A Hybrid Model for Context-Dependent Sentiment Analysis

Abstract

In this thesis, we propose a hybrid model that combined a unigram Naive Bayes bag-of- words machine learning model and the recursive structure of a Recursive Neural Tensor Network model where we would directly encode specifics of linguistic patterns into sentiment detector. The goal was to create a program that could detect sentiment quickly, without the long training time required by RNTN, while still capturing the lexical patterns that were largely missed by the bag-of-words model. The model was tested against the Yelp Academic Dataset, which contains over 300,000 reviews for restaurants and services in the Phoenix, Arizona area. While a bag-of-words model was able to achieve a 53.15% accuracy, the hybrid parse only achieved a 51.96% accuracy. Even so, it was able to make significant improvement in the accuracy for the class of reviews rated 3, increasing its accuracy from 25.89% to 33.76%, while all other classes only suffered a very marginal decrease in accuracy.