Tackling Frame-Semantic Role Labeling through Analogies

Anonymous

Abstract: Analogy making has been claimed to be at the core of cognition and be intimately related with abstraction and categorization. Despite its importance, analogies have only been scantily studied in Natural Language Processing (NLP) with most work being limited to word analogies. Most extant approaches view analogy making as the identification of the fact that pairs $(A,B)$ and $(C,D)$ share the same latent relation without necessarily naming that relation. In this paper we adapt this framework in the context of frame semantics, focusing on the problem of semantic role labeling. For a given target sentence and a predicate we are able to identify all semantic roles, casting the problem as an analogy with a previously annotated sentence of the same frame, but not necessarily of the same predicate. We show that careful selection of the source sentence has the potential to surpass state of the art results while at the same time using a computationally lean model.

Paper Type: long

Research Area: Semantics: Sentence-level Semantics, Textual Inference and Other areas