Composition-based Heterogeneous Graph Multi-channel Attention Network for Multi-aspect Multi-sentiment Classification

Proceedings of the 29th International Conference on Computational Linguistics, pages 6827–6836 October 12–17, 2022. 6827 Composition-based Heterogeneous Graph Multi-channel Attention Network for Multi-aspect Multi-sentiment Classification Hao Niu1, Yun Xiong1∗, Jian Gao2, Zhongchen Miao2, Xiaosu Wang1, Hongrun Ren1, Yao Zhang1, Yangyong Zhu1 1Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University 2Shanghai Financial Futures Information Technology Co., Ltd. 1{hniu18, yunx, xswang19, renhr20, yaozhang, yyzhu}@fudan.edu.cn 2{gaojian, miaozc}@cffex.com.cn Abstract Aspect-based sentiment analysis (ABSA) has drawn more and more attention because of its extensive applications. However, towards the sentence carried with more than one aspect, most existing works generate an aspect-specific sentence representation for each aspect term to predict sentiment polarity, which neglects the sentiment relationship among aspect terms. Besides, most current ABSA methods focus on sentences containing only one aspect term or multiple aspect terms with the same sentiment polarity, which makes ABSA degenerate into sentence-level sentiment analysis. In this paper, to deal with this problem, we construct a hetero- geneous graph to model inter-aspect relation- ships and aspect-context relationships simul- taneously and propose a novel Composition- based Heterogeneous Graph Multi-channel At- tention Network (CHGMAN) to encode the constructed heterogeneous graph. Meanwhile, we conduct extensive experiments on three datasets: MAMS-ATSA, Rest14, and Lap- top14, experimental results show the effective- ness of our method. 1 Introduction Aspect-based sentiment analysis (ABSA) is a sen- timent analysis task, which aims to predict the sen- timent polarity (e.g., POSITIVE, NEGATIVE, NEU- TRAL) towards the given aspect term in a sentence. For example, there are two aspect terms decor and food in the sentence: The decor is not a special at all but their amazing food makes up for it, and these two aspect terms will be assigned with NEG- ATIVE and POSITIVE sentiment polarities respec- tively. Since the sentiments expressed by these two aspect terms are opposite in polarity, it is unreason- able to assign a sentence-level sentiment polarity. In this regard, ABSA can provide more detailed sentimental predictions compared with sentence- level sentiment analysis. ∗Corresponding author In recent years, neural network-based methods (Tang et al., 2016; Wang et al., 2016; Ma et al., 2017; Chen et al., 2017) are used to tackle ABSA task and achieve promising performance. Subse- quently, with the rise of contextual embedding, con- textualized language models, such as BERT (De- vlin et al., 2019), are introduced into ABSA and boost performance. Additionally, in order to utilize the dependency tree information, a lot of graph- based methods (Zhang et al., 2019; Sun et al., 2019; Huang and Carley, 2019; Wang et al., 2020) are pro- posed to tackle this task by encoding both contex- tual and dependency information simultaneously. However, most of these methods focus on the datasets, in which most sentences consist of only one aspect or multiple aspects with the same senti- ment polarity. Under these circumstances, aspect- based sentiment analysis degenerates into sentence- level sentiment analysis. Existing ABSA methods based on these datasets can hardly adapt to the situa- tion, where multiple aspect terms in a sentence …

Content truncated to 500 words