CodeX: Combining an SVM Classifier and Character N-gram Language Models for Sentiment Analysis on Twitter Text

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 520–524, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics CodeX: Combining an SVM Classifier and Character N-gram Language Models for Sentiment Analysis on Twitter Text Qi Han, Junfei Guo and Hinrich Sch¨utze Institute for Nature Language Processing University of Stuttgart Stuttgart, Germany {hanqi, guojf}@ims.uni-stuttgart.de Abstract This paper briefly reports our system for the SemEval-2013 Task 2: sentiment analysis in Twitter. We first used an SVM classifier with a wide range of features, including bag of word features (unigram, bigram), POS fea- tures, stylistic features, readability scores and other statistics of the tweet being analyzed, domain names, abbreviations, emoticons in the Twitter text. Then we investigated the ef- fectiveness of these features. We also used character n-gram language models to address the problem of high lexical variation in Twit- ter text and combined the two approaches to obtain the final results. Our system is robust and achieves good performance on the Twitter test data as well as the SMS test data. 1 Introduction The challenge of the SemEval-2013 Task 2 (Task B) is the “Message Polarity Classification” (Wilson et al., 2013). Specifically, the task was to classify whether a given message has positive, negative or neutral sentiment; for messages conveying both pos- itive and negative sentiment, whichever is stronger should be chosen. In recent years, text messaging and microblog- ging such as tweeting has gained its popularity. Since these short messages are often used not only to discuss facts but also to share opinions and sen- timents, sentiment analysis on this type of data has lately become interesting. However, some features of this type of data make natural language process- ing challenging. For example, the messages are usu- ally short and the language used can be very in- formal, with misspellings, creative spellings, slang, URLs and special abbreviations. Some research has already been done attempting to address these prob- lems, to enable sentiment analysis on this type of data, in particular on Twitter data, and even to use the outcome of sentiment analysis to make predic- tions (Jansen et al., 2009; Barbosa and Feng, 2010; Bifet and Frank, 2010; Davidov et al., 2010; Jiang et al., 2011; Pak and Paroubek, 2010; Saif et al., 2012; Tumasjan et al., 2010). As the research mentioned above, our system used a machine learning based approach for sentiment analysis. Our system combines results from an SVM classifier using a wide range of features as well as votes derived from character n-gram language mod- els to do the final prediction. The rest of this paper is organized as follows. Sec- tion 2 describes the features used for the SVM clas- sifier. Section 3 describes how the votes from char- acter n-gram language models were derived. Section 4 describes the details of our method. And finally section 5 presents the results. 2 Features We pre-processed the tweets as follows: i) tok- enized the tweets using …

Content truncated to 500 words