Hybrid Learning Approach Based Aspect Category Detection for Sentiment Summarization with Co-Occurrence Data

Abstract

User-generated reviews are precious decision-making resources. Identifying the feature categories mentioned in a specified review phrase (e.g. "food" and "service" in restaurant reviews) is a significant task for analyzing sentiment and mining opinion. Most prior researchers hold hand-crafted characteristics and a classification algorithm to achieve the assignment given a predefined aspect category set. The key step to achieve better efficiency is feature engineering that consumes a great deal of human effort and can be volatile when the product domain changes. A hybrid learning method is suggested in this project to automatically learn helpful characteristics for the identification of aspect categories. Specifically, on a big collection of reviews with noisy labels, a Hybrid Aspect Analysis Algorithm is first suggested to achieve ongoing word depictions. We subsequently suggest generating deeper and hybrid characteristics through the stacked neural networks on the word vectors. Finally, a logistic regression classifier is trained to predict the aspect category with hybrid characteristics. The tests are conducted on a SemEval-2014 benchmark dataset. In this paper we achieves the state of the art results with the F1 score of 90.10% on the dataset. Overall, our approach to representation learning outperforms traditional hand-crafted characteristics and embedding algorithms with current words.