Fpa-GCN: enhancing aspect sentiment triplet extraction with feature-rich prediction-aware graph convolutional networks
摘要
Aspect Sentiment Triplet Extraction (ASTE) is a critical yet challenging task within Aspect-Based Sentiment Analysis (ABSA), which is dedicated to discerning aspect terms, their corresponding opinion terms, and sentiment polarities. The applicability of ASTE spans a broad array of sectors, from e-commerce to social media analytics and customer feedback analysis, highlighting its versatility and significance across various industries. However, the prevailing research methodologies frequently fail to capture the slight interplay among the three essential sentiment elements, often resulting in a misalignment that hampers the establishment of accurate contextual relationships. These misalignments often occur as discrepancies in the model’s understanding of contextually related content, deviating from expected associations. To address these challenges, this paper introduces the Feature-Rich Prediction-Aware Graph Convolutional Network (Fpa-GCN) model, meticulously designed for the ASTE task. By harnessing the power of BERT for superior semantic feature extraction and utilizing these features through Biaffine attention, the Fpa-GCN model employs a multi-branch graph convolutional network to effectively encapsulate contextual relations. Additionally, the model incorporates Information Fusion and an advanced Gating mechanism, further enhancing triplet extraction precision. Rigorous evaluations on benchmark datasets affirm the Fpa-GCN model’s supremacy over existing state-of-the-art models, achieving significant F1 score improvements. Specifically, the model achieves gains of 1.31%, 0.11%, 3.7%, and 2.15% on the Restaurant14, Laptop14, Restaurant15, and Restaurant16 datasets, respectively, in experiment group