Title:
Aspect-Based Sentiment Analysis of Online Reviews Using Word and Paragraph Vectors by Deep Learning Methods

Speaker:
ZAINUDIN, Suhaila (UKM)

Abstract:
Aspect-based Sentiment Analysis (ABSA) aims to extract major aspects of an item or product and predict the polarity of each aspect from the reviews. The importance of developing ABSA system is due to easing the process of decision making for customers and also suppliers to monitor their consumers by providing a decomposed view of rated aspects.  Previous methods tend to produce too many non-aspects and miss many of important ones. Domain specific models are often not practical for this task. Also limited works focus on implicit aspects and sentiments. Therefore the goal of this research is to develop an improved ABSA model that finds the most relevant aspect including implicit aspects with more accurate polarity estimation for each aspect in product reviews in different domains. We are going to take advantage of deep learning methods, which has gain significant interest recently. Word2Vec technique will be used for aspect extraction phase, and Doc2Vec technique will use for aspect polarity detection phase. Using these techniques, the system will outperform the previous ones in terms of the accuracy of aspect extraction and polarity prediction. Since deep learning methods do not use additional information such as parser and hand crafted features and sentiment lexicons, we hope to find more relevant aspects and more accurate polarity in multi domains for ABSA.

Extended Abstract:
pdf