BUTknot at SemEval-2016 Task 5: Supervised Machine Learning with Term Substitution Approach in Aspect Category Detection

Proceedings of SemEval-2016, pages 301–305, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics BUTknot at SemEval-2016 Task 5: Supervised Machine Learning with Term Substitution Approach in Aspect Category Detection Jakub Mach´aˇcek Brno University of Technology, Faculty of Information Technology, IT4Innovations Centre of Excellence Boˇzetˇechova 2, 612 66 Brno, Czech Republic qmachacek@stud.fit.vutbr.cz Abstract This paper describes an approach used to solve Aspect Category Detection (Subtask 1, Slot 1) of SemEval 2016 Task 5. The core of the pre- sented system is based on Supervised machine learning using bigram bag-of-words model. The performance is enhanced by several pre- processing methods, most importantly by a term substitution technique. The system has reached a very good performance in compari- son with other submitted systems. 1 Introduction As the Internet more and more becomes the means of expressing opinions about various subjects, the need to effectively process those opinions (subjec- tive information) is becoming more and more impor- tant. Many companies around the world are now in- terested in gathering public opinion and performing strategic moves accordingly. Thus, Sentiment Analy- sis (also known as Opinion mining) has become an important area of interest. Existing systems performing sentiment analysis usually only predict polarity of a given sentiment. While this can be sufficient in a lot of cases, some- times we wish to analyze opinions about different aspects of the same entity. This task is known as Aspect-based Sentiment Analysis. SemEval 2016 Task 5 (Pontiki et al., 2016) con- sists of three subtasks. The first subtask is about sentence-level sentiment analysis and is divided into three slots: Aspect Category Detection, Opinion Tar- get Expression and Sentiment Polarity Detection. There are multiple distinct domains in which par- ticipants were given the opportunity to test their sys- tems. Each domain was available for one of the fol- lowing languages: Arabic, Chinese, Dutch, English, French, Russian, Spanish and Turkish. For all domains, there is a limited, known in ad- vance, list of entities and aspects the system should recognize. Each entity can be associated with only certain aspect(s). The set of all possible entity- aspect pairs is also limited and known in advance. Each submitted system could run in two modes: Constrained (using no external data sources like lex- icons or additional training sets) and Unconstrained (no data source restriction). The system I propose is focused on Aspect Cat- egory Detection only, i.e. it only predicts which as- pects of a given entity a given sentiment has opin- ions about. I decided to participate in the English language, for which two domains were available: restaurant and laptop reviews. There were 12 entity- aspect pairs for the restaurants domain and over 80 for the laptops domain. 2 Approach My approach was inspired by the NLANGP sys- tem (Toh and Su, 2015) which achieved excellent results in SemEval 2015 (Pontiki et al., 2015) with a straightforward solution. I model the task as a multi-label classification with binary relevance transformation, where labels correspond to the entity-aspect pairs. All train and …

Content truncated to 500 words