Title: Impact of Context on End-to-end Metonymy Detection, and Metonymy Detection for Entity Linking
Speaker: Kevin Mathews (HITS)
Metonymy is a figure of speech in which an entity is referred to by another related entity. In this talk, I present two ongoing works on metonymy resolution: (1) The interpretation of a target word relies more on the context than the word itself. To test this claim for end-to-end metonymy detection, we compare two variants: one variant relies primarily on the target word, and the other variant relies primarily on the context. Our results show that masking the target word improves performance, especially in the fine-grained experimental setting. (2) Entity linking (EL) is the task of linking named entity mentions in text to entities in a knowledge base. Existing EL systems rely on prior probability score, context words, and coherence with adjacent entities. We compare the performance of such a system on metonymic mentions with that on non-metonymic (literal) mentions. We show that the performance is inferior on metonymic mentions, even though the overall performance is high.