TwitterHawk: A Feature Bucket Based Approach to Sentiment Analysis

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 640–646, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics TwitterHawk: A Feature Bucket Approach to Sentiment Analysis William Boag, Peter Potash, Anna Rumshisky Dept. of Computer Science University of Massachusetts Lowell 198 Riverside St, Lowell, MA 01854, USA {wboag,ppotash,arum}@cs.uml.edu Abstract This paper describes TwitterHawk, a system for sentiment analysis of tweets which partici- pated in the SemEval-2015 Task 10, Subtasks A through D. The system performed com- petitively, most notably placing 1st in topic- based sentiment classification (Subtask C) and ranking 4th out of 40 in identifying the sen- timent of sarcastic tweets. Our submissions in all four subtasks used a supervised learning approach to perform three-way classification to assign positive, negative, or neutral labels. Our system development efforts focused on text pre-processing and feature engineering, with a particular focus on handling negation, integrating sentiment lexicons, parsing hash- tags, and handling expressive word modifica- tions and emoticons. Two separate classifiers were developed for phrase-level and tweet- level sentiment classification. Our success in aforementioned tasks came in part from lever- aging the Subtask B data and building a single tweet-level classifier for Subtasks B, C and D. 1 Introduction In recent years, microblogging has developed into a resource for quickly and easily gathering data about how people feel about different topics. Sites such as Twitter allow for real-time communication of senti- ment, thus providing unprecedented insight into how well-received products, events, and people are in the public’s eye. But working with this new genre is challenging. Twitter imposes a 140-character limit on messages, which causes users to use novel abbre- viations and often disregard standard sentence struc- tures. For the past three years, the International Work- shop on Semantic Evaluation (SemEval) has been hosting a task dedicated to sentiment analysis of Twitter data. This year, our team participated in four subtasks of the challenge: Contextual Polarity Disambiguation (phrase-level), B: Message Polarity Classification (tweet-level), C: Topic-Based Mes- sage Polarity Classification (topic-based), and D: Detecting Trends Towards a Topic (trending senti- ment). For a more thorough description of the tasks, see Rosenthal et al. (2015). Our system placed 1st out of 7 submissions for topic-based sentiment pre- diction (Subtask C), 3rd out of 6 submissions for de- tecting trends toward a topic (Subtask D), 10th out of 40 submissions for tweet-level sentiment predic- tion (Subtask B), and 5th out of 11 for phrase-level prediction (Subtask A). Our system also ranked 4th out of 40 submissions in identifying the sentiment of sarcastic tweets. Most systems that participated in this task over the past two years have relied on basic machine learn- ing classifiers with a strong focus on developing ro- bust and comprehensive feature set. The top sys- tem for Subtask A in both 2013 and 2014 from NRC Canada (Mohammad et al., 2013; Zhu et al., 2014) used a simple linear SVM while putting great ef- fort into creating and incorporating sentiment lexi- cons as well as carefully handling …

Content truncated to 500 words