Interpreting Logical Metonymy through Dense Paraphrasing

Compositionality has been argued to be both a desirable and perhaps even necessary component of interpreting language, yet there appear to be many linguistic phenomena that do not overtly exhibit semantic compositional behavior. One of the most interesting challenges involves the phenomena of contextual modulations referred to collectively as semantic coercion or logical metonymy. Some models of how we understand, for example, Alex enjoyed her coffee and Jen heard the train incorporate mechanisms that provide “compositional flexibility”, to allow event-selecting and sound-selecting verbs, respectively, to combine with arguments that denote neither. In this paper, we present a computational model that provides for such flexibility in the interpretation of a verb with its arguments, for such coercive contexts in English. Specifically, we argue that such constructions typically have surface structural correlates in the form of dense paraphrases, and that these forms can be used to model the masked content in the coerced compositional context. We present preliminary results using a transformer architecture (BERT) on a masked completion task. This suggests that constructions involving “enriched composition” can in fact be computationally analyzed with attention-based architectures. Our results show that modeling logical metonymy is a challenging task but can be substantially improved by fine-tuning through dense paraphrasing.