FBK: Sentiment Analysis in Twitter with Tweetsted

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 466–470, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics FBK: Sentiment Analysis in Twitter with Tweetsted Md. Faisal Mahbub Chowdhury FBK and University of Trento, Italy fmchowdhury@gmail.com Marco Guerini Trento RISE, Italy marco.guerini@trentorise.eu Sara Tonelli FBK, Trento, Italy satonelli@fbk.eu Alberto Lavelli FBK, Trento, Italy lavelli@fbk.eu Abstract This paper presents the Tweetsted system im- plemented for the SemEval 2013 task on Sen- timent Analysis in Twitter. In particular, we participated in Task B on Message Polar- ity Classification in the Constrained setting. The approach is based on the exploitation of various resources such as SentiWordNet and LIWC. Official results show that our approach yields a F-score of 0.5976 for Twitter mes- sages (11th out of 35) and a F-score of 0.5487 for SMS messages (8th out of 28 participants). 1 Introduction Microblogging is currently a very popular commu- nication tool where millions of users share opinions on different aspects of life. For this reason it is a valuable source of data for opinion mining and sen- timent analysis. Working with such type of texts presents chal- lenges for NLP beyond those typically encountered when dealing with more traditional texts, such as newswire data. Tweets are short, the language used is very informal, with creative spelling and punctua- tion, misspellings, slang, new words, URLs, genre- specific terminology and abbreviations, and #hash- tags. These characteristics need to be handled with specific approaches. This paper presents the approach adopted for the SemEval 2013 task on Sentiment Analysis in Twit- ter, in particular Task B on Message Polarity Clas- sification in the Constrained setting (i.e., using the provided training data only). The goal of Task B on Message Polarity Classi- fication is the following: given a message, decide whether it expresses a positive, negative, or neutral sentiment. For messages conveying both a positive and a negative sentiment, whichever is the stronger sentiment should be chosen. Two modalities are possible: (1) Constrained (us- ing the provided training data only; other resources, such as lexica, are allowed; however, it is not al- lowed to use additional tweets/SMS messages or ad- ditional sentences with sentiment annotations); and (2) Unconstrained (using additional data for train- ing, e.g., additional tweets/SMS messages or addi- tional sentences annotated for sentiment). We par- ticipated in the Constrained modality. We adopted a supervised machine learning (ML) approach based on various contextual and seman- tic features. In particular, we exploited resources such as SentiWordNet (Esuli and Sebastiani, 2006), LIWC (Pennebaker and Francis, 2001), and the lex- icons described in Mohammad et al. (2009). Critical features include: whether the mes- sage contains intensifiers, adjectives, interjections, presence of positive or negative emoticons, pos- sible message polarity based on SentiWordNet scores (Esuli and Sebastiani, 2006; Gatti and Guerini, 2012), scores based on LIWC cate- gories (Pennebaker and Francis, 2001), negated words, etc. 2 System Description Our supervised ML-based approach relies on Sup- port Vector Machines (SVMs). …

Content truncated to 500 words