Ensemble Models for Aspect Category Related ABSA Subtasks

Abstract

Many E-commerce and social networking sites have a vast amount of data shared on them. This data is in the form of text, images, audio, and videos. However, people are more accustomed to sharing their experiences or views, about the products purchased by them using textual data. Usually, the users have a good and/or bitter experience about the particular feature of the product, instead of the product as a whole.  In this paper, we have performed sentiment analysis of reviews at a deeper level which is known as Aspect Based Sentiment Analysis (ABSA). ABSA allows analyzing data at a finer level. For ABSA, the Aspect Category Detection and Aspect Category Polarity are two subtasks of ABSA related to aspect category. These subtasks aim to detect the aspect categories referenced in the review along with the polarity for each of them. In this paper, we focus on these subtasks for Hindi ABSA Dataset. We compare the different ways of representing the review sentence using word vectors. We compare the performance of the Aspect Category Detection and Aspect Category Polarity subtask using two models. Among the two models- Ensemble model and Feed Forward Neural Network model, the Ensemble model provides significant improvement in performance for both subtasks. The Ensemble model with a sentence vector representation reports considerable improvement in F-score over state-of-the-art Aspect Category Detection results for all four major domains. Our proposed Ensemble model for Aspect Category Polarity subtask provides an increase in accuracy in the range of 7% to 14% for three of the four major domains over best state-of-the-art results.