DRHGNN: a dynamic residual hypergraph neural network for aspect sentiment triplet extraction

Abstract

Aspect Sentiment Triple Extraction (ASTE) is an emerging sentiment analysis task. Many existing methods focus on designing a new labeling scheme to enable end-to-end operation of the model. However, these methods overlook the relationships between words in the ASTE task. In this paper, we propose the Dynamic Residual Hypergraph Neural Network (DRHGNN), which fully considers the relationships between words. Specifically, based on the pre-defined ten types of word pair relationships, we employ a graph attention network to model sentence features as a relational graph matrix. Subsequently, we use a dynamic hypergraph network to learn deep features from the transformed graph structure, then constructing relation-aware node representations. Furthermore, we integrate a residual connection to improve the performance of our DRHGNN model. Finally, we design a relationship constraint to dynamically control the number of hyperedges, thereby enhancing the effectiveness of the dynamic hypergraph neural network. Extensive experimental results on benchmark datasets show that our proposed model significantly outperforms state-of-the-art methods, demonstrating the effectiveness and robustness of the model.