A Span Extraction Approach for Information Extraction on Visually-Rich Documents

Information extraction (IE) for visually-rich documents (VRDs) has achieved SOTA performance recently thanks to the adaptation of Transformer-based language models, which shows great potential pre-training methods. In this paper, we present a new approach improve capability model on VRDs. Firstly, introduce query-based IE that employs span instead using common sequence labeling approach. Secondly, extend formulation, propose training task focusing modelling relationships among semantic entities within document. This enables target spans be extracted recursively and can used pre-train or as an downstream task. Evaluation three datasets popular business (invoices, receipts) our proposed method achieves significant improvements compared existing models. The also provides mechanism knowledge accumulation from multiple tasks.