MWE as WSD: Solving Multiword Expression Identification with Word Sense Disambiguation

Recent work in word sense disambiguation (WSD) utilizes encodings of the sense gloss (definition text), in addition to the input words and context, to improve performance. In this work we demonstrate that this approach can be adapted for use in multiword expression (MWE) identification by training a Bi-encoder model which uses gloss and context information to filter MWE candidates produced from a simple rule-based extraction pipeline. We achieve state-of-the-art results in MWE identification on the DiMSUM dataset, and competitive results on the PARSEME 1.1 English dataset using this method. Our model also retains most of its ability to perform WSD, demonstrating that a single model can successfully be applied to both of these tasks. Additionally, we experiment with applying Poly-encoder models to MWE identification and WSD, introducing a modified Poly-encoder architecture which outperforms the standard Poly-encoder on these tasks.