Semantic Coercion in Language: Beyond Distributional Analysis

We argue that lexical representations which are built from evidence of distributional behavior alone are unable to fully explain the rich variation in linguistic meaning in language. Lexical meaning is modulated in context and contextual semantic operations influence the behavior that words exhibit: this is why a context-sensitive lexical architecture is needed in addition to empirical analysis to make sense of corpus data. As a case study that shows how distributional analysis and theoretical modeling can interact, we present a corpus investigation aimed at identifying mechanisms of semantic coercion in predicate-argument constructions, conducted within the Generative Lexicon (GL) model. The analysis demonstrates the ubiquity of the phenomenon and highlights the limits of a theory-blind distributional analysis. In particular, it shows how coercion may alter the distributional behavior of words, allowing them to show up in contexts in which they would otherwise not appear. A descriptive theory of coercion as proposed here is relevant not only for theoretical considerations, but also for computational purposes such as the elaboration of annotation schemes for the automatic recognition and resolution of coercion phenomena in texts.