Dependency-guided aspect extraction with coreference resolution and binary aspect classification

Aspect-Based Sentiment Analysis (ABSA) aims to identify opinion targets (aspects) in text and determine the sentiment expressed toward each of them. Classical ABSA models based on recurrent networks and transformers achieve strong performance on benchmarks, but they often under-extract explicit aspects in syntactically complex sentences and do not generalize well across domains. Pure prompt-based use of Large Language Models (LLMs) offers strong semantic reasoning but tends to miss aspects, is sensitive to prompt design, and incurs non-trivial inference cost. This paper proposes CoreDep, a dependency-driven hybrid framework that integrates coreference resolution, dependency parsing, and a fine-tuned transformer classifier for robust aspect extraction. CoreDep first generates a high-recall candidate set using linguistic structure, then applies a binary classifier to recover precision. We evaluate aspect extraction under fuzzy span matching to account for boundary ambiguity, and demonstrate that the extracted aspects support effective downstream sentiment analysis using both LLMs and supervised classifiers. Experiments on the SemEval-2014 Laptop dataset show that CoreDep achieves an aspect-level precision of 0.8448, recall of 0.7508, and F1-score of 0.7950. We further construct and annotate a new laptop review dataset from Takealot.com containing 3,175 aspect instances, on which our domain-adapted model achieves an F1-score of 0.7766. Overall, the results indicate that combining linguistic structure with lightweight neural validation remains an effective and practical strategy for ABSA in the LLM era.