Exploring Bayesian Uncertainty Modeling for Book Genre Classification

In this paper, we aim to model the Bayesian uncertainty of a model designed to solve the task of book genre classification. Model prediction confidence can judge the predictive quality and usability of predictions made from a machine learning model. This work explores two methods to ascertain model uncertainty using Monte Carlo dropouts and deep ensembling. We apply uncertainty modeling to a bidirectional LSTM model trained on the CMU book summary dataset to perform book genre classification from book summaries. We show how these techniques improve results by 14% from the best baseline model and discuss their feasibility in real-world scenarios.