Fine-grained Question-Answer Sentiment Classification with Hierarchical Graph Attention Network
Published:
User-oriented Question-Answer (QA) text pair plays an increasingly important role in online e-commerce platforms, and expresses sentiment information with complicated semantic relations, causing great challenges for accurate sentiment analysis. To address this problem, we propose a novel hierarchical graph attention network (HGAT) to explore abundant relations. Firstly, we utilize the dependency parser to model relations of sentiment words with consideration of syntactic structures within sub-sentences. Then, to better extract hidden features of these sentiment words, we feed the dependency graph into an improved word-level graph attention network (GAT) that incorporates the learned attention weight with the prior graph edge weight. Besides, the sigmoid self-attention mechanism is applied to aggregate salient word representations. Finally, we establish a graph of all sub-sentences with a strong connection and capture inter-relations and intra-relations through the sentence-level GAT. Extensive experiments show that HGAT can achieve significant improvements in QA-style sentiment classification compared with several baselines.
Recommended citation: Zeng J, Liu T, Jia W, et al. Fine-grained Question-Answer sentiment classification with hierarchical graph attention network[J]. Neurocomputing, 2021, 457: 214-224.