Computational models for metaphor understanding

• Main Author: Chua, Zi Heng
• Other Authors: Erik Cambria
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/165935
• Institution: Nanyang Technological University

A metaphor is a figure of speech used to make the language more vivid and expressive. Metaphor understanding is a complex task for NLP, as it involves recognizing analogies and making inferences between non-literal concepts. Previous literature on Metaphor Understanding focuses mainly on the Metaphor Identification subtask instead of the more challenging Metaphor Interpretation subtask, due to the lack of annotated datasets on paraphrases. In addition, previous works employ complex methods to deal with Multi-Word Expression (MWE) metaphors, which are processed separately from single-word metaphors. This project involves the full Metaphor understanding pipeline. Firstly, significant contributions were made to annotating the novel Metaphor Interpretation dataset. Next, a preliminary bad case analysis was conducted for Metaphor Identification. Finally, this paper proposes 2 Metaphor Interpretation models based on different training paradigms: Classification and Masked Language Modelling (MLM). Our unified processing methods apply to both single-word and MWE metaphors, which simplifies the task. A detailed evaluation is conducted to compare the performances of both proposed models on how generalisable they are in terms of unseen cases and MWEs.