Sentiment Classification in Resource-Scarce Languages by using Label Propagation

Sentiment Classification in Resource-Scarce Languages by using Label Propagation Yong Ren a, Nobuhiro Kaji b, Naoki Yoshinaga b, Masashi Toyoda b, and Masaru Kitsuregawa b aGraduate School of Information Science and Technology, University of Tokyo, Japan renyong@tkl.iis.u-tokyo.ac.jp bInstitute of Industrial Science, University of Tokyo, Japan {kaji,ynaga,toyoda,kitsure}@tkl.iis.u-tokyo.ac.jp Abstract. With the advent of consumer generated media (e.g., Amazon reviews, Twitter, etc.), sentiment classification becomes a heated topic. Previous work heavily relies on a large amount of linguistic resources, which are difficult to obtain in resource-scarce languages. To overcome this problem, we investigate the usefulness of label propagation, which is a graph-based semi-supervised learning method. Extensive experimental evaluation on three real datasets demonstrated that label propagation performs more stable than support vector machines (SVMs) and transductive support vector machines (TSVMs) in a document-level sentiment classification task for resource-scarce languages (Chinese in our case). Keywords: sentiment classification, label propagation, semi-supervised learning 1 Introduction Over the last decade, document-level sentiment classification has attracted much attention from NLP researchers; its potential applications include opinion summarization and opinion mining (Pang and Lee, 2008). Most of the existing methods locate sentiment classification as a supervised classification problem and train a reliable classifier from a large amount of labeled data (Pang et al., 2002; Mullen and Collier, 2004; Matsumoto et al., 2005; Gamon, 2005). The main disadvantage of such supervised approaches is that it is quite expensive in both time and labor to annotate a large amount of training data. Unfortunately, in some languages such as Chinese and Hindi, a sufficient amount of training data is not always available. Sentiment classification becomes a quite challenging problem for such resource-scarce languages. While some studies have tackled this problem (Wan, 2009; Dasgupta and Ng, 2009), they still require substantial human efforts or specific linguistic resources that are only available in particular languages as we will see later in Section 2. We therefore want to develop a low-cost, general method that can be readily applicable to sentiment classification in any languages. In this paper, we explore the use of label propagation (LP) (Zhu and Ghahramani, 2002) in building a document-level sentiment classifier under a minimally-supervised setting, where we have only a small number of labeled reviews other than the target reviews that we want to classify. Having a similarity graph whose vertices are labeled or unlabeled instances (reviews, here) and edges represent the similarity between the vertices, LP infers the label (sentiment polarity) of unlabeled (target) reviews based on labeled reviews that are similar to the target reviews. The key in applying LP to document-level sentiment classification is therefore in the way to represent reviews on the graph and to define the similarity between the them. We thus investigate the impact of the review representation and the similarity measure on the classification performance. The main contributions of our work are summarized as follows: Copyright 2011 by Yong Ren, Nobuhiro Kaji, Naoki Yoshinaga, Masashi Toyoda, and Masaru Kitsuregawa 25th Pacific Asia Conference on Language, Information and Computation, pages 420–429 420 • We evaluate …

Content truncated to 500 words