Document-level Multi-aspect Sentiment Classification by Jointly Modeling Users, Aspects, and Overall Ratings

Proceedings of the 27th International Conference on Computational Linguistics, pages 925–936 Santa Fe, New Mexico, USA, August 20-26, 2018. 925 Document-level Multi-aspect Sentiment Classification by Jointly Modeling Users, Aspects, and Overall Ratings Junjie Li1,2, Haitong Yang3 and Chengqing Zong1,2,4 1 National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2 University of Chinese Academy of Sciences, Beijing, China 3 School of Computer, Central China Normal University, Wuhan 430079, China 4 CAS Center for Excellence in Brain Science and Intelligence Technology {junjie.li, cqzong}@nlpr.ia.ac.cn, htyang@mail.ccnu.edu.cn Abstract Document-level multi-aspect sentiment classification aims to predict user’s sentiment polarities for different aspects of a product in a review. Existing approaches mainly focus on text informa- tion. However, the authors (i.e. users) and overall ratings of reviews are ignored, both of which are proved to be significant on interpreting the sentiments of different aspects in this paper. There- fore, we propose a model called Hierarchical User Aspect Rating Network (HUARN) to consider user preference and overall ratings jointly. Specifically, HUARN adopts a hierarchical architec- ture to encode word, sentence, and document level information. Then, user attention and aspect attention are introduced into building sentence and document level representation. The document representation is combined with user and overall rating information to predict aspect ratings of a review. Diverse aspects are treated differently and a multi-task framework is adopted. Empirical results on two real-world datasets show that HUARN achieves state-of-the-art performances. 1 Introduction The ever-increasing popularity of online consumer review platforms, such as Tripadvisor1 and Yelp2, has led to large amounts of online reviews that are often too numerous for users to analyze. Consequently, there is a growing need for systems analyzing reviews automatically. Lots of approaches (Xia et al., 2011; Socher et al., 2013; Tang et al., 2015a; Yang et al., 2016) usually focus on determining the overall sentiment rating of a review. Actually, not only does a review express the general attitude of reviewer, but it also conveys fine-grained sentiments towards different aspects of corresponding products. Figure 1 shows an example where Bob posts a review about a hotel and gives scores on overall attitude, location, room, and service respectively. The analysis of these aspect ratings could not only benefit mining inter- ested aspects for users, but also help companies better understand the major pros and cons of the product. However, compared with the overall rating, users are less motivated to give aspect ratings. The reviews without aspect ratings are rampant, which are more than 46% in a simple corpus-based statistics3. Ac- cordingly, it is really useful to perform document-level multi-aspect sentiment classification, whose goal is to predict ratings for different aspects in a review (Yin et al., 2017). Multi-task learning (Caruana, 1997; Collobert et al., 2011; Luong et al., 2016) is a straightforward approach for document-level multi-aspect sentiment classification, which shares the input and hidden layers to obtain a document representation as the input of different aspect-specific classifiers. However, the representation fails to capture the differences between aspects. In fact, when we predict …

Content truncated to 500 words