Aspect-level sentiment triplet extraction (ASTE) is a core task in fine-grained sentiment analysis, aiming to jointly extract aspect terms and opinion terms while classifying the sentiment polarity of each aspect term. This task faces several critical challenges, including difficulty in effectively capturing high-order semantic and syntactic dependencies among words; the lack of explicit guidance from sentiment knowledge and insufficient feature interaction in the enumeration-based pairs of aspect and opinion spans, which produces a large number of negative samples and thus degrades quality of pair representations. To tackle these issues, we present an ASTE method based on sentiment-knowledge-enhanced and dual hypergraphs. Specifically, sentiment word embeddings derived from a sentiment lexicon are first fused with BERT contextual representations via a gated mechanism. Then, a multi-head cross-attention based gated fusion module is designed to facilitate feature interaction between the sentiment-enhanced BERT representations and hypergraph vectors constructed from semantic similarity, yielding a multi-source fused representation. Next, we adopt a span enumeration strategy to represent candidate spans for aspect and opinion by incorporating BERT contextual semantics, part-of-speech tags, and other linguistic features, followed by entity classification. Afterwards, we enhance pair modeling through multi-head cross-attention and a syntactic dependency hypergraph convolutional network to capture high-order dependencies. Finally, sentiment polarity is predicted by jointly leveraging the aspect span representation, the opinion span representation, and their element-wise multiplicative interaction. Our model innovatively integrates multi-source feature fusion and semantic–syntactic dependency hypergraph convolution mechanisms. Experiments on four benchmark datasets show that our method achieves MacF1 of 76. 25%, 65. 98%, 78. 81%, and 75. 59%, respectively, surpassing strong baselines and confirming the validity of the proposed framework.
Zhao et al. (Fri,) studied this question.