A local LDA based method for Latent Aspect Rating Analysis on reviews

This paper proposes a Local LDA based method for LARA, which includes two stages, and uses LRR model to infer the latent rating and weight for each aspect based on the overall rating and the review content.

Abstract

The expanding volume of online reviews has made it an important and challenging task to mine detailed information of opinions in those reviews. In many cases, along with the comment, a user also gives an overall rating on the target entity, which in fact could not reflect the detailed opinions on each aspect of the entity. Therefore, Latent Aspect Rating Analysis (LARA) came into being. The goal of LARA is to infer a latent rating and weight for each aspect based on the overall rating and the review content. Although some methods have been applied to solve this problem, they rely too much on the predefinition of aspects with keywords, which needs supervision and may hence introduce some biases. In this paper, we propose a Local LDA based method for LARA, which includes two stages. In the first stage, we employ Local LDA to discover aspects automatically. In the second stage, we use LRR model to infer the latent rating and weight for each of the discovered aspects. The experimental results on the review dataset demonstrate the advantages of the proposed method over the state-of-the-art methods.