A novel approach based on Encoder Decoder technique for detecting Implicit Aspects

Abstract

Natural Language processing is a subset of Artificial Intelligence and one of the most important tasks in today’s world. Different tasks such as prediction, sentiment detection, aspect detection, sarcasm detection, translation from one language to another, emotion detection, etc. fall under Natural Language processing. Customer sentiment gathered from online social media websites gives organizations valuable insights about their products and services. The objective behind gathering the sentiments is to understand the needs and the choices of the customers so as to improvise and enhance the products and services for the customer. In the domain of mobile reviews, customers express their opinions about multiple features of the product. Knowing the sentiment about different features of the product is necessary. Features can be categorized as implicit or explicit. The explicit aspects are clearly mentioned in the review whereas implicit aspects are not mentioned and are indirectly referred to in the review. In aspect detection, recognizing the implicit aspects is very important as the owner of the review might be describing different opinions on different aspects of mobiles. A lot of study has been done on extracting explicit aspects while there are quite a few gaps in research while extracting implicit aspects. In this study, we have used Co-occurrence matrix technique, rule based method and encoder decoder technique with supervised learning method to find the implicit aspects in mobile reviews. The novelty of this research is that in the domain of mobile reviews, encoder decoder technique has been used in conjunction with supervised learning as a backup. Our method can detect explicit as well as implicit aspects. The encoder decoder technique gives us a good performance with an accuracy score of 82% in comparison to the co-occurrence matrix technique and rule based method. Our work will help other researchers working in same domain.