Beyond Benchmarks: Interpretable Aspect-Based Sentiment Analysis for Actionable Insights from Customer Reviews

The pursuit of high performance in Aspect-Based Sentiment Analysis has led to a reliance on complex, large-scale generative models. This trend creates a significant disconnect between academic benchmarks and the needs of enterprise applications, which require transparent, auditable, and actionable insights. This paper introduces Syntactic Pattern Aspect-Based Sentiment Analysis (SP-ABSA), a lexicon-driven framework designed to bridge this interpretability gap. Our primary contribution is a novel, unsupervised two-stage methodology. The offline stage automatically constructs a domain-specific aspect lexicon from a raw corpus using noun chunk extraction and hybrid semantic-morphological clustering. This lexicon then guides the online stage, which employs explicit syntactic pattern checkers on dependency-parsed text to extract sentiment triplets in the format [Aspect, Attribute, Sentiment]. When evaluated on the SemEval-2014 benchmark, SP-ABSA achieves an F1-score of 0.283, a significant improvement over a traditional dependency-rule baseline (0.184). More importantly, our qualitative analysis reveals a systemic divergence between the framework's grammatically-grounded outputs and the salience-focused nature of benchmark annotations. By identifying instances of 'correct but unannotated' insights, we argue that standard metrics are insufficient for evaluating discovery-oriented systems and that interpretability offers a more direct path to business value.