MayAnd at SemEval-2016 Task 5: Syntactic and word2vec-based approach to aspect-based polarity detection in Russian

Proceedings of SemEval-2016, pages 325–329, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics MayAnd at SemEval-2016 Task 5: Syntactic and word2vec-based Approach to Aspect-based Polarity Detection in Russian Vladimir Mayorov and Ivan Andrianov Institute for System Programming of the Russian Academy of Sciences 25 Alexander Solzhenitsyn Street Moscow, Russia vmayorov@ispras.ru, ivan.andrianov@ispras.ru Abstract This paper describes aspect-based polarity de- tection system for Russian, used in aspect- based sentiment analysis task (ABSA) of SemEval-2016 (Task 5, subtask 1, slot 3). The system consists of two independent classifiers: for opinion target expressions and for implicit opinion target mentions. We introduce a set of standard unigram features together with more sophisticated ones: based on sentence syntac- tic structure and based on lemmas vector rep- resentation. Being applied to Russian restaurant reviews, our system achieved best quality among four participants. 1 Introduction Aspect-based sentiment analysis task (ABSA) drew much attention with the growth of Internet retail sites and facilities for customers to make a free text product review. ABSA systems allow businesses to track reputation of their products even more granu- larly than before: per aspect. At the same time, cus- tomers can be provided with per-aspect sentiment summaries of reviews about concurrent products to simplify their choice. SemEval-2016 Task 5 (ABSA) organizers (Pon- tiki et al., 2016) proposed two subtasks: Sentence- level ABSA and Text-level ABSA. First subtask was split to three slots: Aspect Category Detec- tion, Opinion Target Expression, Sentiment Polar- ity. Training data for the task were available for var- ious languages (English, Russian, etc.) and domains (restaurants, laptops, etc.). Slot 3 can be described as follows: given a set of reviews labeled with OTEs (opinion target ex- pressions: start-end-entity-attribute quadruples) de- termine a polarity label for each OTE: positive, neg- ative, neutral (mildly positive or mildly negative sen- timent). Additionally polarity labels should be com- puted for entity-attribute pairs assigned to reviews sentences. This paper describes our team participation in slot 3 evaluation for Russian language Restaurant do- main. The rest of the article is structured as fol- lows: section 2 mentions some recent approaches to aspect-based polarity detection task, section 3 con- tains our system description, section 4 provides our system evaluation results. 2 Related work According to (Andrianov et al., 2015), aspect-based polarity detection methods can be divided to three categories: classifier-based (Zhang and Lan, 2015; Hamdan et al., 2015; Blinov and Kotelnikov, 2015), neural network-based (Tarasov, 2015) and unsuper- vised (Garc´ıa Pablos et al., 2015). Most promis- ing approaches according to recent evaluations re- sults (Pontiki et al., 2015; Loukachevitch et al., 2015) are classifier-based methods, especially semi- supervised ones which bring “external knowledge” (word2vec (Mikolov et al., 2013), Brown clus- ters (Brown et al., 1992), polarity lexicons (Bac- cianella et al., 2010; Wilson et al., 2005)) into the ta- ble. Semi-supervised techniques become extremely useful due to lack of labeled data for the aspect- based sentiment analysis task in general and polarity detection subtask in particular. 325 3 System Description Our aspect-based polarity detection …

Content truncated to 500 words