Evaluating Topic Modeling as Preprocessing for a Sentiment Analysis Task

Classifying the sentiment of documents is a well-studied problem in Natural Language Processing (NLP). The existence of excellent discriminative classifiers like Maxent has pushed the main body of research in the direction of feature engineering. In this paper, I examine an unusual class of features, the document-topic proportions assigned by the Latent Dirichlet Allocation topic model. In particular, I evaluate the relative performance of topic proportions alone, topic proportions together with unigrams, and unigrams alone. My findings are that topic proportions do not add enough information to significantly improve classification performance. I also perform domain adaptation experiments, finding that (although topic proportions do not outperform unigram features) there is some promise in classification based on topic proportions.