Document-Level Multi-Aspect Sentiment Classification as Machine Comprehension

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2044–2054 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Document-Level Multi-Aspect Sentiment Classification as Machine Comprehension Yichun Yin1, Yangqiu Song2, Ming Zhang1 1School of Electronics Engineering and Computer Science, Peking University, Beijing, China 2Department of Computer Science and Engineering, HKUST, Hong Kong {yichunyin, mzhang cs}@pku.edu.cn, yqsong@cse.ust.hk Abstract Document-level multi-aspect sentiment classification is an important task for cus- tomer relation management. In this paper, we model the task as a machine compre- hension problem where pseudo question- answer pairs are constructed by a small number of aspect-related keywords and as- pect ratings. A hierarchical iterative atten- tion model is introduced to build aspect- specific representations by frequent and repeated interactions between documents and aspect questions. We adopt a hi- erarchical architecture to represent both word level and sentence level informa- tion, and use the attention operations for aspect questions and documents alterna- tively with the multiple hop mechanism. Experimental results on the TripAdvisor and BeerAdvocate datasets show that our model outperforms classical baselines. 1 Introduction Document-level sentiment classification is one of the pragmatical sentiment analysis tasks (Pang and Lee, 2007; Liu, 2010). There are many Web sites having platforms for users to input reviews over products or services, such as TripAdvisor, Yelp, Amazon, etc. Most of reviews are very compre- hensive and thus long documents. Analyzing these documents to predict ratings of products or ser- vices is an important complementary way for bet- ter customer relationship management. Recently, neural network based approaches have been de- veloped and become state-of-the-arts for long- document sentiment classification (Tang et al., 2015a,b; Yang et al., 2016). However, predict- ing an overall score for each long document is not enough, because the document can mention dif- “The situation is good, it's very clean, but there is nothing special. Breakfast at downstairs is directly from grocery store. Water pressure is good! A decent choice for sleeping. New York is expensive place!” Cleanliness:5 Room:: 4 Value:: 2 Review Rating Figure 1: Example: hotel review with aspects. ferent aspects of the corresponding product or ser- vice. For example, in Figure 1, there could be dif- ferent aspects for a review of hotel. These aspects help customer service better understand what are the major pros and cons of the product or ser- vice. Compared to the overall rating, users are less motivated to give aspect ratings. Therefore, it is more practically useful to perform document-level multi-aspect sentiment classification task, predict- ing different ratings for each aspect rather than an overall rating. One straightforward approach for document- level multi-aspect sentiment classification multi-task learning (Caruana, 1997). For neuralis networks, we can simply treat each aspect (e.g., rating from one to five) as a classification task, and let different tasks use softmax classifier to extract task-specific representations at the top layer while share the input and hidden layers to mutually en- hance the prediction results (Collobert et al., 2011; Luong et al., 2016). However, such approach ig- nores …

Content truncated to 500 words