SAIL: A hybrid approach to sentiment analysis

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 438–442, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics SAIL: A hybrid approach to sentiment analysis Nikolaos Malandrakis1, Abe Kazemzadeh2, Alexandros Potamianos3, Shrikanth Narayanan1 1 Signal Analysis and Interpretation Laboratory (SAIL), USC, Los Angeles, CA 90089, USA 2 Annenberg Innovation Laboratory (AIL), USC, Los Angeles, CA 90089, USA 3Department of ECE, Technical University of Crete, 73100 Chania, Greece malandra@usc.edu, kazemzad@usc.edu, potam@telecom.tuc.gr, shri@sipi.usc.edu Abstract This paper describes our submission for Se- mEval2013 Task 2: Sentiment Analysis in Twitter. For the limited data condition we use a lexicon-based model. The model uses an af- fective lexicon automatically generated from a very large corpus of raw web data. Statistics are calculated over the word and bigram af- fective ratings and used as features of a Naive Bayes tree model. For the unconstrained data scenario we combine the lexicon-based model with a classifier built on maximum entropy language models and trained on a large exter- nal dataset. The two models are fused at the posterior level to produce a final output. The approach proved successful, reaching rank- ings of 9th and 4th in the twitter sentiment analysis constrained and unconstrained sce- nario respectively, despite using only lexical features. 1 Introduction The analysis of the emotional content of text, is relevant to numerous natural language processing (NLP), web and multi-modal dialogue applications. To that end there has been a significant scientific effort towards tasks like product review analysis (Wiebe and Mihalcea, 2006; Hu and Liu, 2004), speech emotion extraction (Lee and Narayanan, 2005; Lee et al., 2002; Ang et al., 2002) and pure text word (Esuli and Sebastiani, 2006; Strappar- ava and Valitutti, 2004) and sentence (Turney and Littman, 2002; Turney and Littman, 2003) level emotion extraction. The rise of social media in recent years has seen a shift in research focus towards them, particularly twitter. The large volume of text data available is particularly useful, since it allows the use of com- plex machine learning methods. Also important is the interest on the part of companies that are actively looking for ways to mine social media for opinions and attitudes towards them and their products. Sim- ilarly, in journalism there is interest in sentiment analysis for a way to process and report on the public opinion about current events (Petulla, 2013). Analyzing emotion expressed in twitter borrows from other tasks related to affective analysis, but also presents unique challenges. One common is- sue is the breadth of content available in twitter: a more limited domain would make the task easier, however there are no such bounds. There is also a significant difference in the form of language used in tweets. The tone is informal and typographical and grammatical errors are very common, making even simple tasks, like Part-of-Speech tagging much harder. Features like hashtags and emoticons can also be helpful (Davidov et al., 2010). This paper describes our submissions for Se- mEval …

Content truncated to 500 words