Complementary Learning of Aspect Terms for Aspect-based Sentiment Analysis

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 7029–7039 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 7029 Complementary Learning of Aspect Terms for Aspect-based Sentiment Analysis Han Qin♠∗, Yuanhe Tian♥∗, Fei Xia♥, Yan Song♠† ♠The Chinese University of Hong Kong (Shenzhen) ♥University of Washington ♠hanqin@link.cuhk.edu.cn ♥{yhtian, fxia}@uw.edu ♠songyan@cuhk.edu.cn Abstract Aspect-based sentiment analysis (ABSA) aims to predict the sentiment polarity towards a given aspect term in a sentence on the fine-grained level, which usually requires a good understanding of contextual information, especially appropriately distinguishing of a given aspect and its contexts, to achieve good performance. However, most existing ABSA models pay limited attention to the modeling of the given aspect terms and thus result in inferior results when a sentence contains multiple aspect terms with contradictory sentiment polarities. In this paper, we propose to improve ABSA by complementary learning of aspect terms, which serves as a supportive auxiliary task to enhance ABSA by explicitly recovering the aspect terms from each input sentence so as to better understand aspects and their contexts. Particularly, a discriminator is also introduced to further improve the learning process by appropriately balancing the impact of aspect recovery to sentiment prediction. Experimental results on five widely used English benchmark datasets for ABSA demonstrate the effectiveness of our approach, where state-of-the-art performance is observed on all datasets.‡ Keywords: aspect-based sentiment analyses, complementary learning 1. Introduction Aspect-based sentiment analysis (ABSA) is an impor- tant task in natural language processing (NLP). It aims to detect the sentiment polarity of a given aspect term in an input sentence. Normally, a good understanding of the information concerning the aspect term, includ- ing the boundaries of the aspect term, its context, and the words contained in the aspect term, becomes es- sential for ABSA. For instance, the example sentence in Figure 1 has two different aspect terms (i.e., “price” and “service”), where the context word “reasonable” shapes “price” to be positive while “poor” determines “service” to be negative. In this case, with the same sentential context, the aspect terms have contradictory sentiment polarities, which raise the bar for ABSA and often cause inferior performance of models without careful treatment to the aspect terms and their contex- tual information. To incorporate such information into an ABSA model, most previous studies (Song et al., 2019; Zeng et al., 2019; Phan and Ogunbona, 2020; Yang and Zeng, 2020; Veyseh et al., 2020; Chen et al., 2020a) concate- nate aspect term(s) directly to the end of an input sen- tence with a special token serving as the separator (e.g., “[SEP]” for BERT-based models (Devlin et al., 2019)) and feed the resulted sentence+aspect pair into an en- coder. This simple and straightforward approach has been proved to be rather effective; however, it only em- phasizes the aspect term to the ABSA model without further understanding of the boundary and context of a given aspect term. Therefore, an appropriate approach *Equal contribution. †Corresponding author. ‡Our code is available at …

Content truncated to 500 words