A Collaborative Framework for Structure Identification over Print Documents

Abstract

We describe Texture, a framework for data extraction over print documents that allows end-users to construct data extraction rules over an inferred document structure. To effectively infer this structure, we enable developers to contribute multiple heuristics that identify different structures in English print documents, crowd-workers and annotators to manually label these structures, and end-users to search and decide which heuristics to apply and how to boost their performance with the help of ground-truth data collected from crowd-workers and annotators. Texture's design supports each of these different user groups through a suite of tools. We demonstrate that even with a handful of student-developed heuristics, we can achieve reasonable precision and recall when identifying structures across different document collections.