Large Language Model-Driven Hard Negatives Generation for Contrastive Aspect-Based Sentiment Analysis

Abstract

While sentiment analysis has advanced significantly, fine-grained sentiment classification such as aspect-based sentiment analysis (ABSA), continues to present challenges. These difficulties primarily stem from data scarcity and the inherent complexities of identifying sentiments specific to different aspects within a text. The recent emergence of generative artificial intelligence (AI), particularly large language models (LLMs), has fundamentally altered the landscape of data management and integration. To better capture nuanced sentiment differences, this paper proposes a novel approach that leverages LLMs to generate “hard negatives” for ABSA. We introduce the sentiment-aware chain-of-thought (SACoT) framework, which emulates a human-like reasoning process to manipulate the sentiment polarity of aspect terms. This is achieved by integrating chain-of-thought (CoT) reasoning with a few-shot prompting strategy, guiding the LLM to produce sentences with the desired sentiment towards specific aspects. Furthermore, we employ an LLM-as-a-Judge framework to refine these generated hard negatives, thereby enhancing their quality for subsequent use in supervised fine-tuning and contrastive learning. Extensive experiments on benchmark datasets demonstrate that our proposed approach, LLM-SCADA (LLM-guided sentiment-context-aspect data augmentation), significantly improves model performance, especially in transitioning from neutral to positive sentiments (NEU → POS). An in-depth ablation study further underscores the distinct contributions of the generated hard negatives, the contrastive learning approach, and the refinement process. These findings highlight the potential of LLMs to create sentiment-shifted samples based on aspect terms, paving the way for more robust and sophisticated techniques in the field of ABSA.