bwbaugh : Hierarchical sentiment analysis with partial self-training

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 539–542, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics bwbaugh : Hierarchical sentiment analysis with partial self-training Wesley Baugh Department of Computer Science University of North Texas brianbaugh@my.unt.edu Abstract Using labeled Twitter training data from SemEval-2013, we train both a subjectivity classifier and a polarity classifier separately, and then combine the two into a single hier- archical classifier. Using additional unlabeled data that is believed to contain sentiment, we allow the polarity classifier to continue learn- ing using self-training. The resulting system is capable of classifying a document as neutral, positive, or negative with an overall accuracy of 61.2% using our hierarchical Naive Bayes classifier.1 1 Introduction Many people use social networks, such as Twitter, to connect and communicate with others. Users of social networks often share their experiences, such as watching a recent movie or tv show, reading a book, or a newly tried product or service. In addi- tion, social networks provide an avenue for discus- sion of current events, such as politics. Many people and companies are often concerned with how others perceive their product—which is sometimes them- selves, as is the case for politicians—or their service. By understanding and reacting to what the consumer is thinking, they can attempt to maximize their good press as well as to help minimize the bad. It would therefore be useful to use the information users of social networks share to perform sentiment analysis in order to understand how people perceive targets of interest. 1 A working demo of the system will be available for a short time at: http://infertweet.bwbaugh.com In general, sentiment analysis often involves the use of machine learning, especially Naive Bayes, SVM, and MaxEnt classifiers [Jose]. Features gen- eral include n-grams and POS tags [Go et al., 2009; Pak and Paroubek, 2010; Jose], as well as senti- ment lexicons [Jose]. Go et al. [2009] achieved around 82.5% accuracy for positive-negative polar- ity detection, Jose achieved around 76% accuracy for subjective-objective classification, and Pak and Paroubek [2010] achieved around 70% accuracy for a combined subjectivity-polarity classifier. While determining whether a document known to be subjective is positive or negative (polarity detec- tion) is relatively easy, a currently more difficult task in sentiment analysis is identifying whether a docu- ment is subjective or objective (subjectivity analy- sis). Many approaches simply ignore the objective class [Go et al., 2009], which does not work for real world problems as there are a substantial amount of documents that are either partially or wholly objec- tive [Koppel and Schler, 2006]. Many previous methods focus on either subjectiv- ity analysis or polarity detection. Our method incor- porates both subtasks into a single overall system in order to perform sentiment analysis. 2 Background The sentiment analysis in Twitter task of SemEval- 2013 [Wilson et al., 2013] provides 9,864 la- beled tweets from Twitter to be used as a train- ing dataset. Each instance …

Content truncated to 500 words