Syntactic-Guided Chain of Thought for Iterative Implicit and Explicit Target Detection in Aspect-Based Sentiment Analysis

Prompt engineering is essential for optimizing the performance of large-language models (LLMs), particularly in tasks requiring complex interpretations such as aspect-based sentiment analysis (ABSA). However, existing methodologies often struggle to detect implicit targets, especially in multi-opinion sentences where sentiments are directed toward aspects that are not explicitly mentioned. This study addresses this gap by proposing the Iterative Syntactic-Guided Chain of Thought (IS-COT) framework, which integrates dependency parsing with modular prompt engineering to enhance LLMs’ reasoning capabilities. IS-COT leverages syntactic structures and iterative refinement to detect both explicit and implicit targets while resolving ambiguities in multi-opinion contexts. Experimental evaluations on benchmark datasets, Sem-Eval 2015 (Res15) and Sem-Eval 2016 (Res16), demonstrated the effectiveness of the framework, achieving superior performance with 80.43 F1 scores on Res15 and 84.47 F1 scores on Res16, significantly outperforming state-of-the-art models. These results highlight IS-COT’s potential of IS-COT as a comprehensive and interpretable solution for ABSA, addressing the critical limitations of existing approaches and advancing the field through innovative syntactic and semantic integration.