A Data Preprocessing Method to Classify and Summarize Aspect-Based Opinions Using Deep Learning

Abstract

Opinion summarization is based on aspect analyses of products, events or topics, which is a very interesting topic in natural language processing. Opinions are often expressed in various different ways in regards to objects. Therefore, it is important to express the characteristics of a product, event or topic in a final summary compiled by an automatic summarizing system. This paper proposes a method for conducting data preprocessing on the sentence level of a text using Convolutional Neural Networks. The corpus includes Vietnamese opinions on cars collected from social networking sites, forums, online newspapers and the websites of automobile dealers. The data processing phase will standardize terms for aspects that occur in opinion expressing aspects of the product. These aspects are used by manufacturers. Similarly, the standardization will be performed for both positive and negative terms used in opinions. The sentiment terms in the opinions will be replaced by standardized sentiment terms expressing the same sentiment polarities as those being replaced. This standardization is performed with the support of a semantic and sentiment ontology which has a tree hierarchy in the case of cars. This ontology ensures that the semantics and sentiment of the original opinion are not changed. The experimental results of the paper show that the proposed method gives better results than using no data preprocessing method for deep learning.