UFRGS: Identifying Categories and Targets in Customer Reviews

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 725–729, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics UFRGS: Identifying Categories and Targets in Customer Reviews Anderson Kauer Institute of Informatics – UFRGS Porto Alegre – RS – Brazil aukauer@inf.ufrgs.br Viviane P. Moreira Institute of Informatics – UFRGS Porto Alegre – RS – Brazil viviane@inf.ufrgs.br Abstract This paper reports on our participation in SemEval-2015 Task 12, which was devoted to Aspect-Based Sentiment Analysis. Partic- ipants were required to identify the category (entity and attribute), the opinion target, and the polarity of customer reviews. The system we built relies on classification algorithms to identify aspect categories and on a set of rules to identify the opinion target. We propose a two-phase classification approach for category identification and use a simple method for po- larity detection. Our results outperform the baseline in many cases, which means our sys- tem could be used as an alternative for aspect classification. 1 Introduction Aspect Based Sentiment Analysis aims at discover- ing the opinions or sentiments expressed by a user on the different aspects of a given entity (Hu and Liu, 2004; Liu, 2012). Recently, a number of meth- ods and techniques have been developed to tackle this task and some of them rely on syntactic depen- dencies to locate the opinion target (Kim and Hovy, 2004; Qiu et al., 2011; Liu et al., 2013). A syntac- tic parser takes a natural language sentence as input and outputs the relationships between the words in the sentence. Figure 1 shows the dependency tree for the sentence “The phone has a good screen.” and the grammatical relations of each token (det, subj, mod, obj). We explore using grammatical relations to help identify the opinion targets. In this paper, we describe a system which took part on SemEval-2015, and the way it was applied can’t write enough good things about this camera,” “things” is extracted as an aspect because it is modified by the opinion word “good.” However, “things” is very unlikely to be a product aspect and thus should be pruned. We propose to use WordNet [28] to automatically generate a list of general words using three typical general words “thing,” “person,” and “place” as seeds. By extending the DP method with the knowledge that a general word is normally not an aspect, we obtain a major improvement in the precision with almost no drop in recall on a widely used benchmark data set. In summary, we make two contributions: (1) We propose to employ Answer Set Programming (ASP) – a variant of Logic Programming – to implement syntactical approach based aspect extraction. Our implementation of the DP method is more elegant and efficient, and it has only 8 rules, while a Java implementation has about 510 lines of code. The ASP based implementation can process about 3000 sentences per second, while the Java implementation only processes about 300 sentences per second. The preciseness and simplicity of the logic programming rules enable the …

Content truncated to 500 words