AMI&ERIC: How to Learn with Naive Bayes and Prior Knowledge: an Application to Sentiment Analysis

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 364–368, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics AMI&ERIC: How to Learn with Naive Bayes and Prior Knowledge: an Application to Sentiment Analysis Mohamed Dermouche1,2, Leila Khouas1, 1AMI Software R&D 1475 av. A. Einstein 34000 Montpellier, France mde@amisw.com lkh@amisw.com Julien Velcin2 and Sabine Loudcher2 2Universit´e de Lyon, ERIC (Lyon 2) 5 av. P. Mend`es-France 69676 Bron Cedex, France julien.velcin@univ-lyon2.fr sabine.loudcher@univ-lyon2.fr Abstract In this paper, we describe our system that par- ticipated in SemEval-2013, Task 2.B (senti- ment analysis in Twitter). Our approach con- sists of adapting Naive Bayes probabilities in order to take into account prior knowledge (represented in the form of a sentiment lex- icon). We propose two different methods to efficiently incorporate prior knowledge. We show that our approach outperforms the clas- sical Naive Bayes method and shows compet- itive results with SVM while having less com- putational complexity. 1 Introduction With the advent of Internet microblogging, social networks, like Twitter1 and Facebook2, have brought about a real revolution in our way of communi- cating. People share their opinions of everyday life without taboos or restrictions thanks to the anonymity offered by these tools, which makes them a valuable source of information rather rich of sub- jective data. These data can be mined using sen- timent analysis as a means to understand people’s feelings towards a political cause or what people are thinking about a product or a service. Recent works showed that Twitter sentiments can be correlated to box-office revenues (Asur and Huberman, 2010) or political polls (O’Connor et al., 2010). Machine learning methods, like Naive Bayes (NB) and Support Vector Machines (SVM), have been widely used in sentiment analysis (Pang et al., 1http://www.twitter.com/ 2http://www.facebook.com/ 2002; Pak and Paroubek, 2010). One major problem with these methods, and in particular NB, is that the model is built only on the learning data which can lead to overfitting. In this paper, we describe our ap- proach that participated in SemEval-2013, Task 2.B (sentiment analysis in Twitter) (Wilson et al., 2013). Our approach consists of learning with both NB and prior knowledge. We show that our approach out- performs the classical NB method and gives com- petitive results compared to SVM while having less computational complexity. The remainder of this paper is organized as fol- lows: prior works on sentiment analysis are dis- cussed in Section 2. The proposed approach is de- tailed in Section 3. Then, experiments and results are given in Section 4 and 5. 2 Background Sentiment analysis is a text mining task which deals with the feelings expressed explicitly or implicitly in a textual content. It concerns subjectivity anal- ysis (subjective/objective), opinion mining (posi- tive/negative/neutral), strength analysis, etc. Al- though the term “sentiment analysis” includes all these tasks, it often refers to opinion mining. Sen- timent analysis methods can be categorized into ma- chine learning, linguistic and hybrid methods. Machine learning methods are …

Content truncated to 500 words