A Variational Approach to Weakly Supervised Document-Level Multi-Aspect Sentiment Classification

Proceedings of NAACL-HLT 2019, pages 386–396 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 386 A Variational Approach to Weakly Supervised Document-Level Multi-Aspect Sentiment Classification Ziqian Zeng1, Wenxuan Zhou2, Xin Liu1, and Yangqiu Song1 1Department of CSE, Hong Kong University of Science and Technology, HK 2Department of CS, University of Southern California, CA, USA 1{zzengae, xliucr, yqsong}@cse.ust.hk 2{zhouwenx}@usc.edu Abstract In this paper, we propose a variational ap- proach to weakly supervised document-level multi-aspect sentiment classification. Instead of using user-generated ratings or annotations provided by domain experts, we use target- opinion word pairs as “supervision.” These word pairs can be extracted by using depen- dency parsers and simple rules. Our objec- tive is to predict an opinion word given a tar- get word while our ultimate goal is to learn a sentiment polarity classifier to predict the sen- timent polarity of each aspect given a docu- ment. By introducing a latent variable, i.e., the sentiment polarity, to the objective func- tion, we can inject the sentiment polarity clas- sifier to the objective via the variational lower bound. We can learn a sentiment polarity clas- sifier by optimizing the lower bound. We show that our method can outperform weakly super- vised baselines on TripAdvisor and BeerAd- vocate datasets and can be comparable to the state-of-the-art supervised method with hun- dreds of labels per aspect. 1 Introduction Document-level multi-aspect sentiment classifica- tion (DMSC) aims to predict the sentiment polar- ity of each aspect given a document which con- sists of several sentences describing one or more aspects (Wang et al., 2010, 2011; Yin et al., 2017). Solving the DMSC task is useful for providing both recommendations for users and suggestions for business owners on customer review platforms. Aspect based sentiment classification (Tang et al., 2016a,b; Wang et al., 2016b; Chen et al., 2017; Ma et al., 2017; Wang et al., 2018) was usu- ally done by supervised learning, where aspect- level annotations should be provided. Aspect-level annotations are not easy to obtain. Even when the platform provides the function to rate for different aspects, users are less likely to submit all of them. For example, about 37% of the aspect ratings are missing on TripAdvisor. If we can solve DMSC task without using aspect-level annotations, it can save human effort to annotate data or collect user- generated annotations on the platform. Existing weakly supervised approaches (Wang et al., 2010, 2011) use overall polarities instead of aspect polarities as “supervision.” Compared with the polarity of each aspect, it is relatively easy to obtain overall polarities. Specifically, they min- imize the square loss between the overall polar- ity and the weighted sum of all aspect polarities. However, when users only care about a particu- lar rare aspect, e.g., childcare services, these ap- proaches cannot estimate parameters of the rare aspect incrementally. They have to re-collect doc- uments which mentioned this rare aspect and es- timate parameters of all aspects based on the new corpus. In addition, these approaches assume the …

Content truncated to 500 words