Active Hidden Markov Models for Information Extraction

Information extraction from HTML documents requires a classifier capable of assigning semantic labels to the words or word sequences to be extracted. If completely labeled documents are available for training, well-known Markov model techniques can be used to learn such classifiers. In this paper, we consider the more challenging task of learning hidden Markov models (HMMs) when only partially (sparsely) labeled documents are available for training. We first give detailed account of the task and its appropriate loss function, and show how it can be minimized given an HMM. We describe an EM style algorithm for learning HMMs from partially labeled data. We then present an active learning algorithm that selects "difficult" unlabeled tokens and asks the user to label them. We study empirically by how much active learning reduces the required data labeling effort, or increases the quality of the learned model achievable with a given amount of user effort.