Gradiant-Analytics: Training Polarity Shifters with CRFs for Message Level Polarity Detection

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 539–544, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics Gradiant-Analytics: Training Polarity Shifters with CRFs for Message Level Polarity Detection H´ector Cerezo-Costas, Diego Celix-Salgado Gradiant - Galician Research and Development Center in Advanced Telecommunications Edificio CITEXVI, local 14 Vigo, Pontevedra 36310, SPA {hcerezo, dcelix}@gradiant.org Abstract In this paper we present our solution for ob- taining sentiment at message-level of short sentences. The system combines the use of polarity dictionaries and Conditional Random Fields to obtain syntactic and semantic fea- tures, which are afterwards fed to a statistical classifier in order to obtain the sentence polar- ity. To improve results, an ensemble of clas- sifiers was employed by combining the indi- vidual outputs with majority voting strategy. Our solution was evaluated in the SemEval 2015 Task 10, subtask B: Sentiment Analysis in Twitter, achieving competitive performance in all testsets. 1 Introduction Sentiment Analysis (SA) is a hot-topic in the aca- demic world, and also in the industry. In SA, a la- bel is automatically assigned to a piece of content carrying the polarity of the composition. The rel- evance for the web industry is clear, as new ser- vices promote content sharing among users. The number of registers generated by these services is paramount, discouraging manual analysis. Hence, automatic systems capable of processing this infor- mation have great value for the industry. Many ser- vices, such as web advertisement, recommendation, and mail campaigns (to name a few) could benefit from the information gathered with polarity analysis of user content. This work is focused on message-level sentiment analysis, that is, the objective is the assignment of polarity to a small piece a text, typically one or two sentences with less than 140 characters. This restric- tion is motivated by the popularity of microblogging services such as Twitter. Here, users write messages of up to 140 characters to share information, their opinions or their feelings with other users. Those messages are shared in real time, and are a sample of the public opinion. Therefore, these small compo- sitions published in microblogging sites can be ana- lyzed to deduce the opinions about any topic of in- terest. Nevertheless, automatic systems are not perfect. The results of the sentiment analysis in short sen- tences is not completely reliable. State of the art solutions are still far from being comparable to hu- man performance, though very promising results were obtained recently using deep learning systems (Socher et al., 2013; Tang et al., 2014), and a careful selection of features with Support Vector Machines (SVM) (Zhu et al., 2014) or other statistical classi- fiers (Go et al., 2009). This paper describes our sentiment classifier for short sentences and the results in our participation in the SemEval 2015 competition. We have imple- mented a supervised solution for learning the polar- ity of short messages. We made extensive use of se- quential Conditional Random Fields (CRFs) in order to obtain the scope of polarity modifiers …

Content truncated to 500 words