Testing the Effectiveness of Knowledge Editing in Financial Polarity Editing

Abstract

A major challenge for large language models (LLMs) is adapting to domain-specific knowledge and context efficiently. In this study, we propose a domain adaptation method using knowledge editing for sentiment analysis, where the criteria for positive and negative classifications change over time and across contexts (e.g., before and after the COVID-19 pandemic). While previous studies have demonstrated the effectiveness of knowledge editing, its use in sentiment classification remains underexplored. We explore a novel training approach leveraging knowledge editing and evaluate its effectiveness in a Japanese sentiment analysis task. Specifically, we propose an enhanced training method applying Rank-One Model Editing (ROME) to the LLaMA model and assess its performance in a zero-shot setting. The results show our approach achieves higher accuracy in sentiment classification than conventional methods. This study provides the first empirical validation of knowledge editing in sentiment analysis and highlights its potential as an efficient domain adaptation technique.