A Two-Stage Classifier for Sentiment Analysis

International Joint Conference on Natural Language Processing, pages 897–901, Nagoya, Japan, 14-18 October 2013. A Two-Stage Classifier for Sentiment Analysis Dai Quoc Nguyen and Dat Quoc Nguyen and Son Bao Pham Faculty of Information Technology University of Engineering and Technology Vietnam National University, Hanoi {dainq, datnq, sonpb}@vnu.edu.vn Abstract In this paper, we present a study applying re- ject option to build a two-stage sentiment po- larity classification system. We construct a Naive Bayes classifier at the first stage and a Support Vector Machine at the second stage, in which documents rejected at the first stage are forwarded to be classified at the second stage. The obtained accuracies are comparable to other state-of-the-art results. Furthermore, experiments show that our classifier requires less training data while still maintaining rea- sonable classification accuracy. 1 Introduction The rapid growth of the Web supports human users to easily express their reviews about such entities as prod- ucts, services, events and their properties as well as to find and evaluate the others’ opinions. This brings new challenges for building systems to categorize and un- derstand the sentiments in those reviews. In particular, document-level sentiment classifica- tion systems aim to determine either a positive or neg- ative opinion in a given opinionated document (Tur- ney, 2002; Liu, 2010). In order to construct these systems, classification-based approaches (Pang et al., 2002; Pang and Lee, 2004; Mullen and Collier, 2004; Whitelaw et al., 2005; Kennedy and Inkpen, 2006; Martineau and Finin, 2009; Maas et al., 2011; Tu et al., 2012; Wang and Manning, 2012) utilizing machine learning to automatically identify document-level sen- timent polarity are still mainstream methods obtaining state-of-the-art performances. It is because of possibly combining various features such as: bag of words, syn- tactic and semantic representations as well as exploit- ing lexicon resources (Wilson et al., 2005; Ng et al., 2006; Taboada et al., 2011) like SentiWordNet (Bac- cianella et al., 2010). In these systems, Naive Bayes (NB) and Support Vector Machine (SVM) are often ap- plied for training learning models as they are frequently used as baseline methods in task of text classification (Wang and Manning, 2012). Although NBs are very fast classifiers requiring a small amount training data, there is a loss of accuracy due to the NBs’ conditional independence assumption. On the other hand, SVMs achieve state-of-the-art results in various classification tasks; however, they may be slow in the training and testing phases. In pattern recognition systems, reject option (Chow, 1970; Pudil et al., 1992; Fumera et al., 2000; Fumera et al., 2004) is introduced to improve classification re- liability. Although it is very useful to apply reject op- tion in many pattern recognition/classification systems, it has not been considered in a sentiment classification application so far. In this paper, we introduce a study combining the advantages of both NB and SVM classifiers into a two- stage system by applying reject option for document- level sentiment classification. In the first stage of our system, a NB classifier, which is trained based on a …

Content truncated to 500 words