EAD: An Exploration-Analysis-Disambiguation Reasoning Framework for Word Sense Disambiguation with Small LLMs

Forum topic · 小凯 · 2026-03-07

Summary

Researchers propose the Exploration-Analysis-Disambiguation (EAD) framework, a three-stage explicit reasoning method that enables low-parameter large language models to perform Word Sense Disambiguation (WSD). WSD remains a key NLP challenge for rare or ambiguous words where context alone is insufficient, and LLMs often rely on frequent-sense bias rather than genuine contextual analysis. EAD addresses this by guiding the model through three stages: exploring potential senses, analyzing contextual relevance, and final disambiguation. Experiments show the framework significantly improves WSD performance on standard benchmarks using models with fewer than 10 billion parameters. The paper (arXiv:2603.05400) is authored by Junyu Lu, Yanan Zheng, Chen Gong, and colleagues, and was posted to zhichai.net on 2026-03-07.

Paper Overview

Field: NLP Authors: Junyu Lu, Yanan Zheng, Chen Gong, Yibo Zhang, Shengnan Li, Yining Zhang, Daimeng Wei, Zhiqiang Zhang, Jiajun Zhang Published: 2026-03-05 arXiv: 2603.05400

Abstract

Word Sense Disambiguation (WSD) remains a key challenge in Natural Language Processing (NLP), especially when dealing with rare or ambiguous words where context alone is insufficient. While large language models (LLMs) have shown promise, their performance on WSD tasks is often hindered by their tendency to rely on frequent sense biases rather than contextual analysis.

The EAD Framework

This paper proposes an Exploration-Analysis-Disambiguation (EAD) reasoning framework that enables low-parameter LLMs to perform explicit reasoning for WSD. The framework consists of three stages:

1. Exploration — exploring potential senses of the target word 2. Analysis — analyzing contextual relevance of each candidate sense 3. Disambiguation — making the final sense decision

Results

Experiments show that this approach significantly improves WSD performance on standard benchmarks while using models with fewer than 10 billion parameters, demonstrating that explicit structured reasoning can substitute for raw model scale on this task.

---

*Auto-collected on 2026-03-07*