SWASH: A Naive Bayes Classifier for Tweet Sentiment Identification

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 626–630, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics SWASH: A Naive Bayes Classifier for Tweet Sentiment Identification Ruth Talbot1, Chloe Acheampong2 and Richard Wicentowski1 1Swarthmore College, Swarthmore, PA USA 2Ashesi University, Accra, Ghana {rtalbot1, richardw}@cs.swarthmore.edu chloeacheampong@gmail.com Abstract This paper describes a sentiment classifica- tion system designed for SemEval-2015, Task 10, Subtask B. The system employs a con- strained, supervised text categorization ap- proach. Firstly, since thorough preprocess- ing of tweet data was shown to be effective in previous SemEval sentiment classification tasks, various preprocessessing steps were in- troduced to enhance the quality of lexical in- formation. Secondly, a Naive Bayes classi- fier is used to detect tweet sentiment. The classifier is trained only on the training data provided by the task organizers. The system makes use of external human-generated lists of positive and negative words at several steps throughout classification. The system pro- duced an overall F-score of 59.26 on the of- ficial test set. 1 Introduction Over the past few years, an increasing number of people have begun to express their opinion through social networks and microblogging services. Twit- ter, as one of the most popular of these social net- works, has become a major platform for social com- munication, allowing its users to send and read short messages called ‘tweets’. Tweets have become im- portant in a variety of tasks, including the predic- tion of election results (O’Connor et al., 2010). The emergence of online expressions of opinion has at- tracted interest in sentiment analysis of tweets in both academia and industry. Sentiment analysis, also known as opinion mining, focuses on computa- tional treatments of sentiments (emotions, attitudes, opinions) in natural language text. In this paper we describe our submission to Task 10, subtask B: Mes- sage Polarity Classification. The task is defined as: ‘Given a message, classify whether the message is of positive, negative, or neutral sentiment. For a mes- sage conveying both a positive and negative senti- ment, whichever is the stronger sentiment should be chosen’ (Rosenthal et al., 2015). This paper describes a system which utilizes a Naive Bayes classifier to determine the sentiment of tweets. This paper describes the resources used, the system details, including preprocessing steps taken, feature extraction and classifier implemented, and the test runs and end results. 2 Resources 2.1 Labeled Tweets This system is constrained, and the only training data used is the sentiment labeled training data pro- vided by the task organizers. The training data we used includes 8142 tweets, each labeled as positive, negative or neutral. 2.2 Sentiment Lexicon Our system relies on an external lexicon of approx- imately 6800 tokens labeled as either positive or negative (Liu et al., 2005). The lexicon consists of words that humans have tagged as having either strongly negative or strongly positive sentiment. If a word in a tweet is preidentified as highly positive or negative, we add a special feature to the tweet’s fea- tures to …

Content truncated to 500 words