Unsupervised Sentiment Classification: A Hybrid Sentiment-Topic Model Approach

With the large volume of text available online it is becoming impractical to use supervised machine learning methods that require a sizeable training set of labelled data. In this paper we introduced a new sentiment-topic model called the hybrid sentiment-topic model (HST). The HST model is a completely unsupervised sentiment classification method that allows for the topical context of words in documents to be accounted for when classifying sentiment. The only input needed for the model is a list of positive seed words, a list of negative seed words, and the number of topics. The HST model differs from similar models as it ensures that each objective topic discovered has both a positive sentiment-topic and negative sentiment-topic associated with it; other similar models do not guarantee symmetric sentiment-topics. The HST model performs three functions, firstly, it discovers objective topics in a corpus of text; secondly, it finds a positive and negative sentiment-topic for each objective topic; and finally, it performs sentiment classification. The HST model is tested using a dataset consisting of movie reviews and a dataset of social media posts. For each dataset a variety of seed word lists and different numbers of topics are tested; the HST model is then compared against similar sentiment-topic models. In all experiments conducted, the HST model was found to outperform similar sentiment-topic models in terms of classification accuracy by a noticeable margin. Additionally, the HST model was found to converge faster than similar models and the accuracy was found to be more stable during the Gibbs sampling process.