OPTWIMA: Comparing Knowledge-rich and Knowledge-poor Approaches for Sentiment Analysis in Short Informal Texts

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 460–465, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics OPTWIMA: Comparing Knowledge-rich and Knowledge-poor Approaches for Sentiment Analysis in Short Informal Texts Alexandra Balahur European Commission Joint Research Centre Via E. Fermi 2749 21027 Ispra (VA), Italy {alexandra.balahur}@jrc.ec.europa.eu Abstract The fast development of Social Media made it possible for people to no loger remain mere spectators to the events that happen in the world, but become part of them, comment- ing on their developments and the entities in- volved, sharing their opinions and distribut- ing related content. This phenomenon is of high importance to news monitoring systems, whose aim is to obtain an informative snap- shot of media events and related comments. This paper presents the strategies employed in the OPTWIMA participation to SemEval 2013 Task 2-Sentiment Analysis in Twitter. The main goal was to evaluate the best settings for a sentiment analysis component to be added to the online news monitoring system. We describe the approaches used in the com- petition and the additional experiments per- formed combining different datasets for train- ing, using or not slang replacement and gener- alizing sentiment-bearing terms by replacing them with unique labels. The results regarding tweet classification are promising and show that sentiment generaliza- tion can be an effective approach for tweets and that SMS language is difficult to tackle, even when specific normalization resources are employed. 1 Introduction Sentiment analysis is the Natural Language Process- ing (NLP) task dealing with the detection and clas- sification of sentiments in texts. Usually, the classes considered are “positive”, “negative” and “neutral”, although in some cases finer-grained categories are added (e.g. “very positive” and “very negative”) or only the “positive” and “negative” classes are taken into account. This task has received a lot of interest from the re- search community in the past years. The work done regarded the manner in which sentiment can be clas- sified from texts pertaining to different genres and distinct languages, in the context of various applica- tions, using knowledge-based, semi-supervised and supervised methods [Pang and Lee, 2008]. The re- sult of the analyses performed have shown that the different types of text require specialized methods for sentiment analysis, as, for example, sentiments are not conveyed in the same manner in newspaper articles and in blogs, reviews, forums or other types of user-generated contents [Balahur et al., 2010]. In the light of these findings, dealing with senti- ment analysis in tweets and SMS (that we can gener- ally call “short informal texts”) requires an analysis of the characteristics of such texts and the design of adapted methods. Our participation in the SemEval 2013 Task 2 [Wilson et al., 2013] had as objective to test how well our proposed methods for sentiment analysis for short informal texts (especially tweets) would perform. The two subtasks proposed in this com- petition were: a) the classification of sentiment from snippets from tweets …

Content truncated to 500 words