JoWUA

Abstract

Automated extraction of structured information from handwritten financial forms is an important and less targeted challenge in document intelligence. Public datasets to date are mainly limited to printed financial documents - invoices, receipts and standard business forms - but do not capture the hierarchical structure, handwriting style variation, or multiple languages seen in actual banking, especially in developing economies where form filling is done manually. To fill this gap, the FinScan Document Dataset is presented as a new dataset containing 4,046 annotated financial documents across five form types: Account Opening Forms, Loan Applications, KYC Forms, cheques, and salary slips. Each document is rendered at 1,198 × 1,978 pixels and annotated in FUNSD format with this dataset contains a total of 319,850 token-level bounding boxes and three-class NER labels: B-QUESTION, B-ANSWER, O. A hierarchical evaluation protocol is proposed that aligns six essential dimensions for validating the corpus: Document Image Quality Assessment (DIQA), OCR/HTR performance, layout recognition, table structure validation, Key Information Extraction (KIE), and document-level correctness-forming the first comprehensive evaluation framework in financial document understanding that integrates all six dimensions. Experiments show an overall CER of 0.58%, mean layout IoU of 0.7612 and KIE-F1 of 0.8690. Cascade analysis shows that layout false negatives are the most common failure type, accounting for 99.5% of document-level failures. This analysis identifies layout improvement as the most impactful research direction for advancing financial document benchmarking. FinScan functions as a research benchmark for cloud-based financial document processing services. In this context, handwritten banking forms are ingested, extracted, and returned as structured JSON through web APIs. Future work will address cloud scalability and microservice integration.