BiSyn-GAT+: Bi-Syntax Aware Graph Attention Network for Aspect-based Sentiment Analysis

Findings of the Association for Computational Linguistics: ACL 2022, pages 1835 - 1848 May 22-27, 2022 c⃝2022 Association for Computational Linguistics BiSyn-GAT+: Bi-Syntax Aware Graph Attention Network for Aspect-based Sentiment Analysis Shuo Liang 1 , Wei Wei 2, �, Xian-Ling Mao 3, Fei Wang 4, Zhiyong He 5 1,2 Cognitive Computing and Intelligent Information Processing (CCIIP) Laboratory, School of Computer Science and Technology, Huazhong University of Science and Technology 3 School of Computer Science and Technology, Beijing Institute of Technology 4 Institute of Computig Technology, Chinese Academy of Sciences 5 Naval University of Engineering 1 shuoliang@hust.edu.cn, 2 weiw@hust.edu.cn, 3 maoxl@bit.edu.cn, 4 wangfei@ict.ac.cn, 5 moonmon_pub@outlook.com Abstract Aspect-based sentiment analysis (ABSA) is a fine-grained sentiment analysis task that aims to align aspects and corresponding sentiments for aspect-specific sentiment polarity infer- ence. It is challenging because a sentence may contain multiple aspects or complicated (e.g., conditional, coordinating, or adversative) rela- tions. Recently, exploiting dependency syntax information with graph neural networks has been the most popular trend. Despite its suc- cess, methods that heavily rely on the depen- dency tree pose challenges in accurately mod- eling the alignment of the aspects and their words indicative of sentiment, since the de- pendency tree may provide noisy signals of unrelated associations (e.g., the “conj” rela- tion between “great” and “dreadful” in Fig- ure 2). In this paper, to alleviate this prob- lem, we propose a Bi-Syntax aware Graph Attention Network (BiSyn-GAT+). Specif- ically, BiSyn-GAT+ fully exploits the syn- tax information (e.g., phrase segmentation and hierarchical structure) of the constituent tree of a sentence to model the sentiment-aware context of every single aspect (called intra- context) and the sentiment relations across as- pects (called inter-context) for learning. Ex- periments on four benchmark datasets demon- strate that BiSyn-GAT+ outperforms the state- of-the-art methods consistently. 1 Introduction Aspect-based sentiment analysis (ABSA) aims to identify the sentiment polarity towards a given as- pect in the sentence. Many previous works (Yang et al., 2018; Li et al., 2019) mainly focus on ex- tracting sequence features via Recurrent Neural Networks (RNNs) or Convolution Neural Networks (CNNs) with attention mechanisms, which often assume that words closer to the target aspect are � Corresponding Author (a) The food is great but the service and the environment are dreadful. (b) The food is great but the service and the environment are quite the opposite. positive negative neutral Figure 1: Examples of ABSA task. Each underlined aspect is classified to corresponding sentiment polarity. nsubj cop cc conj nsubj cop det det cc conj det but The food is great the service and the environment are dreadful <root> Figure 2: Dependency tree of “The food is great but the service and the environment are dreadful”. Two sepa- rate ellipses encircle its two clauses. The “conj” edge between “great” and “dreadful” is a noise. more likely to be related to its sentiment. However, the assumption might not be valid as exemplified in Figure 1 (a), “service” is obviously closer to “great” rather than “dreadful”, and these methods may …

Content truncated to 500 words