Unsupervised Training of HMM Structure and Parameters for OCRed List Recognition and Ontology Population

Machine learning based approaches to information extraction and ontology population often require a large number of manually selected and annotated examples in order to learn a mapping from facts asserted in text to structured facts asserted in an ontology. In this paper, we propose ListReader which provides a way to train the structure and parameters of a hidden Markov model (HMM) using text selected and labeled completely automatically. This HMM is capable of recognizing lists of records in OCRed and other text documents and associating subsets of identical fields across related record templates. The training method we employ is based on a novel unsupervised active grammar-induction framework that, after producing an HMM wrapper, uses an efficient active sampling process to complete the mapping from the HMM wrapper to ontology by requesting annotations from a user for automatically-selected examples. We measure performance of the final HMM in terms of F-measure of extracted information and manual annotation cost and show that ListReader (HMM) learns faster than a state-of-the-art baseline (CRF) and an alternate version of ListReader that induces a regular expression wrapper.