Cost-Effective Information Extraction from Lists in OCRed Historical Documents
Thomas L. Packer, David W. Embley · 2015
To work well, machine-learning-based approaches to information extraction and ontology population often require a large number of manually selected and annotated examples. In this paper, we propose ListReader which provides a way to train the structure and parameters of a Hidden Markov Model (HMM) without requiring any labeled training data. The induced HMM is a wrapper---a function that hides within it the complexities of low-level processing---in ListReader's case the complexities of information extraction from OCRed historical documents. The HMM wrapper is capable of recognizing lists of records in text documents and associating subsets of identical fields across related record templates. The algorithmic training method we employ is based on a novel unsupervised active grammar-induction framework. The training produces an HMM wrapper and uses an efficient active sampling process to complete the mapping from 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 learns faster and better than a state-of-the-art baseline and an alternate version of ListReader that induces a regular-expression wrapper.