SWAT-CMW: Classification of Twitter Emotional Polarity using a Multiple-Classifier Decision Schema and Enhanced Emotion Tagging

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 669–672, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics SWAT-CMW: Classification of Twitter Emotional Polarity using a Multiple-Classifier Decision Schema and Enhanced Emotion Tagging Riley Collins, Daniel May, Noah Weinthal, and Richard Wicentowski Swarthmore College Swarthmore, PA 19081 {rcollin4, dmay1, nweinth1, richardw}@cs.swarthmore.edu Abstract In this paper, we describe our approach to Se- meval 2015 task 10 subtask B, message level sentiment detection. Our system implements a variety of classifiers and data preparation tech- niques from previous work. The set of features and classifiers used in the final system pro- duced consistently strong results using cross- validation on the provided training data. Our final system achieved an F-score of 57.60 on the provided test data. The overall best per- forming system had an F-score of 64.84. 1 Introduction With the unprecedented growth of social media in the past decade, more individuals than ever before have a means to share their opinions and broad- cast their voice. As the number of readily available opinions grows, a challenge of academic and com- mercial importance emerges. Namely, if the senti- ment of social media communications can be reli- ably determined algorithmically, a deeply informa- tive dataset can be developed. Such data can be used in a variety of applications, from predicting elec- tion results to seeing how well a new product is re- ceived. However, this task is greatly complicated by inconsistencies in spelling, grammar, lexicon, and other linguistic phenomena found in online commu- nications. The SemEval 2015 Task 10 Subtask B (Rosenthal et al., 2015) challenges participants to determine the sentiment polarity of posts on the so- cial media site Twitter. Specifically, the task is to decide whether the sentiment of a given tweet is pos- itive, negative, or neutral. In this paper we present an approach to this task which synthesizes a number of different preprocessing techniques and classifica- tion methods, which we use to classify the tweets. Our approach was inspired by several approaches to previous iterations of this task. The winning team in 2014, TeamX, used several preprocessors includ- ing text normalization, lexical sense mapping, clus- tering, and word sense disambiguation to train a ma- chine learner to determine emotion (Miura et al., 2014). Ultimately, we hypothesized that a success- ful approach relies not only on the choice of a good classifier, but also in large part upon the prepara- tion of data for that classifier. This hypothesis led us to place high value on our prepocessing, and as such we focused our energy on implementing strate- gies that would lead to improvements within exist- ing classifiers, a decision which ultimately led to the creation of our decision schema. 2 System Description 2.1 Preprocessing Our system makes use of various preprocessing steps in order to reduce the dimensionality of the data set and improve overall performance. These steps included: • Tokenization using Twokenizer (Gimpel et al., 2011), a tokenizer designed specifically for Tweets. • Case-folding so …

Content truncated to 500 words