Review Classification Using Semantic Features and Run-Time Weighting

Review Classification Using Semantic Features and Run-Time Weighting Chung-chi Huang a, Meng-chiech Lee b, Zhe-nan Lin b, and Jason S. Chang b aInstitute of Information Systems and Applications, National Tsing Hua University, HsinChu, Taiwan 300, R.O.C. u901571@gmail.com bDepartment of Computer Science, National Tsing Hua University, HsinChu, Taiwan 300, R.O.C. {iamrabbit37, rexkimta, jason.jschang}@gmail.com Abstract. We introduce a method for learning to assign suitable sentiment ratings to review articles. In our approach, reviews are transformed into collections of n-gram and semantic word class features aimed at maximizing the probability of classifying them into accurate ratings. The method involves automatically segmenting review articles into sentences and automatically estimating associations between features and sentiment ratings via machine learning techniques. At run-time, a simple weighting strategy is performed to give extra weights to features in potential evaluative sentences (e.g., the first, the last sentences and sentences with adverbs) from others. Experiments show that word class information alleviates data sparseness problem facing higher-level n-grams (e.g., bigrams and trigrams) and that our model using both training-time n-gram and semantic features and run-time weighting mechanism outperforms a strong baseline with surface n-gram features by 2.5% relatively. Keywords: sentiment, semantic orientations, classification, sentiment ratings, machine learning techniques. 1 Introduction With the accessibility to the Internet in recent years, any Web user can easily be an article reader or a content provider. Web contents can be coarsely categorized into two groups: ones on knowledge and ones experience. Knowledge of any discipline is provided on on-line forums or on certain Web pages (e.g., WIKIPEDIA1) by the collaborative effort of a number of domain experts. On the other hand, more and more Internet users author their personal experiences on the Web. Experiences of using a high-tech product, dining in a restaurant, watching a movie and so on are documented. These publicly-available remarks are especially valuable in terms of marketing and recommendation. Take Yahoo!奇摩生活+2 (Yahoo! Kimo Life Style) for example. It is a platform where Web users can author their experiences with or opinions toward the service-providing businesses including restaurants, hotels, amusement parks and etc. Figure 1 shows a restaurant review on Yahoo! Kimo Life Style. In this application, the overall evaluation on a service provider is rated via a number of stars. More stars, more positive the evaluation. We present a model that automatically learns to rate review articles according to their semantic orientations. Reviews are ranked from one to five with larger integers meaning Copyright 2009 by Chung-chi Huang, Meng-chiech Lee, Zhe-nan Lin, and Jason S. Chang 1 http://en.wikipedia.org/wiki/Wiki 2 http://tw.lifestyle.yahoo.com/ 200 23rd Pacific Asia Conference on Language, Information and Computation, pages 200–209 reviewers’ attitude toward the service is more positive and with smaller ones meaning reviewers’ experience with the enterprise is more negative. During training, our model leverages machine learning techniques (e.g., maximum entropy or conditional random fields) to estimate associations between various ratings (i.e., one to five) and features, including surface n-gram features and semantic features. Semantic classes of words are exploited in our model in order to alleviate …

Content truncated to 500 words