GRAPH ATTENTION NETWORKS WITH STRUCTURAL ATTENTION MECHANISM FOR ASPECT-BASED SENTIMENT CLASSIFICATION

Aspect-based sentiment classification has caused increasingly attention in domestic and foreign research because of the wide range of applications. More and more neural network models was present for the task. Although these methods have achieved promising performances, there are still some limitations. These approaches may not consider the syntactic structure of sentences and long-term word dependencies, and some unrelated context words may be misjudged as powerful words for predicting sentiment polarity, which greatly affects the effectiveness of the model. To address this problem, we propose a novel structural attention mechanism and graph attention network (SSA-GAT), which captures the contextual fragments related to the sentiment of the aspect terms, and employs dependency tree to obtain the syntactic information of the sentence and word dependencies. Our approach can effectively combine semantic relations, syntactic dependencies and long-distance dependence on words. Experimental results on three benchmarking have shown that the effectiveness of the proposed method compared with several state-of-the-art models.