ELiRF-UPV at SemEval-2017 Task 4: Sentiment Analysis using Deep Learning

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 723–727, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics ELiRF-UPV at SemEval-2017 Task 4: Sentiment Analysis using Deep Learning Jos´e- ´Angel Gonz´alez, Ferran Pla, Llu´ıs-F. Hurtado Universitat Polit`ecnica de Val`encia Cam´ı de Vera sn, 46022, Val`encia {jogonba2|fpla|lhurtado}@dsic.upv.es Abstract This paper describes the participation of ELiRF-UPV team at task 4 of Se- mEval2017. Our approach is based on the use of convolutional and recurrent neural networks and the combination of general and specific word embeddings with polar- ity lexicons. We participated in all of the proposed subtasks both for English and Arabic languages using the same system with small variations. 1 Introduction Twitter has become a source of a huge amount of information which introduces great possibili- ties of research in the field of Sentiment Anal- ysis. Sentiment Analysis or Opinion Mining, is a research area within Natural Language Pro- cessing whose aim is to identify the underlying emotion of a certain document, sentence or as- pect (Liu, 2012). Sentiment Analysis systems has been applied, among other, for classifying reviews (Turney, 2002; Pang et al., 2002), for generating aspect-based summaries (Hu and Liu, 2004), or political tendency identification (Pla and Hurtado, 2014). SemEval-2017 task 4 organizers proposed five different subtasks. All five subtasks are related to sentiment analysis at global level in Twitter, but each one of them has significant differences. Ad- ditionally, in the 2017 edition, the five subtasks were also proposed in Arabic. Altogether, the par- ticipants could address ten different challenges. Subtask A consists in predicting the message polarity as positive, negative, or neutral. In sub- tasks B and C, given a message and a topic systems should assign the message in a two-point scale or in a five-point scale respectively. Subtasks D and E address the problem of tweet quantification, that is, given a set of tweets about a given topic, estimate the distribution of the tweets across two-point scale or in a five-point scale re- spectively. The rest of this paper is organized as follows. Section 2 describes the general system architec- ture proposed in this work. Section 3 presents both the variations on the general system introduced to address the different subtasks and the results ob- tained in the subtasks. Finally, section 4 presents some conclusions and the future work. 2 System description In this section, we describe the system architec- ture we used for all the Sentiment Analysis sub- tasks. This system is based on the use of convolu- tional and recurrent neural networks and the com- bination of general and specific word embeddings (Mikolov et al., 2013b,a) with polarity lexicons. Slight modifications of the system have been applied to adapt it to each subtask. These mod- ifications are motivated by the characteristics of each subtask and the available resources. The system combines three Convolutional Re- current Neural Network (CRNN) (Zhou et al., 2002) in order to learn high level abstractions (Le- cun et al., 2015) from noisy representations …

Content truncated to 500 words