NLANGP: Supervised Machine Learning System for Aspect Category Classification and Opinion Target Extraction

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 496–501, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics NLANGP: Supervised Machine Learning System for Aspect Category Classification and Opinion Target Extraction Zhiqiang Toh Institute for Infocomm Research 1 Fusionopolis Way Singapore 138632 ztoh@i2r.a-star.edu.sg Jian Su Institute for Infocomm Research 1 Fusionopolis Way Singapore 138632 sujian@i2r.a-star.edu.sg Abstract This paper describes our system used in the Aspect Based Sentiment Analysis Task 12 of SemEval-2015. Our system is based on two supervised machine learning algorithms: sig- moidal feedforward network to train binary classifiers for aspect category classification (Slot 1), and Conditional Random Fields to train classifiers for opinion target extraction (Slot 2). We extract a variety of lexicon and syntactic features, as well as cluster features induced from unlabeled data. Our system achieves state-of-the-art performances, rank- ing 1st for three of the evaluations (Slot 1 for both restaurant and laptop domains, and Slot 1 & 2) and 2nd for Slot 2 evaluation. 1 Introduction The amount of user-generated content on the web has grown rapidly in recent years, prompting in- creasing interests in the research area of sentiment analysis and opinion mining. Most previous work is concerned with detecting the overall polarity of a sentence or paragraph, regardless of the target enti- ties (e.g. restaurants) and their aspects (e.g. food). By contrast, the Aspect Based Sentiment Analysis task of SemEval 2014 (SE-ABSA14) is concerned with identifying the aspects of given target enti- ties and the sentiment expressed towards each as- pect (Pontiki et al., 2014). The SemEval-2015 Aspect Based Sentiment Analysis (SE-ABSA15) task is a continuation of SE-ABSA14 (Pontiki et al., 2015). The SE- ABSA15 task features a number of changes that address issues raised in SE-ABSA14 and also en- courage further in-depth research. For example, (1) instead of isolated (potentially out of context) sen- tences, the input datasets will contain entire reviews; (2) information linking aspect terms and aspect cat- egories are now provided; (3) besides in-domain ABSA (Subtask 1), SE-ABSA15 will include an out-of-domain ABSA subtask (Subtask 2). We participate in Subtask 1 of SE-ABSA15, namely aspect category classification (Slot 1) and opinion target extraction (Slot 2). We also partici- pate in the evaluation which assesses whether a sys- tem identifies both the aspect categories and opinion targets correctly (Slot 1 & 2). For Slot 1, we model the problem as a multi-class classification problem where binary classifiers are trained to predict the aspect categories. We follow the one-vs-all strategy and train a binary classifier for each category in the training set. Each classifier is trained using sigmoidal feedforward network with 1 hidden layer. For Slot 2, we follow the approach of Toh and Wang (2014) by modeling the problem as a sequential labeling task, using Conditional Random Fields (CRF) as the training algorithm. For Slot 1 & 2, we perform a simple combination of Slot 1 pre- dictions and Slot 2 predictions. The remainder of this paper is structured as fol- lows. In Section 2, …

Content truncated to 500 words