Lightly-supervised learning methods for one-class text classification

Abstract

This thesis introduces a lightly-supervised learning method to train text classifiers with very little manual labelling effort. We adapt two previous state-of-theart lightly-supervised models, generalized expectation (GE) criteria (Druck et al. (2008)) and multinomial naïve Bayes (MNB) with priors (Settles (2011)) to oneclass classification problem. Users just need to label a handful of keywords for the target category. We also combine the two aforementioned models by letting MNB automatically augment the list of GE constraints. In addition, we ensemble two families of classifiers to improve the accuracy further. We successfully applied our model to a real-world problem of online advertising. On a corpus of online advertising data, the proposed model achieved the top macro average F₁ of 0.69 and closed 50% gap between previous state-of-the-art lightly-supervised models and a fully-supervised model MaxEnt model.