RTRGO: Enhancing the GU-MLT-LT System for Sentiment Analysis of Short Messages

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 497–502, Dublin, Ireland, August 23-24, 2014. RTRGO: Enhancing the GU-MLT-LT System for Sentiment Analysis of Short Messages Tobias G¨unther Retresco GmbH retresco.de email@tobias.io Jean Vancoppenolle ferret go GmbH ferret-go.com jean.vcop@gmail.com Richard Johansson University of Gothenburg www.svenska.gu.se richard.johansson@gu.se Abstract This paper describes the enhancements made to our GU-MLT-LT system (G¨unther and Furrer, 2013) for the SemEval-2014 re-run of the SemEval-2013 shared task on sentiment analysis in Twitter. The changes include the usage of a Twitter-specific to- kenizer, additional features and sentiment lexica, feature weighting and random sub- space learning. The improvements result in an increase of 4.18 F-measure points on this year’s Twitter test set, ranking 3rd. 1 Introduction Automatic analysis of sentiment expressed in text is an active research area in natural language pro- cessing with obvious commercial interest. In the simplest formulation of the problem, sentiment analysis is framed as a categorization problem over documents, where the set of categories is typically a set of polarity values, such as posi- tive, neutral, and negative. Many approaches to document-level sentiment classification have been proposed. For an overview see e.g. Liu (2012). Text in social media and in particular microblog messages are a challenging text genre for senti- ment classification, as they introduce additional problems such as short text length, spelling vari- ation, special tokens, topic variation, language style and multilingual content. Following Pang et al. (2002), most sentiment analysis systems have been based on standard text categorization tech- niques, e.g. training a classifier using some sort of bag-of-words feature representation. This is also true for sentiment analysis of microblogs. Among This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ the first to work specifically with Twitter1 data were Go et al. (2009), who use emoticons as labels for the messages. Similarly, Davidov et al. (2010), Pak and Paroubek (2010), and Kouloumpis et al. (2011) use this method of distant supervision to overcome the data acquisition barrier. Barbosa and Feng (2010) make use of three different senti- ment detection websites to label messages and use mostly non-lexical features to improve the robust- ness of their classifier. Bermingham and Smeaton (2010) investigate the impact of the shortness of Tweets on sentiment analysis and Speriosu et al. (2011) propagate information from seed labels along a linked structure that includes Twitter’s follower graph. There has also been work on lexicon-based approaches to sentiment analysis of microblogs, such as O’Connor et al. (2010), Thel- wall et al. (2010) and Zhang et al. (2011). For a detailed discussion see G¨unther (2013). In 2013, the International Workshop on Se- mantic Evaluation (SemEval) organized a shared task on sentiment analysis in Twitter (Nakov et al., 2013) to enable a better comparison of dif- ferent approaches for sentiment analysis of mi- croblogs. The shared task consisted of two sub- tasks: one on recognizing contextual polarity of a given subjective expression …

Content truncated to 500 words