A Neural Network based Model for Determining Overall Aspect Weights in Opinion Mining and Sentiment Analysis

Opinion mining and sentiment analysis has recently attracting many studies in the field of text mining and knowledge discovery. Its task is to identify and analyze subjective information of customers' opinion from social media sources in the Web. The two kinds of objects of this study are products and product's aspects which are given in reviews from sources such as social networks, merchant sites, blogs, forums. This paper focuses on determining the important degree of aspects given a set of reviews. Suppose that each given review is assigned with aspects' ratings and overall rating. Under our opinion, the overall rating of a review is derived from its aspect ratings. This observation inspires us to formulate these factors in a neural network. This proposed model can generate the weights for aspects which reflect aspects' important degrees. Doing experiment for this model, we use a dataset of 397528 reviews of 2558 hotels which are collected from an well known tourist website - tripadvisor.com. Five common aspects are used include "cleanliness", "location", "service", "room" and "value". The obtained experimental result shows that our proposed method outperforms some well known studies for the same problem such as the probabilistic rating regression method or the frequency-based method.