Reproduce “FLERT: Document-Level Features for Named Entity Recognition” on OpenResearch

About this paper

Stefan Schweter, Alan Akbik

Researchers at Humboldt-Universität zu Berlin systematically evaluated the impact of document-level context on transformer-based Named Entity Recognition (NER), achieving F1-score improvements of approximately 1.15 percentage points by extending context beyond sentence boundaries. Their "FLERT" approach, particularly when fine-tuning transformers and explicitly enforcing document boundaries, led to new state-of-the-art results on several CoNLL-03 datasets for German and Dutch.

Key insights

• Document-level features consistently enhance NER performance across architectures and languages, contributing an average F1-score increase of approximately 1.15 percentage points.
• Fine-tuning transformer models with document-level context generally outperforms feature-based LSTM-CRF approaches, showing an average F1-score advantage of about 2 percentage points.
• Explicitly enforcing document boundaries (truncating context at document ends) is crucial for performance, improving F1-scores and suggesting that transformers do not inherently learn to respect these structural separations.

Summary by alphaXiv.