ASVUniOfLeipzig: Sentiment Analysis in Twitter using Data-driven Machine Learning Techniques

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 450–454, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics ASVUniOfLeipzig: Sentiment Analysis in Twitter using Data-driven Machine Learning Techniques Robert Remus Natural Language Processing Group, Department of Computer Science, University of Leipzig, Germany rremus@informatik.uni-leipzig.de Abstract This paper describes University of Leipzig’s approach to SemEval-2013 task 2B on Sen- timent Analysis in Twitter: message polar- ity classification. Our system is designed to function as a baseline, to see what we can accomplish with well-understood and purely data-driven lexical features, simple general- izations as well as standard machine learning techniques: We use one-against-one Support Vector Machines with asymmetric cost fac- tors and linear “kernels” as classifiers, word uni- and bigrams as features and additionally model negation of word uni- and bigrams in word n-gram feature space. We consider gen- eralizations of URLs, user names, hash tags, repeated characters and expressions of laugh- ter. Our method ranks 23 out of all 48 partic- ipating systems, achieving an averaged (pos- itive, negative) F-Score of 0.5456 and an av- eraged (positive, negative, neutral) F-Score of 0.595, which is above median and average. 1 Introduction In SemEval-2013’s task 2B on Sentiment Analysis in Twitter, given a Twitter message, i.e. a tweet, the goal is to classify whether this tweet is of positive, negative, or neutral polarity (Wilson et al., 2013), i.e. the task is a ternary polarity classification. Due to Twitter’s growing popularity, the availabil- ity of large amounts of data that go along with that and the fact, that many people freely express their opinion on virtually everything using Twitter, re- search on sentiment analysis in Twitter has received a lot of attention lately (Go et al., 2009; Pak and Paroubek, 2010). Language is usually used casu- ally in Twitter and exhibits interesting properties. Therefore, some studies specifically address certain issues, e.g. a tweet’s length limitation of 140 char- acters, some studies leverage certain language char- acteristics, e.g. the presence of emoticons etc. Davidov et al. (2010) identify various “sentiment types” defined by Twitter hash tags (e.g. #bored) and smileys (e.g. :S) using words, word n-grams, punctuation marks and patterns as features. Bar- bosa and Feng (2010) map words to more general representations, i.e. part of speech (POS) tags and the words’ prior subjectivity and polarity. Addi- tionally, they count the number of re-tweets, hash tags, replies, links etc. They then combine the out- puts of 3 online sources of labeled but noisy and bi- ased Twitter data into a more robust classification model. Saif et al. (2012) also address data sparsity via word clustering methods, i.e. semantic smooth- ing and sentiment-topics extraction. Agarwal et al. (2011) contrast a word unigram model, a tree ker- nel model and a model of various features, e.g. POS tag counts, summed up prior polarity scores, pres- ence or absence of capitalized text, all applied to bi- nary and ternary polarity classification. Kouloumpis et al. (2011) show …

Content truncated to 500 words