Noise-Aware Named Entity Recognition for Historical VET Documents

TL;DR

This work tackles robust named entity recognition (NER) in historical German VET documents afflicted by OCR noise. It introduces Noise-Aware Training (NAT) with three data variants—noisy, clean, and artificial—combined with transfer learning and multi-stage fine-tuning to build a domain-adaptive NER model capable of identifying multiple entity types (JOB_TITLE, JOB_TITLE_GROUP, SKILL, SUBJECT, ACTIVITY). The artificial model, trained on synthetically perturbed data, achieves the best overall $F_1$ score (77.9%), with notable performance for job titles (up to $F_1$ ≈ 87.9%), and reveals how data source alignment and entity type alignment influence results. The approach demonstrates the value of domain-specific and noise-aware fine-tuning for robust information extraction from historical documents and provides publicly available code to support reproducibility and adoption. The work also discusses limitations and avenues for improvement, including OCR improvement, layout analysis, and extending multi-entity recognition in the VET domain.

Abstract

This paper addresses Named Entity Recognition (NER) in the domain of Vocational Education and Training (VET), focusing on historical, digitized documents that suffer from OCR-induced noise. We propose a robust NER approach leveraging Noise-Aware Training (NAT) with synthetically injected OCR errors, transfer learning, and multi-stage fine-tuning. Three complementary strategies, training on noisy, clean, and artificial data, are systematically compared. Our method is one of the first to recognize multiple entity types in VET documents. It is applied to German documents but transferable to arbitrary languages. Experimental results demonstrate that domain-specific and noise-aware fine-tuning substantially increases robustness and accuracy under noisy conditions. We provide publicly available code for reproducible noise-aware NER in domain-specific contexts.

Figures (5)

• Figure 1: Example of an OCR error in a job title.
• Figure 2: Document processing pipeline.
• Figure 3: Multi-stage training process.
• Figure 4: Precision, recall, and $F_1$ score over time.
• Figure 5: Confusion matrix for entity recognition.