NLANGP at SemEval-2016 Task 5: Improving Aspect Based Sentiment Analysis using Neural Network Features

Proceedings of SemEval-2016, pages 282–288, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics NLANGP at SemEval-2016 Task 5: Improving Aspect Based Sentiment Analysis using Neural Network Features 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 submitted to Aspect Based Sentiment Analysis Task 5 of SemEval-2016. Our system consists of two components: binary classifiers trained using single layer feedforward network for aspect category classification (Slot 1), and sequential labeling classifiers for opinion target extrac- tion (Slot 2). Besides extracting a variety of lexicon features, syntactic features, and clus- ter features, we explore the use of deep learn- ing systems to provide additional neural net- work features. Our system achieves the best performances on the English datasets, ranking 1st for four evaluations (Slot 1 for both restau- rant and laptop domains, Slot 2, and Slot 1 & 2). 1 Introduction Sentiment analysis and opinion mining have gained increasing interests in recent years due to the contin- uous growing of user-generated content on the In- ternet. Traditionally, the primary focus of the re- search has been on the detection of the overall senti- ment of a sentence or paragraph. However, such ap- proach is unable to handle conflicting sentiment for different aspects of the same entity. Hence, a more fine-grained approach, known as Aspect-Based Sen- timent Analysis (ABSA), is proposed. The goal is to correctly identify the aspects of entities and the polarity expressed for each aspect. The SemEval-2016 Aspect Based Sentiment Analysis (SE-ABSA16) task is a continuation of the same task in 2015 (Pontiki et al., 2015). Be- sides sentence-level ABSA (Subtask 1), it provides datasets to allow participants to work on text-level ABSA (Subtask 2). In addition, additional datasets in languages other than English are available (Pon- tiki et al., 2016). We participate in Subtask 1 of SE-ABSA16, where we submitted results for Slot 1 (aspect cat- egory classification), Slot 2 (opinion target extrac- tion), and Slot 1 & 2 (assessing whether a system correctly identifies both Slot 1 and Slot 2) for the English datasets. Our work is based on our previous machine learn- ing system described in Toh and Su (2015), en- hanced using additional features learned from neu- ral networks. For Slot 1, we treat the problem as a multi-class classification problem where aspect cat- egories are predicted via a set of binary classifiers. The one-vs-all strategy is used to train a binary clas- sifier for each category found in the training data. Each classifier is trained using a single layer feed- forward network. We enhance the system by adding neural network features learned from a Deep Con- volutional Neural Network system. For Slot 2, we treat the problem as a sequential labeling task, where sequential labeling classifiers are trained using Con- ditional Random Fields (CRF). The output of a Re- current Neural Network system is used as additional features. To generate Slot …

Content truncated to 500 words