A Framework for predicting item ratings based on Aspect Level Sentiment Analysises

With the increase of review snippets on the internet, review rating prediction has become an interesting research problem which if solved will lead to a wide range of applications. Existing methods like Sentiment strength-based methods, Matrix Factorization, Deep neural network-based methods can solve this problem satisfactorily with supervision. However, these techniques suffer from data sparsity problem, and need extensive training using manually curated training sets. In order to overcome these limitations, we propose an unsupervised framework for item rating prediction. This framework uses Latent Dirichlet Allocation to extract topics from reviews and labels each topic with the aspect name. We estimate aspect ratings and aspect importance scores of each product, which are then used to predict the overall rating of a product as a linear combination of these aspect ratings weighted by corresponding aspect importance scores. We validate our model on a review dataset comprising of thousand reviews about two products. The performance metrics used to validate the framework are Confusion Matrix, F-measure, Root Mean Squared Error.