SZTE-NLP: Sentiment Detection on Twitter Messages

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 549–553, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics SZTE-NLP: Sentiment Detection on Twitter Messages Viktor Hangya, G´abor Berend, Rich´ard Farkas University of Szeged Department of Informatics hangyav@gmail.com, {berendg,rfarkas}@inf.u-szeged.hu Abstract In this paper we introduce our contribution to the SemEval-2013 Task 2 on “Sentiment Analysis in Twitter”. We participated in “task B”, where the objective was to build mod- els which classify tweets into three classes (positive, negative or neutral) by their con- tents. To solve this problem we basically fol- lowed the supervised learning approach and proposed several domain (i.e. microblog) spe- cific improvements including text preprocess- ing and feature engineering. Beyond the su- pervised setting we also introduce some early results employing a huge, automatically anno- tated tweet dataset. 1 Introduction In the past few years, the popularity of social me- dia has increased. Many studies have been made in the area (Jansen et al., 2009; O’Connor et al., 2010; Bifet and Frank, 2010; Sang and Bos, 2012). Peo- ple post messages on a variety of topics, for example products, political issues, etc. Thus a big amount of user generated data is created day-by-day. The man- ual processing of this data is impossible, therefore automatic procedures are needed. In this paper we introduce an approach which is able to assign sentiment labels to Twitter messages. More precisely, it classifies tweets into positive, neg- ative or neutral polarity classes. The system partici- pated in the SemEval-2013 Task 2: Sentiment Anal- ysis in Twitter, Task–B Message Polarity Classifica- tion (Wilson et al., 2013). In our approach we used a unigram based supervised model because it has been shown that it works well on short messages like tweets (Jiang et al., 2011; Barbosa and Feng, 2010; Agarwal et al., 2011; Liu, 2010). We reduced the size of the dictionary by normalizing the messages and by stop word filtering. We also explored novel features which gave us information on the polarity of a tweet, for example we made use of the acronyms in messages. In the “constrained” track of Task–B we used the given training and development data only. For the “unconstrained” track we downloaded tweets using the Twitter Streaming API1 and automatically anno- tated them. We present some preliminary results on exploiting this huge dataset for training our classi- fier. 2 Approach At the beginning of our experiments we used a unigram-based supervised model. Later on, we re- alized that the size of our dictionary is huge, so in the normalization phase we tried to reduce the number of words in it. We investigated novel fea- tures which contain information on the polarity of the messages. Using these features we were able to improve the precision of our classifier. For imple- mentation we used the MALLET toolkit, which is a Java-based package for natural language processing (McCallum, 2002). 2.1 Normalization One reason for the unusually big dictionary size is …

Content truncated to 500 words