Harnessing Domain Insights: A Prompt Knowledge Tuning Method for Aspect-Based Sentiment Analysis

Abstract

Aspect-Based Sentiment Analysis (ABSA) endeavors to predict the sentiment polarity of specific aspects in the given reviews. Recently, prompt tuning has been widely explored and achieved remarkable success for improving semantic comprehension in many NLP tasks. However, most existing methods consider semantic tuning for various tasks while neglecting the domain knowledge, such as commonsense background knowledge, leading to poor semantic quality and inferior model performance. To bridge the gap, in this paper, we conduct a systematic study of  Prompt Tuning with Domain Knowledge (PTDK) for ABSA, which aims to design efficient prompts that guide the model to learn the knowledge of specific aspects in ABSA. Specifically, we first fine-tune Large Language Models (LLMs) using hard prompts, which enhances the ability to extract enriched domain insights from the knowledge base. Additionally, we employ a Co-occurrence Gate to meticulously filter and refine this domain knowledge, which significantly strengthens the domain representation capacity of prompt template. Concurrently, we construct a hybrid prompt template, designed for the collaborative tuning of soft prompts along with the integration of specific masks across various domain vectors. Experimental results on three public datasets demonstrate that our method consistently outperforms current state-of-the-art baselines in all cases, with an average increase in accuracy of 1.15% and an enhancement of the F1-score by 1.03%.

Keywords: LLMs, prompt tuning, domain knowledge base, co-occurrence gate, hybrid prompt