ECNU: Extracting Effective Features from Multiple Sequential Sentences for Target-dependent Sentiment Analysis in Reviews

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 736–741, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics ECNU: Extracting Effective Features from Multiple Sequential Sentences for Target-dependent Sentiment Analysis in Reviews Zhihua Zhang, Man Lan∗ Shanghai Key Laboratory of Multidimensional Information Processing Department of Computer Science and Technology, East China Normal University, Shanghai 200241, P. R. China 51131201039@ecnu.cn, mlan@cs.ecnu.edu.cn∗ Abstract This paper describes our systems submitted to the target-dependent sentiment polarity clas- sification subtask in aspect based sentimen- t analysis (ABSA) task (i.e., Task 12) in Se- mEval 2015. To settle this problem, we ex- tracted several effective features from three se- quential sentences, including sentiment lexi- con, linguistic and domain specific features. Then we employed these features to con- struct classifiers using supervised classifica- tion algorithm. In laptop domain, our systems ranked 2nd out of 6 constrained submissions and 2nd out of 7 unconstrained submissions. In restaurant domain, the rankings are 5th out of 6 and 2nd out of 8 respectively. 1 Introduction Reviews express opinions of customers towards var- ious aspects of a product or service. Mining cus- tomer reviews (i.e., opinion mining) has emerged as an interesting new research direction in recen- t years. Since sentiment expressed in reviews usu- ally adheres to specific categories or target terms, it is much meaningful to identify the sentiment target and its orientation, which helps users gain precise sentiment insights on specific sentiment target. Unlike most existing sentiment analysis method- s which try to detect the polarity of a sentence or a review, the aspect based sentiment analysis task (ASBA) shared as task 12 in SemEval 2015 is aim- ing at addressing the category- or target- dependent sentiment analysis in reviews. There are two types of subtasks organized in ASBA. The first aspect de- tection subtask is to identify the sentiment adherent from reviews, i.e., the category (i.e., entity-attribute (E-A) pair) or opinion target expression (OTE) in re- views. In most cases, the customers may not ex- plicitly indicate the entity and attribute words in re- views but the opinion target expression is a segment of review. For example, in a given review: “The pizza is overpriced and soggy.”, target=“pizza”, category=“FOOD-QUALITY”. Its category label FOOD-QUALITY does not exist in reviews, while it- s OTE word “pizza” is explicitly present in reviews. The second sentiment polarity classification subtask is to assign a polarity label (i.e., positive, negative or neutral) for every E-A pair or OTE identified from the given reviews. We participated the second type subtask, i.e., performing sentiment polarity classifi- cation on reviews. There are two domains in this sentiment analysis subtask, i.e, laptop and restauran- t. In laptop domain, only E-A pairs are annotated and provided in reviews while in restaurant domain, both E-A pairs and OTE are provided. Comparing with laptop reviews, the restaurant reviews provide the annotated surface words adhering to sentimen- t. Therefore we speculate that the performance in restaurant domain would be much better than that of laptop domain. …

Content truncated to 500 words