Utilizing Large Twitter Corpora to Create Sentiment Lexica

Utilizing Large Twitter Corpora to Create Sentiment Lexica Valerij Fredriksen, Brage Jahren, Björn Gambäck Department of Computer Science Norwegian University of Science and Technology Trondheim, Norway valerij92@gmail.com, brage.jahren@gmail.com, gamback@ntnu.no Abstract The paper describes an automatic Twitter sentiment lexicon creator and a lexicon-based sentiment analysis system. The lexicon creator is based on a Pointwise Mutual Information approach, utilizing 6.25 million automatically labeled tweets and 103 million unlabeled, with the created lexicon consisting of about 3 000 entries. In a comparison experiment, this lexicon beat a manually annotated lexicon. A sentiment analysis system utilizing the created lexicon, and handling both negation and intensification, produces results almost on par with sophisticated machine learning-based systems, while significantly outperforming those in terms of run-time. Keywords: Pointwise Mutual Information, Sentiment lexica, Lexicon-based sentiment analysis 1 Introduction A popular social medium providing opinionated texts is the micro-blogging service Twitter. On Twitter, users can post textual entries of up to 140 characters, commonly called tweets. Each day, approximately 500 million new tweets are posted; a fraction of those are made available through Twitter’s public API. Large datasets can therefore easily be acquired, making Sentiment Analysis (SA) of tweets par- ticularly popular. The present study shows that two features in particular stand out when building Twitter sentiment classification systems: the effect of using sentiment lexica in the clas- sification process and the run-time performance. The use of sentiment lexica is shown to be the single most valu- able system component in terms of overall system perfor- mance. Without using sentiment lexica, the performance dropped as much as 4%. In general, most present-day Twit- ter specific lexica only contain word unigrams and bigrams. Here, in contrast, the effect of including longer phrases in sentiment lexica is explored, as well as utilizing the long phrases in the classification process, while striving to cre- ate a lexicon-based SA system with real time performance. Section 2 gives an overview of the state-of-art in automatic creation of sentiment lexica and lexicon-based SA, while Section 3 describes the datasets used. In Section 4, the over- all architecture of both the lexicon-based SA system and the lexicon creator system are detailed. Section 5 includes the tests conducted on the created lexicon and lexicon-based SA system, which are further discussed in Section 6. Fi- nally, Section 7 concludes and outlines future work. 2 Related Work Lexicon-based sentiment analysis is the task of performing SA solely based on sentiment lexica. The overall perfor- mance of a lexicon-based SA system is hence closely re- lated to the quality of the sentiment lexica, but also to how the sentiment lexicon information is used. In recent years, most SA systems use a Machine Learning (ML) approach, with sentiment lexica as a feature among others. The num- ber of new lexicon-based SA systems that have appeared is therefore few. Although the task of classification differs between a lexicon-based and an ML-based SA system, the sentiment lexica features in ML systems are grounded in the same ideas as lexicon-based SA systems. The common …

Content truncated to 500 words