NTNU: Domain Semi-Independent Short Message Sentiment Classification

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 430–437, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics NTNU: Domain Semi-Independent Short Message Sentiment Classification Øyvind Selmer Mikael Brevik Bj¨orn Gamb¨ack Lars Bungum Department of Computer and Information Science Norwegian University of Science and Technology (NTNU) Sem Sælands vei 7–9, NO–7491 Trondheim, Norway {oyvinsel,mikaelbr}@stud.ntnu.no {gamback,larsbun}@idi.ntnu.no Abstract The paper describes experiments using grid searches over various combinations of ma- chine learning algorithms, features and pre- processing strategies in order to produce the optimal systems for sentiment classification of microblog messages. The approach is fairly domain independent, as demonstrated by the systems achieving quite competitive results when applied to short text message data, i.e., input they were not originally trained on. 1 Introduction The informal texts in microblogs such as Twitter and on other social media represent challenges for traditional language processing systems. The posts (“tweets”) are limited to 140 characters and often contain misspellings, slang and abbreviations. On the other hand, the posts are often opinionated in nature as a very result of their informal character, which has led Twitter to being a gold mine for sen- timent analysis (SA). SA for longer texts, such as movie reviews, has been explored since the 1990s;1 however, the limited amount of attributes in tweets makes the feature vectors shorter than in documents and the task of analysing them closely related to phrase- and sentence-level SA (Wilson et al., 2005; Yu and Hatzivassiloglou, 2003). Hence there are no guarantees that algorithms that perform well on document-level SA will do as well on tweets. On the other hand, it is possible to exploit some of the special features of the web language, e.g., emoticons 1See Pang and Lee (2008); Feldman (2013) for overviews. and emotionally loaded abbreviations. Thus the data will normally go through some preprocessing before any classification is attempted, e.g., by filtering out Twitter specific symbols and functions, in particular retweets (reposting another user’s tweet), mentions (’@’, tags used to mention another user), hashtags (’#’, used to tag a tweet to a certain topic), emoti- cons, and URLs (linking to an external resource, e.g., a news article or a photo). The first system to re- ally use Twitter as a corpus was created as a student course project at Stanford (Go et al., 2009). Pak and Paroubek (2010) experimented with sentiment clas- sification of tweets using Support Vector Machines and Conditional Random Fields, benchmarked with a Na¨ıve Bayes Classifier baseline, but were unable to beat the baseline. Later, and as Twitter has grown in popularity, many other systems for Twitter Senti- ment Analysis (TSA) have been developed (see, e.g., Maynard and Funk, 2011; Mukherjee et al., 2012; Saif et al., 2012; Chamlertwat et al., 2012). Clearly, it is possible to classify the sentiment of tweets in a single step; however, the approach to TSA most used so far is a two-step strategy where the first step is subjectivity classification and …

Content truncated to 500 words