ECNU at SemEval-2016 Task 5: Extracting Effective Features from Relevant Fragments in Sentence for Aspect-Based Sentiment Analysis in Reviews

Proceedings of SemEval-2016, pages 361–366, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics ECNU at SemEval-2016 Task 5: Extracting Effective Features from Relevant Fragments in Sentence for Aspect-Based Sentiment Analysis in Reviews Mengxiao Jiang1, Zhihua Zhang1, Man Lan1,2∗ 1Department of Computer Science and Technology, East China Normal University, Shanghai, P.R.China 2Shanghai Key Laboratory of Multidimensional Information Processing {51151201080, 51131201039}@ecnu.cn, mlan@cs.ecnu.edu.cn∗ Abstract This paper describes our systems submitted to the Sentence-level and Text-level Aspect- Based Sentiment Analysis (ABSA) task (i.e., Task 5) in SemEval-2016. The task involves two phases, namely, Aspect Detection phase and Sentiment Polarity Classification phase. We participated in the second phase of both subtasks in laptop and restaurant domains, which focuses on the sentiment analysis based on the given aspect. In this task, we extract- ed four types of features (i.e., Sentiment Lexi- con Features, Linguistic Features, Topic Mod- el Features and Word2vec Feature) from cer- tain fragments related to aspect rather than the whole sentence. Then the proposed features are fed into supervised classifiers for senti- ment analysis. Our submissions rank above average. 1 Introduction Aspect-Based Sentiment Analysis task (ABSA), i.e., task 5 in SemEval-2016, is an interesting task, which focuses on the sentiment analysis based on the target and certain categories. The organizers set up three subtasks, i.e., Sentence-level ABSA (i.e, Subtask 1), Text-level ABSA (i.e., Subtask 2), and Out-of do- main ABSA (i.e., Subtask 3). For subtask 1 and 2, the training data in each domain is provided, while no labeled data is provided for subtask 3. Given an opinionated document in a domain, both subtask 1 and subtask 2 are grouped into two phases, i.e., As- pect Detection phase and Sentiment Polarity Classi- fication phase. We participated in the second phase of these two subtasks, aiming to identify the senti- ment polarity for each given aspect which is made up of <E#A, OTE> in two domains. Specifically, for Sentence-level ABSA, focusing on identifying all the opinion tuples (i.e., <E#A, OTE, polarity>) in each sentence of the reviews, the Aspect Detection phase contains two slots. The Slot1 is to identify every entity (i.e., E) and at- tribute (i.e., A) pair (also named as category, e.g., RESTAURANT-PRICES) according to given sen- tence, and the Slot2 focuses on detecting Opinion Target Expression (i.e., OTE or target in short). For example, in “Pizza here is consistently good”, the participants are required to recognize Pizza as OTE and FOOD#QUALITY as E#A. The second phase, i.e., Sentiment Polarity Classification (Slot3), is to determine the sentiment polarity (i.e., positive, nega- tive, or neutral) for each aspect (i.e., <E#A, OTE>). As for Text-level ABSA, aiming at identifying the opinion tuples (i.e., <E#A, polarity>) expressed in each review, the Aspect Detection phase is to identi- fy the E#A pairs and the second phase is to assign the sentiment label (positive, negative, neutral or con- flict) for each detected E#A. The conflict label is as- signed when the dominant sentiment of the opinion is not clear. In previous work, (Kim et …

Content truncated to 500 words