Text-guided tabular learning for case processing

Contact person: Martin Jullum

Status: Project

This project addresses an “inference gap” in automated support and filtering for medium- and high-stakes case processing, ranging from customer and citizen services to fraud detection and criminal investigations. In these settings, caseworkers and investigators typically produce case notes and reports explaining why a case receives a particular outcome -- for example, why a complaint is upheld or a claim is judged fraudulent. However, automated systems used to select or prioritize new cases are usually trained only on these outcomes and the structured tabular data available at the time. As a result, the models receive little guidance about the reasoning and information that actually matter in practice. In this project, we aim to develop a Text-Guided Tabular Learning framework, based on LUPI (Learning Using Privileged Information) and cross-modal knowledge distillation, to bridge this gap. A text-aware “teacher” model uses historical reports to guide the training of a regular tabular “student” model, effectively transferring the caseworker’s intuition into an automated system that does not require text when assessing new cases. The results could help users identify relevant cases more accurately, reduce false positives, and make better use of the expert knowledge already captured through case processing, while taking relevant legal and ethical considerations into account.