Kea: Sentiment Analysis of Phrases Within Short Texts

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 380–384, Dublin, Ireland, August 23-24, 2014. Kea: Sentiment Analysis of Phrases Within Short Texts Ameeta Agrawal, Aijun An Department of Computer Science and Engineering York University, Toronto, Canada M3J 1P3 {ameeta, aan}@cse.yorku.ca Abstract Sentiment Analysis has become an in- creasingly important research topic. This paper describes our approach to building a system for the Sentiment Analysis in Twit- ter task of the SemEval-2014 evaluation. The goal is to classify a phrase within a short piece of text as positive, negative or neutral. In the evaluation, classifiers trained on Twitter data are tested on data from other domains such as SMS, blogs as well as sarcasm. The results indicate that apart from sarcasm, classifiers built for sentiment analysis of phrases from tweets can be generalized to other short text do- mains quite effectively. However, in cross- domain experiments, SMS data is found to generalize even better than Twitter data. 1 Introduction In recent years, new forms of communication such as microblogging and text messaging have become quite popular. While there is no limit to the range of information conveyed by tweets and short texts, people often use these messages to share their sen- timents. Working with these informal text gen- res presents challenges for natural language pro- cessing beyond those typically encountered when working with more traditional text genres. Tweets and short texts are shorter, the language is very informal, with creative spelling and punctuation, misspellings, slang, new words, URLs, and genre- specific terminology such as, RT for “re-tweet” and #hashtags for tagging (Rosenthal et al., 2014). Although several systems have tackled the task of analyzing sentiment from entire tweets, the task of analyzing sentiments of phrases (a word This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ or more) within a tweet has remained largely un- explored. This paper describes the details of our system that participated in the subtask A of Semeval-2014 Task 9: Sentiment Analysis in Twitter (Rosenthal et al., 2014). The goal of this task is to determine whether a phrase within a mes- sage is positive, negative or neutral in that context. Here, a message indicates any short informal piece of text such as a tweet, SMS data, or a sentence from Live Journal blog, which is a social network- ing service where Internet users keep an online di- ary. A phrase could be a word or a few consecutive words within a message. The novelty of this task lies in the fact that a model built using only Twitter data is used to clas- sify instances from other short text domains such as SMS and Live Journal. Moreover, a short test corpus of sarcastic tweets is also used to test the performance of the sentiment classifier. The main contributions of this paper include a) developing a sentiment analysis classifer for phrases; b) training on Twitter …

Content truncated to 500 words