AVAYA: Sentiment Analysis on Twitter with Self-Training and Polarity Lexicon Expansion

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 333–340, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics AVAYA: Sentiment Analysis on Twitter with Self-Training and Polarity Lexicon Expansion Lee Becker, George Erhart, David Skiba and Valentine Matula Avaya Labs Research 1300 West 120th Avenue Westminster, CO 80234, USA {beckerl,gerhart,dskiba,matula}@avaya.com Abstract This paper describes the systems submitted by Avaya Labs (AVAYA) to SemEval-2013 Task 2 - Sentiment Analysis in Twitter. For the constrained conditions of both the message polarity classification and contextual polarity disambiguation subtasks, our approach cen- ters on training high-dimensional, linear clas- sifiers with a combination of lexical and syn- tactic features. The constrained message po- larity model is then used to tag nearly half a million unlabeled tweets. These automati- cally labeled data are used for two purposes: 1) to discover prior polarities of words and 2) to provide additional training examples for self-training. Our systems performed compet- itively, placing in the top five for all subtasks and data conditions. More importantly, these results show that expanding the polarity lexi- con and augmenting the training data with un- labeled tweets can yield improvements in pre- cision and recall in classifying the polarity of non-neutral messages and contexts. 1 Introduction The past decade has witnessed a massive expansion in communication from long-form delivery such as e-mail to short-form mechanisms such as mi- croblogging and short messaging service (SMS) text messages. Simultaneously businesses, media out- lets, and investors are increasingly relying on these messages as sources of real-time information and are increasingly turning to sentiment analysis to dis- cover product trends, identify customer preferences, and categorize users. While a variety of corpora ex- ist for developing and evaluating sentiment classi- fiers for long-form texts such as product reviews, there are few such resources for evaluating senti- ment algorithms on microblogs and SMS texts. The organizers of SemEval-2013 task 2, have be- gun to address this resource deficiency by coordi- nating a shared evaluation task for Twitter sentiment analysis. In doing so they have assembled corpora in support of the following two subtasks: Task A - Contextual Polarity Disambiguation “Given a message containing a marked in- stance of a word or phrase, determine whether that instance is positive, negative or neutral in that context.” Task B - Message Polarity Classification “Given a message, classify whether the message is of positive, negative, or neutral sentiment. For messages conveying both a positive and negative sentiment, whichever is the stronger sentiment should be chosen.” This paper describes the systems submitted by Avaya Labs for participation in subtasks A and B. Our goal for this evaluation was to investigate the usefulness of dependency parses, polarity lexicons, and unlabeled tweets for sentiment classification on short messages. In total we built four systems for SemEval-2013 task 2. For task B we developed a constrained model using supervised learning, and an unconstrained model that used semi-supervised learning in the form of self-training and polarity …

Content truncated to 500 words