Conference Paper/Proceeding/Abstract 29 views
An Exploration-Analysis-Disambiguation Reasoning Framework for Word Sense Disambiguation with Low-Parameter LLMs
Deshan Sumanathilaka 
,
Nicholas Micallef ,
Julian Hough
,
Nicholas Micallef ,
Julian Hough
The Fifteenth biennial Language Resources and Evaluation Conference (LREC)
Swansea University Authors:
Deshan Sumanathilaka , Nicholas Micallef , Julian Hough
-
PDF | Accepted Manuscript
Download (266.77KB)
Abstract
Word Sense Disambiguation (WSD) remains a key challenge in Natural Language Processing (NLP), especially when dealing with rare or domain-specific senses that are often misinterpreted. While modern high-parameter Large Language Models (LLMs) such as GPT-4-Turbo have shown state-of-the-art WSD perfor...
| Published in: | The Fifteenth biennial Language Resources and Evaluation Conference (LREC) |
|---|---|
| Published: |
|
| URI: | https://cronfa.swan.ac.uk/Record/cronfa71528 |
| Abstract: |
Word Sense Disambiguation (WSD) remains a key challenge in Natural Language Processing (NLP), especially when dealing with rare or domain-specific senses that are often misinterpreted. While modern high-parameter Large Language Models (LLMs) such as GPT-4-Turbo have shown state-of-the-art WSD performance, their computational and energy demands limit scalability. This study investigates whether low-parameter LLMs (4B parameters) can achieve comparable results through fine-tuning strategies that emphasize reasoning-driven sense identification. Using the FEWS dataset augmented with semi-automated, rationale-rich annotations, we fine-tune eight small-scale open-source LLMs (e.g. Gemma and Qwen). Our results reveal that CoT-based reasoning combined with neighbour-word analysis achieves performance comparable to GPT-4-Turbo in zero-shot settings. Importantly, Gemma-3-4B and Qwen-3-4B models consistently outperform all medium-parameter baselines and state-of-the-art models on FEWS, with robust generalization to unseen senses. Furthermore, evaluation on the unseen ''Fool Me If You Can'' dataset confirms strong cross-domain adaptability without task-specific fine-tuning. This work demonstrates that with carefully crafted reasoning-centric fine-tuning, low-parameter LLMs can deliver accurate WSD while substantially reducing computational and energy demands. |
|---|---|
| Keywords: |
Word Sense Disambiguation, Low-parameter LLMs, Reasoning-driven Fine-tuning |
| College: |
Faculty of Science and Engineering |