred a specific adjective order for a given stimulus was predicted by the stimulus' robustness to overeager abstraction. A two-alternative forced choice and a naturalness rating task showed that exemplar-based variables also make a larger contribution to the explained variation in Dutch adjective order preferences than the linear precedence between adjective classes for normal native speakers.;The main finding is that online abstraction plays a crucial role in explaining the generalization of linguistic patterns at the syntax--semantics interface that involve word combinations. The implication is that models aiming to explain the generalization of linguistic patterns involving multiple words should incorporate a component that captures effects of online abstraction over concrete exemplars.