CIS-positive: A Combination of Convolutional Neural Networks and Support Vector Machines for Sentiment Analysis in Twitter

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 527–532, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics CIS-positive: Combining Convolutional Neural Networks and SVMs for Sentiment Analysis in Twitter Sebastian Ebert and Ngoc Thang Vu and Hinrich Schütze Center for Information and Language Processing University of Munich, Germany {ebert|thangvu}@cis.lmu.de, inquiries@cislmu.org Abstract This paper describes our automatic sentiment analysis system – CIS-positive – for SemEval 2015 Task 10 “Sentiment Analysis in Twit- ter”, subtask B “Message Polarity Classifica- tion”. In this system, we propose to normalize the Twitter data in a way that maximizes the coverage of sentiment lexicons and minimizes distracting elements. Furthermore, we inte- grate the output of Convolutional Neural Net- works into Support Vector Machines for the polarity classification. Our system achieves a macro F1 score of the positive and negative class of 59.57 on the SemEval 2015 test data. 1 Introduction On the Internet, text containing different forms of sentiment appears everywhere. Mining this informa- tion supports many types of interest groups. Com- panies, for instance, are interested in user feedback about the advantages and drawbacks of their prod- ucts. Users want to read short reviews or ratings of hotels they want to book for their next vaca- tion. Politicians try to predict the outcome of the next presidential election. An automatic sentiment analysis system can support all these different re- quirements. One source of these types of informa- tion covering many domains and topics is the social networking service Twitter. Its popularity and the users’ productivity in creating new text makes it an interesting research topic. However, Twitter intro- duces specific challenges as we will see next. In general, automatic sentiment analysis is chal- lenging due to many different factors, such as am- biguous word senses, context dependency, sarcasm, etc. Specific properties of Twitter text make this task even more challenging. The limit of 140 character per message leads to countless acronyms and ab- breviations. Moreover, the vast majority of tweets is of informal character and contains intentional miss-spellings and wrong use of grammar. Hence, the out-of-vocabulary (OOV) rate of Twitter text is rather high, which leads to information loss. One of the SemEval 2015 shared tasks – Task 10: Sentiment Analysis in Twitter – addresses these challenges (Rosenthal et al., 2015). We participated in Subtask B the “Message Polarity Classification” task. The goal is to predict the polarity of a given tweet into positive, negative, or neutral. The task organizers provided tweet IDs and corresponding la- bels to have a common ground for training polarity classification systems. More information about the task, its other subtasks as well as information about how the data was selected can be found in (Rosen- thal et al., 2015). In this paper, we present our sentiment analysis system for SemEval 2015 - Task 10. Our system addresses the above mentioned challenges in two ways. First, we normalize the text to maximize the coverage of sentiment lexicons and minimize dis- tracting elements such as user …

Content truncated to 500 words