UWB at SemEval-2016 Task 7: Novel Method for Automatic Sentiment Intensity Determination

Proceedings of SemEval-2016, pages 481–485, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UWB at SemEval-2016 Task 7: Novel Method for Automatic Sentiment Intensity Determination Ladislav Lenc, Pavel Kr´al and V´aclav Rajtmajer Dept. of Computer Science & Engineering Faculty of Applied Sciences University of West Bohemia Plzeˇn, Czech Republic {llenc,pkral,rajtmajv}@kiv.zcu.cz Abstract We present a novel method for determining sentiment intensity. The main goal is to as- sign a phrase a score from 0 to 1 which indi- cates the strength of its association with pos- itive sentiment. The proposed model uses a rich set of features with Gaussian pro- cesses regression model that computes the fi- nal score. The system was evaluated on the data from 7th task of SemEval 2016. Our regression model trained on the development data reached Kendall rank correlation of 0.659 on general English phrases and 0.414 on En- glish Twitter test data. 1 Introduction A great part of today’s communication takes place on the Internet. Many companies make their busi- ness on the Web, it is possible to read newspa- pers through this media, etc. All these topics are tightly connected with forums, reviews and com- ments where users express their opinions and feel- ings. This great amount of short messages is a very rich source of information. For example companies can survey how people appreciate their goods or ser- vices. Social media such as Twitter or Facebook can also provide large amount of data for opinion and sentiment analysis. Sentiment analysis can be seen as a part of opin- ion mining that gives the affective part of an opin- ion (Kim and Hovy, 2004). The short texts are usu- ally assigned either positive or negative sentiment (i.e. sentiment polarity). It can be done for exam- ple by text categorization techniques. The minimum cuts are used for this task in (Pang and Lee, 2004). Another sentiment analysis approach based on La- tent Dirichlet Allocation (LDA) is proposed in (Li et al., 2010). Many researchers evaluate their approaches on movie or product reviews. Lately, with the boom of social networks, there are also applications that analyze sentiment on Twitter (Wang et al., 2011). An important progress was made thanks to the Se- mEval workshop (Rosenthal et al., 2015) in a senti- ment analysis task. The task of determining senti- ment intensity was introduced as a part of SemEval 2015 task 10. The goal of this task is to associate a word (or a short phrase) with a score that indi- cates its sentiment level. A score of 1 means that the phrase is fully positive whereas 0 is assigned to negative ones. This approach allows much more fine-grained sentiment evaluation. This year, senti- ment intensity determination is the objective of the task 7 (Kiritchenko et al., 2016) and aims at three separate datasets: General English phrases, English Twitter and Arabic Twitter. An interesting method that uses no linguistic re- sources and is based on word embeddings was pro- posed in (Astudillo et al., …

Content truncated to 500 words