Multi-Domain Aspect Extraction Using Support Vector Machines

Multi-Domain Aspect Extraction using Support Vector Machines Nadheesh Jihan, Yasas Senarath, Dulanjaya Tennekoon, Mithila Wickramarathne, and Surangika Ranathunga Department of Computer Science and Engineering, University of Moratuwa, Katubedda 10400, Sri Lanka {nadheeshj.13, wayasas.13, dulanjayatennekoon.13, mithwick.13, surangika}@cse.mrt.ac.lk Abstract This paper describes a system to extract aspect categories for the task of aspect based sentiment analysis. This system can extract both implicit and explicit aspects. We propose a one-vs-rest Support Vector Machine (SVM) classifier preceded by a state of the art preprocessing pipeline. We present the use of mean embeddings as a feature along with two other new features to significantly improve the accuracy of the SVM classifier. This solution is extensible to customer reviews in different domains. Our results outperform the best recorded F1 score in the SemEval-2016 Task 5 dataset consisting of customer reviews from restaurant and laptop domains. Keywords: Aspect Extraction, Sentiment Analysis, Supervised Machine Learning, SVM, Preprocessing, Mean Embedding 1. Introduction The Internet has become the means of expressing opinion and view of consumers of products and services. Information contained in these reviews is of great value to other consumers as well as the companies that own those products and services. However, consumer reviews are often unstructured and noisy. Manual analysis of this huge amount of data for information is impossible. Automatic sentiment analysis of customer reviews has therefore, become a priority for the research community in the recent years. Conventional sentiment analysis of text focuses on the opinion of the entire text or the 308 The 2017 Conference on Computational Linguistics and Speech Processing ROCLING 2017, pp. 308-322  The Association for Computational Linguistics and Chinese Language Processing sentence. In the case of consumer reviews, it has been observed that customers often talk about multiple aspects of an entity and express an opinion on each aspect separately rather than expressing opinion towards the entity as a whole [1]. Aspect Based Sentiment Analysis (ABSA) has emerged to tackle this issue. The goal of Aspect Based Sentiment Analysis is to identify aspects present in the text, and the opinion expressed for each aspect [2]. One of the most crucial tasks of ABSA is to extract aspects from the review text. The state of the art systems have trouble in working with multiple domains, detecting multiple aspects in a single sentence, handling a large number of hierarchical aspects and detecting implicit aspects where the aspect is to be inferred from the context [3]. The objective of our research is to develop new techniques that would be able to perform aspect extraction from customer reviews with high accuracy, across multiple domains. A one-vs-rest Support Vector Machine (SVM) classifier and a list of carefully selected features are at the core of our supervised machine learning approach for aspect extraction. We identified that when Mean Embeddings are provided as a feature to the SVM classifier, results get improved significantly. The system was further enhanced using a clever text pre-processing pipeline complemented with context sensitive spell correction. Our system is able to outperform the …

Content truncated to 500 words