mib at SemEval-2016 Task 4a: Exploiting lexicon based features for Sentiment Analysis in Twitter

Proceedings of SemEval-2016, pages 133–138, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics MIB at SemEval-2016 Task 4a: Exploiting lexicon-based features for sentiment analysis in Twitter Vittoria Cozza IIT-CNR Pisa Italy v.cozza@iit.cnr.it Marinella Petrocchi IIT-CNR Pisa Italy m.petrocchi@iit.cnr.it Abstract This work presents our team solution for task 4a (Message Polarity Classification) at the Se- mEval 2016 challenge. Our experiments have been carried out over the Twitter dataset pro- vided by the challenge. We follow a super- vised approach, exploiting a SVM polyno- mial kernel classifier trained with the chal- lenge data. The classifier takes as input ad- vanced NLP features. This paper details the features and discusses the achieved results. 1 Introduction Revealing the sentiment behind a text is motivated by several reasons, e.g., to figure out how many opinions on a certain topic are positive or negative. Also, it could be interesting to span positivity and negativity across a n-point scale. As an example, a five-point scale is now widespread in digital sce- narios where human ratings are involved: Amazon, TripAdvisor, Yelp, and many others, adopt the scale for letting their users rating products and services. Under the big hat of sentiment analysis (Liu, 2012), polarity recognition attempts to classify texts into positive or negative, while the rating infer- ence task tries to identify different shades of pos- itivity and negativity, e.g., from strongly-negative, to strongly-positive. There currently exists a num- ber of popular challenges on the matter, as those in- cluded in the SemEval series on evaluations of com- putational semantic analysis systems1. Both polarity recognition and rating inference have been applied 1https://en.wikipedia.org/wiki/SemEval to recommendation systems. Recently, Academia has been focusing on the feasibility to apply senti- ment analysis tasks to very short and informal texts, such as tweets (see, e.g. (Rosenthal et al., 2015)). This paper shows the description of the system that we have set up for participating into the Semeval 2016 challenge in (Nakov et al., 2016b), task 4a (Message Polarity Classification). We have adopted a supervised approach, a SVM polynomial kernel classifier trained with the data provided by the chal- lenge, after extracting lexical and lexicon features from such data. The paper is organised as follows. Next section briefly addresses related work in the area. Section 3 describes the features extracted from the training data. In Section 4, we present the results of our at- tempt to answer to the challenge. Finally, we give concluding remarks. 2 Related work In the last recent years, the Semeval tasks series challenges the polarity evaluation of tweets. This represents a detachment from the traditional polar- ity detection task. Tweets usually features the use of an informal language, with mispellings, new words, urls, abbreviations and specific symbols (like RT for “re-tweet” and # for hashtags, which are a type of tagging for Twitter messages). Existing approaches and open issues on how to handle such new chal- lenges are in related work like (Kouloumpis et al., 2011; Barbosa and Feng, 2010). At the 2015 challenge (Rosenthal …

Content truncated to 500 words