Leveraging dependency and constituent graphs for aspect sentiment triplet extraction

Abstract

Aspect Sentiment Triplet Extraction task (ASTE) aims to extract aspect terms, opinion terms, and determine their corresponding sentiment polarity from the text. Most current studies overlook the impact of dependency noise and sentence structure noise, while a few studies attempt to incorporate constituent features to mitigate such noise. However, they lack fine-grained fusion and alignment between dependency and constituent features. To address the above issue, this paper proposes a method that leverages dependency and constituent graphs (Dual-GNN). First, the model uses GCN to learn the dependency features and employs HGNN to capture the constituent features. Then, we enhance the dependency features with dependency related features and the constituent features with constituent related features. Additionally, we design a fine-grained word-level fusion and alignment matrices that combine dependency and constituent features to reduce the impact of noise and enable fine-grained triplet extraction. Finally, we adopt an efficient table-filling decoding strategy to extract the triplets. We conducted experimental validation on the ASTE-Data-v1, ASTE-Data-v2, and DMASTE datasets. The main results show that, compared with baseline methods, Dual-GNN achieves an F1 score improvement of 0.7 %-2.1 % on the ASTE-Data-v1 dataset and 0.6 %-1.5 % on the ASTE-Data-v2 dataset. Constituent features not only effectively reduce the impact of dependency noise and sentence structure noise but also help the model perceive multi-word term boundaries and accurately pair aspect terms with opinion terms. Combining the advantages of both dependency and constituent features enables more effective execution of the ASTE task.