ECNU: A Combination Method and Multiple Features for Aspect Extraction and Sentiment Polarity Classification

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 252–258, Dublin, Ireland, August 23-24, 2014. ECNU: A Combination Method and Multiple Features for Aspect Extraction and Sentiment Polarity Classification Fangxi Zhang, Zhihua Zhang, Man Lan∗ Department of Computer Science and Technology East China Normal University 51111201041,51131201039@ecnu.cn; mlan@cs.ecnu.edu.cn∗ Abstract This paper reports our submissions to the four subtasks of Aspect Based Sentimen- t Analysis (ABSA) task (i.e., task 4) in SemEval 2014 including aspect term ex- traction and aspect sentiment polarity clas- sification (Aspect-level tasks), aspect cat- egory detection and aspect category sen- timent polarity classification (Category- level tasks). For aspect term extraction, we present three methods, i.e., noun phrase (NP) extraction, Named Entity Recogni- tion (NER) and a combination of NP and NER method. For aspect sentiment classi- fication, we extracted several features, i.e., topic features, sentiment lexicon features, and adopted a Maximum Entropy classifi- er. Our submissions rank above average. 1 Introduction Recently, sentiment analysis has attracted a lot of attention from researchers. Most previous work attempted to detect overall sentiment polarity on a text span, such as document, paragraph and sen- tence. Since sentiments expressed in text always adhere to objects, it is much meaningful to iden- tify the sentiment target and its orientation, which helps user gain precise sentiment insights on spe- cific sentiment target. The aspect based sentiment analysis (ABSA) task (Task 4) (Pontiki et al., 2014) in SemEval 2014 is to extract aspect terms, determine its se- mantic category, and then to detect the sentimen- t orientation of the extracted aspect terms and its category. Specifically, it consists of 4 subtasks. The aspect term extraction (ATE) aims to extrac- t the aspect terms from the sentences in two giv- 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 details: http://creativecommons.org/licenses/by/4.0/ en domains (laptop and restaurant). The aspec- t category detection (ACD) is to identify the se- mantic category of aspects in a predefined set of aspect categories (e.g., food, price). The aspect term polarity (ATP) classification is to determine whether the sentiment polarity of each aspect is positive, negative, neutral or conflict (i.e., both positive and negative). The aspect category po- larity (ACP) classification is to determine the sen- timent polarity of each aspect category. We partic- ipated in these four subtasks. Generally, there are three methods to extract as- pect terms: unsupervised learning method based on word frequency ((Ku et al., 2006), (Long et al., 2010)), supervised machine learning method (Kovelamudi et al., 2011) and semi-supervised learning method (Mukherjee and Liu, 2012) where only several user interested category seeds are given and used to extract more categorize aspect terms. Since sentiments always adhere to entities, several researchers worked on polarity classifica- tion of entity. For example, (Godbole et al., 2007) proposed a system that assigned scores represent- ing positive or negative opinion to each distinc- t entity in the corpus. (Kim et al., 2013) presented …

Content truncated to 500 words