A Simple Proposal for Sentiment Analysis on Movies Reviews with Hidden Markov Models

Abstract

Sentiment analysis of texts is the field of study which analyses and studies opinions, sentiments, value judgments, affections and emotions in texts like blogs, news and treating of products, organisations, events and topics. If information on subjective content is required, such as the emotion aroused by an event, computer techniques must be applied to analyse the pattern of public opinion. A common technique for analysing texts is the “Bag of Words”, which provides good results assuming that the words are independent of one another. In this work we propose the use of Hidden Markov Chains to determine the polarity of the opinions expressed on movie reviews. We propose a method for simulating hidden states through clustering techniques; we then carry out a sensitivity analysis of the model in which we apply variations to model parameters such as the number of hidden states or the number of words used. The results show that our proposal gives a 3% improvement over the basic model using F-score for real databases of public opinion.Sentiment analysis of texts is the field of study which analyses and studies opinions, sentiments, value judgments, affections and emotions in texts like blogs, news and treating of products, organisations, events and topics. If information on subjective content is required, such as the emotion aroused by an event, computer techniques must be applied to analyse the pattern of public opinion. A common technique for analysing texts is the “Bag of Words”, which provides good results assuming that the words are independent of one another. In this work we propose the use of Hidden Markov Chains to determine the polarity of the opinions expressed on movie reviews. We propose a method for simulating hidden states through clustering techniques; we then carry out a sensitivity analysis of the model in which we apply variations to model parameters such as the number of hidden states or the number of words used. The results show that our proposal gives a 3% improvement over the basic model using F-score for real databases of public opinion.