The Bag-of-Opinions Method for Review Rating Prediction from Sparse Text Patterns

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 913–921, Beijing, August 2010 The Bag-of-Opinions Method for Review Rating Prediction from Sparse Text Patterns Lizhen Qu Max-Planck Institute for Informatics lqu@mpii.mpg.de Georgiana Ifrim Bioinformatics Research Centre ifrim@birc.au.dk Gerhard Weikum Max-Planck Institute for Informatics weikum@mpii.mpg.de Abstract The problem addressed in this paper is to predict a user’s numeric rating in a prod- uct review from the text of the review. Un- igram and n-gram representations of text are common choices in opinion mining. However, unigrams cannot capture impor- tant expressions like “could have been bet- ter”, which are essential for prediction models of ratings. N-grams of words, on the other hand, capture such phrases, but typically occur too sparsely in the train- ing set and thus fail to yield robust pre- dictors. This paper overcomes the limita- tions of these two models, by introducing a novel kind of bag-of-opinions represen- tation, where an opinion, within a review, consists of three components: a root word, a set of modifier words from the same sen- tence, and one or more negation words. Each opinion is assigned a numeric score which is learned, by ridge regression, from a large, domain-independent cor- pus of reviews. For the actual test case of a domain-dependent review, the re- view’s rating is predicted by aggregat- ing the scores of all opinions in the re- view and combining it with a domain- dependent unigram model. The paper presents a constrained ridge regression al- gorithm for learning opinion scores. Ex- periments show that the bag-of-opinions method outperforms prior state-of-the-art techniques for review rating prediction. 1 Introduction 1.1 Motivation Opinion mining and sentiment analysis has be- come a hot research area (Pang and Lee, 2008). There is ample work on analyzing the sentiments of online-review communities where users com- ment on products (movies, books, consumer elec- tronics, etc.), implicitly expressing their opinion polarities (positive, negative, neutral), and also provide numeric ratings of products (Titov and McDonald, 2008b; Lerman et al., 2009; Hu and Liu, 2004; Titov and McDonald, 2008a; Pang and Lee, 2005; Popescu and Etzioni, 2005a). Al- though ratings are more informative than polari- ties, most prior work focused on classifying text fragments (phrases, sentences, entire reviews) by polarity. However, a product receiving mostly 5- star reviews exhibits better customer purchase be- havior compared to a product with mostly 4-star reviews. In this paper we address the learning and prediction of numerical ratings from review texts, and we model this as a metric regression problem over an appropriately defined feature space. Formally, the input is a set of rated documents (i.e., reviews), {xi, yi}N i=1, where xi is a sequence of word-level unigrams (w1, ..., wl) and yi ∈ R is a rating. The goal is to learn a function f(x) that maps the word vector x into a numerical rating ˆy, which indicates both the polarity and strength of the opinions expressed in a document. Numerical review rating prediction is harder than classifying by polarity. Consider …

Content truncated to 500 words