<![CDATA[Regulatory compliance activities across the financial sector produce a vast and growing amount of unstructured evidentiary material, such as loan agreements, non-disclosure agreements, annual reports, disclosure forms, scanned correspondence, and more — all of which auditors and regulators must review to assure compliance with disclosure, anti-money-laundering and data-protection requirements. Text-only Natural Language Processing pipelines are unable to handle this type of content because the facts relevant to compliance are often embedded in the format of a page, such as signature blocks, tabular disclosures and heading to clauses. This paper summarizes the results of two and a half years of work, including twenty-five papers from the field of transformer language modeling, the field of layout-aware document representation learning, the field of financial natural language processing and the field of applied regulatory-technology systems, all of which were published between 2015 and 2022. It explores how the performance of text-only encoders jumped from an F1 score of around sixty percent on benchmarks for form understanding to the layout-aware multimodal architectures that exceeded the ninety percent mark on the same benchmark, and reviews applied systems that lowered the manual anti-money-laundering review workload by more than fifty percent while maintaining detection rates above eighty percent. The review also outlines five benchmark datasets that test extraction tasks that are related to compliance and suggests a generalized six-stage extraction pipeline extracted from the reviewed architectures. The results show that layout-conditioned pretraining yields the highest improvements, not just a parameter scale, and that the human verification is a constant characteristic of all the deployed systems studied.]]>
Copyrights © 2023
