Form Understanding and Information Extraction Using LayoutLMv3

Scanned forms are turned into labelled text blocks (header, question, answer, other) and question -> answer pairs. On the 199 noisy scanned forms of FUNSD, Tesseract 5.5.3 OCR, the specification's layout rules, a Random Forest on segment features and two layout-aware transformers - LayoutLMv3-base and LiLT, fine-tuned on a standard laptop CPU - are compared with ground-truth words and with real OCR words. LayoutLMv3 reaches an entity F1 of 0.800 (3 classes) against 0.731 for LiLT; with Tesseract words it falls to 0.518. A web application lets a person review and approve every extraction before it becomes a searchable record.

Category: Information Science and Technology

Information Science & Technology Project IST10 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.