Syntax-Aware Graph Attention Network for Aspect-Level Sentiment Classification

Proceedings of the 28th International Conference on Computational Linguistics, pages 799–810 Barcelona, Spain (Online), December 8-13, 2020 799 Syntax-Aware Graph Attention Network for Aspect-Level Sentiment Classification Lianzhe Huang, Xin Sun, Sujian Li, Linhao Zhang and Houfeng Wang MOE Key Lab of Computational Linguistics Peking University, Beijing, 100871, China {hlz,sunx5,lisujian,zhanglinhao,wanghf}@pku.edu.cn Abstract Aspect-level sentiment classification aims to distinguish the sentiment polarities over aspect terms in a sentence. Existing approaches mostly focus on modeling the relationship between the given aspect words and their contexts with attention, and ignore the use of more elaborate knowledge implicit in the context. In this paper, we exploit syntactic awareness to the model by the graph attention network on the dependency tree structure and external pre-training knowledge by BERT language model, which helps to model the interaction between the context and aspect words better. And the subwords of BERT are integrated into the dependency tree graphs, which can obtain more accurate representations of words by graph attention. Experiments demonstrate the effectiveness of our model. 1 Introduction Aspect-level sentiment analysis (ABSA) (Liu, 2012), as a fundamental natural language processing task, has attracted lots of attentions recently. Its typical application is to classify the target aspects of a prod- uct in a customer review into the sentiment polarities of positive, neutral, or negative according to the contexts (Hu and Liu, 2004) . Taking the sentence “The performance of this laptop is excellent, but the screen is terrible” as example, the correct ABSA results would be to judge the aspect word performance as positive, and screen as negative. The early approaches of aspect-level sentiment analysis mainly use handcrafted features to train a statistical classifier (Kiritchenko et al., 2014). With the rapid progress of deep neural network (DNN), lots of work based on DNN have been proposed (Dong et al., 2014; Tang et al., 2015b; Ma et al., 2017). These studies mostly depend on recurrent neural networks (RNNs) (Bahdanau et al., 2014) to model the semantic relationship between the aspect words and their contexts. However, their performances are affected by the RNN’s limited ability of capturing long-distance dependencies (Werbos, 1990) and loss of syntax information (Sun et al., 2019). To overcome the shortages of RNN-based models mentioned above, there have been a lot of work using various methods like BERT language model and graph neural networks recently. Since BERT language model (Devlin et al., 2018) based on Transformer can alleviate long-distance dependence of RNN, some studies have designed BERT-based models for the ABSA task and achieved outstanding results (Song et al., 2019; Rietzler et al., 2019; Zhao et al., 2019). To the best of our knowledge, most of these work just utilize BERT simply as an embedding layer, which sequentially models the position embedddings of the context and cannot utilize the syntactic information. Another improvement is the use of GNN in ABSA (Sun et al., 2019), since GNN has the capability to model linguistic structures like dependency trees other than only sequential data. With such a capability, the model can build the …

Content truncated to 500 words