UFAL at SemEval-2016 Task 5: Recurrent Neural Networks for Sentence Classification

Proceedings of SemEval-2016, pages 367–371, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UFAL at SemEval-2016 Task 5: Recurrent Neural Networks for Sentence Classification Aleˇs Tamchyna and Kateˇrina Veselovsk´a Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, Prague, Czech Republic {tamchyna,veselovska}@ufal.mff.cuni.cz Abstract This paper describes our system for aspect- based sentiment analysis (ABSA). We partic- ipate in Subtask 1 (sentence-level ABSA), fo- cusing specifically on aspect category detec- tion. We train a binary classifier for each category. This year’s addition of multiple languages makes language-independent ap- proaches attractive. We propose to utilize neural networks which should be capable of discovering linguistic patterns in the data au- tomatically, thereby reducing the need for language-specific tools and feature engineer- ing. 1 Introduction Aspect-based sentiment analysis (ABSA) refers to the identification of specific entities and their as- pects (aspect terms, opinion targets) in text and to the classification of their polarity. Typically, ABSA is applied to user reviews from various fields, such as consumer electronics, hotels or restaurants. In ABSA, we assume that the general target of evalua- tion has several aspects (e.g. food quality for restau- rants) and we attempt to identify users’ opinions on these individual aspects. Unlike the more general task of sentiment analysis where the goal would be to classify the polarity of entire sentences (or even whole reviews), in ABSA, we need to take a more fine-grained approach and consider also the internal structure of the given sentences. As for the sentence-level ABSA, the aim is to identify all opinion tuples present in the sentence, taking into account the context of the whole review. In our work, we focus on identifying aspect term cat- egories; these are composed of entities (e.g. FOOD) and their attribute labels (e.g. QUALITY or PRICE). Both the entities and the attribute labels were as- signed based on predefined inventories. We apply our method on several languages cover- ing the following domains: • Arabic: hotels • Dutch: restaurants • English: consumer electronics and restaurants • French: restaurants • Russian: restaurants • Spanish: restaurants • Turkish: restaurants Our submission was the best system for Russian and Turkish but did not achieve noteworthy results in other domains/languages. 2 Related Work The previous results for different ABSA SemEval tasks are discussed in Pontiki et al. (2014) and Pon- tiki et al. (2015). So far, most researchers in the field have focused on traditional machine learning approaches, such as various probabilistic methods (see Agarwal and Mittal (2016)), and/or employed deterministic methods, e.g. subjectivity lexicons (Taboada et al., 2011). Because more languages were added this year, approaches such as neural net- works, which require less language-specific data and engineering, become attractive. 367 Neural networks have been used for sentiment analysis. Particularly, in last year’s SemEval Twit- ter sentiment classification task, several submissions applied convolutional networks (Toh and Su, 2015; Ebert et al., 2015). In our work, we use recurrent networks instead. Our motivation for this decision …

Content truncated to 500 words