Multi-aspects Rating Prediction Using Aspect Words and Sentences

Multi-aspects Rating Prediction Using Aspect Words and Sentences Takuto Nakamuta Kazutaka Shimada Department of Artificial Intelligence Kyushu Institute of Technology 680-4 Kawazu Iizuka Fukuoka 820-8502 Japan shimada@pluto.ai.kyutech.ac.jp Abstract In this paper we propose a method for a rat- ing prediction task. Each review consists of several ratings for a product, namely aspects. To predict the ratings of the aspects, we uti- lize not only aspect words, but also aspect sen- tences. First, our method detects aspect sen- tences by using a machine learning technique. Then, it incorporates words extracted from as- pect sentences with aspect word features. For estimating aspect likelihood of each word, we utilize the variance of words among aspects. Finally, it generates classifiers for each aspect by using the extracted features based on the aspect likelihood. Experimental result shows the effectiveness of features from aspect sen- tences. 1 Introduction As the World Wide Web rapidly grows, a huge num- ber of online documents are easily accessible on the Web. Finding information relevant to user needs has become increasingly important. The most important information on the Web is usually contained in the text. We obtain a huge number of review documents that include user’s opinions for products. Buying products, users usually survey the product reviews. More precise and effective methods for evaluating the products are useful for users. Many researchers have recently studied extraction and classification of opinions, namely sentiment analysis (Pang and Lee, 2008). For sentiment analysis, one of the most primitive studies is to classify a document into two classes; positive and negative opinions (Pang et al., 2002; Turney, 2002). One simple extension of p/n classifi- cation is a rating prediction task. It is a finer-grained task, as compared with the p/n classification. Several researchers have challenged rating prediction tasks in reviews (Goldberg and Zhu, 2006; Li et al., 2011; Okanohara and Tsujii, 2005; Pang and Lee, 2005). They are called “seeing stars.” These tasks han- dled an overall rating in the prediction. However, each review contains many descriptions about sev- eral aspects of a product. For example, they are “per- formance”, “user-friendliness” and “portability” for laptop PCs and “script”, “casting” and “music” for movies. Since reviewers judge not only the over- all polarity for a product but also details for it, pre- dicting stars of several aspects in a review is also one of the most important tasks in sentiment analy- sis, instead of a single overall rating. There are sev- eral studies to predict some stars in a review, namely “seeing several stars” or “aspect ratings” (Gupta et al., 2010; Pappas and Popescu-Belis, 2014; Shimada and Endo, 2008; Snyder and Barzilay, 2007). In this paper we propose a method for a rating pre- diction task with some aspects. In other words, we focus on multi-scale and multi-aspects rating predic- tion for reviews. We handle video game reviews with seven aspects and zero to five stars. Here we also focus on feature extraction for the prediction. The most common approach is usually based on fea- …

Content truncated to 500 words