Textual Entailment as a Framework for Differentiating Metonymy from Metaphor
Speaker: Kevin Mathews (HITS)
The figures of speech metonymy and metaphor entail a transfer of meaning between two entities. In the case of metonymy, the entities tend to be dissimilar but contiguous, whereas in the case of metaphor, the entities tend to be similar to each other. Our goal is to predict whether a word in a piece of text is metonymic or metaphoric. We explore whether the large pre-trained language models are able to establish metonymic/metaphoric links between the literal meaning of a word and its contextual meaning. For this purpose, we recast this problem as Textual Entailment Recognition, which checks whether a hypothesis entails, contradicts or is not related to a premise. We use the literal (most common) meaning of the word to be disambiguated as the premise, and the input text as the hypothesis. In this talk, I will present our model and discuss various baselines that examine the effectiveness of our approach.