A supervised method for ranking reviews based on latent structure features
Jiawei He, Kai Niu, Zhiqiang He, Siye Wang, Zhisong Bie · 2016
Previous research on reviews ranking has two branches: supervised method and unsupervised method. In this paper, we propose a supervised method to predict the review's importance, and the latent structure features are used to represent the review's writing styles. Our method mainly consists of three stages: Extracting lexical features to represent the review's lexical structure; Using lexical features as the input of a variant LDA topic model to learn latent structure features to represent review's writing styles; Obtaining the importance of a single review by a classifier. The experimental results show that our method performs better than the unsupervised baseline in accuracy, and our method is also more robust when compared with ordinary supervised baseline.