Toward Knowledge Integration With Large Language Model for End-to-End Aspect-Based Sentiment Analysis in Social Multimedia

Abstract

Aspect-based sentiment analysis (ABSA) aims to identify specific sentiment elements in social multimedia content. To address aspect extraction and sentiment prediction together, recent studies have utilized a sequence tagging approach, mainly leveraging pretrained language models (PLMs) with specific architecture and auxiliary subtasks. However, these approaches often overlook task-related knowledge and struggle to scale across different domains. With advances in large language models (LLMs), there is a rising trend in constructing generative ABSA models. Nevertheless, these techniques tend to emphasize specific frameworks and overlook comprehensive knowledge representation. To address these challenges while leveraging the advantages of LLM and PLM-based methods, we propose a hybrid knowledge integration framework (HFABGKI). It employs a parameter-efficient fine-tuning technique, allowing for plug-and-play integration with existing LLMs. To bridge the LLM and PLM-based models, HF-ABGKI incorporates a global label semantic representation for potential aspect tokens, in which a simplified gating mechanism is proposed to filter useful information. Experimental results from six public social multimedia datasets demonstrate that our approach can accurately extract aspect terms and predict their sentiment polarity, achieving state-of-the-art performance compared to existing ABSA methods.