Prompting Large Language Models for Multiword Expression recognition

Proper handling of multiword expressions (MWEs) is a key aspect of developing linguistically sound natural language processing applications. Nevertheless, Large Language Models (LLMs) are not built specifically considering MWEs from the first layers. However, due to their presence in training data, higher layers of LLMs may develop features useful for MWE identification and analysis. This chapter explores the usefulness of LLMs when dealing with MWEs. Theoretical considerations are presented, starting with the tokenization mechanism. Then experimental results, specifically involving MWEs, are given. The main contribution of this work is in its empirical investigation of how LLMs treat MWEs (MWE identification) using zero-shot and Chain-of-Thought prompting. We provide a contrastive analysis of open-source LLMs and their performance in the task of MWE identification.