SWATAC: A Sentiment Analyzer using One-Vs-Rest Logistic Regression

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 636–639, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics SWATAC: A Sentiment Analyzer using One-Vs-Rest Logistic Regression Yousef Alhessi and Richard Wicentowski Swarthmore College Swarthmore, PA 19081 USA {yalhess1, richardw}@cs.swarthmore.edu Abstract This paper describes SWATAC, a system built for SemEval-2015’s Task 10 Subtask B, namely the Message Polarity Classification Task. Given a tweet, the system classifies the sentiment as either positive, negative, or neu- tral. Several preprocessing tasks such as nega- tion detection, spell checking, and tokeniza- tion are performed to enhance lexical infor- mation. The features are then augmented with external sentiment lexicons. Classification is done with Logistic Regression using a one-vs- rest configuration. For the test runs, the sys- tem was trained using only the provided train- ing tweets. The classifier was successful, with an F1 score of 58.43 on the official 2015 test data, and an F1 score of 66.64 on the Twitter 2014 progress data. 1 Introduction Since 2006, Twitter has grown into a ubiquitous global social platform. Millions of users compose Twitter messages, which are known as “tweets”, to express their opinions and sentiments about the world around them. These tweets turn into valu- able resources for sentiment analysis, a field that fo- cuses on analyzing the attitude of speakers or writ- ers towards a certain topic. Working with this infor- mal text genre opens up a new realm of challenges in the natural language processing world. This pa- per describes a tweet sentiment classifier which has been applied to Subtask B of SemEval-2015 Task 10 (Rosenthal et al., 2015). The tweets generated by users contain Internet slang, unconventional punctu- ation and spelling, and typos, which require a differ- ent set of preprocessing tools than traditional genres like newswire text. After preprocessing the tweets, classifying them into categories of positive, negative, and neutral presents another challenge. Many sentiment appli- cations make use of lexicons to supply features to the system, populating a list of positive and neg- ative types. Some publicly available sources in- clude the MPQA Subjectivity Lexicon (Wilson et al., 2005), the Opinion Lexicon (Liu et al., 2005), and the Sentiment140 Lexicon (Mohammad et al., 2013). While some of these lexicons do not target tweets as their analysis subject, they each provide a mapping from n-grams to sentiment labels, which proves to be helpful in building our tweet sentiment analyzer. After preprocessing, the system performs the clas- sification task. The classifier we use is a one- vs-rest logistic regression classifier, so the sys- tem uses three binary classifiers: positive/not- positive, negative/not-negative, and neutral/not- neutral. The classifier also over-samples the low- frequency classes, learning from the same number of examples of each class overall. The accompanying sections of the papers are or- ganized as follows: Section 2 describes resources such as the lexicons used in the system. It also out- lines the system design and the APIs that the system adopts. Section 3 describes the test runs and evalu- ates …

Content truncated to 500 words