Span Labeling with Sentiment-Aware GCN and Multi-Channel Contrastive Learning for Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE), a critical fine-grained task in Aspect-Based Sentiment Analysis (ABSA), aims to extract triplets of aspect terms, opinion terms, and their sentiment polarity from user reviews. Despite significant advancements in this field in recent years, existing methods still face two major challenges: 1) Current syntax-dependency-based approaches typically treat dependency structures as static inputs, overlooking the dynamic interplay between syntactic relations and sentiment expressions; 2) Insufficient modeling capacity for span boundaries and overall expressions undermines the completeness and accuracy of triplet extraction in contexts involving multi-word phrases. To address these, we propose SGMCL, an enhanced span labeling method combining Sentiment-aware Graph Convolution Network (SGCN) and Multi-channel Contrastive Learning (MCL). Specifically, we dynamically adjust the weights of dependency edges using sentiment scores, enabling the features learned by GCN to focus more on sentiment-related structures, which effectively enhances the model’s ability to perceive multi-word expressions and long-distance dependencies. Meanwhile, we introduce a multi-channel contrastive learning mechanism to distinguish features between single words and multi-word span types for aspects and opinions, which effectively improves the model’s performance in scenarios involving multi-word expressions and complex boundaries. The span labeling scheme organically fuses syntactic structural and semantic information. Experiments on benchmark datasets show significant improvements in triplet extraction completeness and accuracy.