Knowledge-Based Neural Pre-training for Intelligent Document Management

Banks are usually large and complex companies that face a number of challenges to support the rapid and effective sharing of information and content across their organizations. Extracting complex metadata from raw bank documents is therefore central to support intelligent data indexing, information circulation and to promote more complex predictive capabilities, e.g., compliance assessment problems. In this paper, we present a weakly-supervised neural methodology for creating semantic metadata from bank documents. It exploits a neural pre-training method optimized against legacy semantic resources able to minimize the training effort. We studied an application to business process design and management in banks and tested the method on documents from the Italian banking community. The measured impact of the proposed training approach to process-related metadata creation confirms its applicability.