GTI: An Unsupervised Approach for Sentiment Analysis in Twitter

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 533–538, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics GTI: An Unsupervised Approach for Sentiment Analysis in Twitter Milagros Fern´andez-Gavilanes, Tamara ´Alvarez-L´opez, Jonathan Juncal-Mart´ınez, Enrique Costa-Montenegro, Francisco Javier Gonz´alez-Casta˜no GTI Research Group AtlantTIC Centre, School of Telecommunication Engineering, University of Vigo 36310 Vigo, Spain {milagros.fernandez, talvarez, jonijm, kike}@gti.uvigo.es, javier@det.uvigo.es Abstract This paper presents the approach of the GTI Research Group to SemEval-2015 task 10 on Sentiment Analysis in Twitter, or more specif- ically, subtasks A (Contextual Polarity Disam- biguation) and B (Message Polarity Classifi- cation). We followed an unsupervised depen- dency parsing-based approach using a senti- ment lexicon, created by means of an auto- matic polarity expansion algorithm and Nat- ural Language Processing techniques. These techniques involve the use of linguistic pe- culiarities, such as the detection of polar- ity conflicts or adversative/concessive subor- dinate clauses. The results obtained confirm the competitive and robust performance of the system. 1 Introduction The domain of sentiment analysis has received in- creasing attention in recent years (Liu, 2012), partic- ularly due to the growth of the Internet and content generated by users of social networks and other plat- forms. Some of these, such as Twitter, allow people to express their opinions using colloquial, compact language. The result is a new form of expression that may in the long term become a source of ex- tremely valuable information. An increasing num- ber of companies are now focusing their market- ing campaigns on online comments, sentiments, and opinions of brands from clients or potential clients, and some are even trying to predict the acceptance and rejection of certain products using this informa- tion (Jansen et al., 2009). Even though the approaches used for this pur- pose are numerous and varied, broadly divided into two categories:they can be supervised machine-learning and unsupervised semantic-based approaches. The former are often classifiers built from features of a “bag of words” representation (Hu and Liu, 2004; Pak et al., 2010). In other words, they consist of automatically analyzing n-grams in search of recurrent combinations of opinion words. The latter aim at capturing and modeling linguistic knowledge through the use of dictionaries (Taboada et al., 2011) containing words that are tagged with their semantic orientation. These methods detect the words present in a text using different strategies in- volving lexics, syntax or semantics (Quinn et al., 2010) and then aggregate their values. Such meth- ods usually combine two or more levels of analysis. In recent years, work on sentiment classification using different types of texts has shown that special- ized methods are required. For example, emotions are not conveyed in the same manner in newspaper articles as in blogs, reviews, forums or other types of user-generated content (Balahur, 2013). Dealing with sentiment in Twitter, thus, requires an analy- sis of the characteristics of tweets and the design of adapted methods. This paper presents a method for sentiment analy- sis in English that uses dependency parsing to deter- …

Content truncated to 500 words