A Report on the First Workshop on Document Intelligence (DI) at NeurIPS 2019

Abstract

The first workshop on Document Intelligence (DI-2019) was held on December 14, 2019 at NeurIPS 2019 conference in Vancouver, Canada. The report summarizes the workshop, with a summary of the talks, papers and posters presented, and discusses common themes, issues and open questions that came up in the workshop.

Data-Efficient Information Extraction from Documents with Pre-trained Language Models

“…Such documents cover a wide variety of types, including invoices, purchase orders, receipts, vendor contracts, financial reports and employment agreements. To cope with the increasing volume of business documents to process, academic and industrial practitioners have leveraged AI techniques to automatically read, understand and interpret them [24]. This research topic, recently referred to as Document Intelligence (DI), comprises multiple disciplines ranging from Natural Language Processing, Computer Vision over Information Retrieval to Knowledge Representation and Reasoning among others.…”

“…To tackle the diversity and complexity of document structure and content, current Information Extraction (IE) approaches employ deep neural networks that learns from annotated documents. Yet, as for many tasks in DI, labeling documents is a challenge in IE since it involves significant human expertise in the targeted application domain [24]. Besides, the extraction objectives are highly specific to the type of documents to process, hindering the reusability of a trained IE model.…”

SynthTIGER: Synthetic Text Image GEneratoR Towards Better Text Recognition Models

“…Optical character recognition (OCR) is a technology extracting machine-encoded texts from text images. It is a fundamental function for visual understanding and has been used in diverse real-world applications such as automatic number plate recognition [16], business document recognition [15,4,3] and passport recognition [9]. In the deep learning era [10,11], OCR performance has been dramatically improved by learning from large-scale data consisting of image and text pairs.…”