Perry-DLC/upy-tds-bert-ner-conll2003

BERT NER on CoNLL-2003

Intended use

English newswire named entity recognition (PER, ORG, LOC, MISC) for a course experiment. Not a general-domain or safety-critical extractor.

Model and training

• Base: bert-base-cased
• Dataset: lhoestq/conll2003; splits train/validation/test; original shared task: CoNLL-2003.
• Adaptation: full; initial seed 42; 3 epochs; batch 16; maximum sequence length 256.
• Two compared alternatives: frozen BERT with trained token head and full fine-tuning. Head learning rate 1e-3; encoder learning rate 2e-5 when trained.
• Labels: O, B-PER, I-PER, B-ORG, I-ORG, B-LOC, I-LOC, B-MISC, I-MISC. Supervise the first subtoken only; other positions -100.

Results

Entity-level strict IOB2 span F1 on held-out test: 0.9120; precision 0.9113; recall 0.9127; token accuracy 0.9824. Validation F1: 0.9455. Single seed; close differences may be noise.

Limitations

English Reuters news from 1996; domain and time shift can hurt performance. One label per word; truncated sentences beyond 256 subtokens lose supervised words. Named entities and noisy labels may differ across datasets. Review the upstream dataset terms before reuse or redistribution.

References

• Tjong Kim Sang & De Meulder (2003), CoNLL-2003.
• Devlin et al. (2019), BERT.
• Hugging Face token classification.