DLIREC: Aspect Term Extraction and Term Polarity Classification System

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 235–240, Dublin, Ireland, August 23-24, 2014. DLIREC: Aspect Term Extraction and Term Polarity Classification System Zhiqiang Toh Institute for Infocomm Research 1 Fusionopolis Way Singapore 138632 ztoh@i2r.a-star.edu.sg Wenting Wang magicwwt@gmail.com Abstract This paper describes our system used in the Aspect Based Sentiment Analysis Task 4 at the SemEval-2014. Our system con- sists of two components to address two of the subtasks respectively: a Conditional Random Field (CRF) based classifier for Aspect Term Extraction (ATE) and a linear classifier for Aspect Term Polarity Classi- fication (ATP). For the ATE subtask, we implement a variety of lexicon, syntac- tic and semantic features, as well as clus- ter features induced from unlabeled data. Our system achieves state-of-the-art per- formances in ATE, ranking 1st (among 28 submissions) and 2rd (among 27 submis- sions) for the restaurant and laptop domain respectively. 1 Introduction Sentiment analysis on document and sentence level no longer fulfills user’s needs of getting more accurate and precise information. By perform- ing sentiment analysis at the aspect level, we can help users gain more insights on the sentiments of the various aspects of the target entity. Task 4 of SemEval-2014 provides a good platform for (1) aspect term extraction and (2) aspect term polar- ity classification. For the first subtask, we follow the approach of Jakob and Gurevych (2010) by modeling term ex- traction as a sequential labeling task. Specifically, we leverage on semantic and syntactic resources to extract a variety of features and use CRF as the learning algorithm. For the second subtask, we This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence de- tails: http://creativecommons.org/licenses/ by/4.0/ simply treat it as a multi-class classification prob- lem where a linear classifier is learned to predict the polarity class. Our system achieves good per- formances for the first subtask in both domains, ranking 1st for the restaurant domain, and 2nd for the laptop domain. The remainder of this paper is structured as fol- lows: In Section 2, we describe our ATE system in detail, including experiments and result analy- sis. Section 3 describes the general approach of our ATP system. Finally, Section 4 summarizes our work. 2 Aspect Term Extraction This subtask is to identify the aspects of given tar- get entities in the restaurant and laptop domains. Many aspect terms in the laptop domain con- tains digits or special characters such as “17 inch screen” and “screen/video resolution”; while in the restaurant domain, aspects in the sentences are specific for a type of restaurants such as “pizza” for Italian restaurants and “sushi” for Japanese restaurants. We model ATE as a sequential labeling task and extract features to be used for CRF training. Besides the common features used in traditional Named Entity Recognition (NER) systems, we also utilize extensive external resources to build various name lists and word clusters. 2.1 Preprocessing Following the traditional BIO …

Content truncated to 500 words