[LVIC-LIMSI]: Using Syntactic Features and Multi-polarity Words for Sentiment Analysis in Twitter

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 418–424, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics [LVIC-LIMSI]: Using Syntactic Features and Multi-polarity Words for Sentiment Analysis in Twitter Morgane Marchand1,2, Alexandru Lucian Ginsca1, Romaric Besanc¸on1, Olivier Mesnard1 (1) CEA-LIST, DIASI, LVIC CEA SACLAY - Nano-INNOV - Bt. 861 - Point courrier 173 91191 Gif-sur-Yvette Cedex, France (2) LIMSI-CNRS Bat 508, BP133,91403 Orsay Cedex morgane.marchand@cea.fr; alexandru.ginsca@cea.fr romaric.besancon@cea.fr; olivier.mesnard@cea.fr Abstract This paper presents the contribution of our team at task 2 of SemEval 2013: Sentiment Analysis in Twitter. We submitted a con- strained run for each of the two subtasks. In the Contextual Polarity Disambiguation subtask, we use a sentiment lexicon approach combined with polarity shift detection and tree kernel based classifiers. In the Message Polarity Clas- sification subtask, we focus on the influence of domain information on sentiment classifica- tion. 1 Introduction In the past decade, new forms of communication, such as microblogging and text messaging have emerged and became ubiquitous. These short mes- sages are often used to share opinions and sentiments. The Sentiment Analysis in Twitter task promotes re- search that will lead to a better understanding of how sentiment is conveyed in tweets and texts. In this paper, we describe our contribution at task 2 of Se- mEval 2013 (Wilson et al., 2013). For the Contextual Polarity Disambiguation subtask, covered in section 2, we use a system that combines a lexicon based approach to sentiment detection with two types of supervised learning methods, one used for polarity shift identification and one for tweet segment classi- fication in the absence of lexicon words. The third section presents the Message Polarity Classification subtask. We focus here on the influence of domain information on sentiment classification by detecting words that change their polarity across domains. 2 Task A: Contextual Polarity Disambiguation In this section we present our approach for the con- textual polarity disambiguation task in which, given a message containing a marked instance of a word or a phrase, the system has to determine whether that instance is positive, negative or neutral in that con- text. For this task, we submitted a single run using only the tweets provided by the organizers. 2.1 System description Based on the predominant strategy, sentiment anal- ysis systems can be divided into those that focus on sentiment lexicons together with a set of rules and those that rely on machine learning techniques. For this task, we use a mixed approach in which we first filter the tweets based on the occurrences of words from a sentiment lexicon and then apply different supervised learning methods on the grounds of this initial classification. In Figure 1 we detail the work- flow of our system. We use the + , − and ∗ symbols to denote a positive, negative and neutral tweet seg- ment, respectively. Also, we use the a → b notation when referring to a polarity shift from a to …

Content truncated to 500 words