AsphodelRem/uz-ner-rembert

Uzbek exact-span NER (RemBERT + domain-adaptive MLM)

Named entity recognition for Uzbek text with exact character offsets. Three classes: ORG (organizations and brands), NAME (people), GEO (geographic entities).

Handles both Latin and Cyrillic scripts, plus Russian and English fragments.

Metrics

Exact-span match on a held-out set (1500 documents, 7698 entities). An entity counts only when the class and both boundaries match exactly.

class	precision	recall	F1
ORG	0.8872	0.8665	0.8767
NAME	0.9182	0.8814	0.8994
GEO	0.9153	0.9096	0.9124
micro	0.9065	0.8862	0.8962

Intended use

Extracting people, organizations and locations from Uzbek news and social media text. Exact character offsets make the output suitable for downstream analytics: entity frequency and co-occurrence statistics, media monitoring, building entity indexes over document collections, and linking mentions back to their exact position in the source text.

Important: weights alone are not enough

Tags are predicted in the BILOU scheme and decoded with the Viterbi algorithm using the transition matrix in transitions.json. That file is not a standard Hugging Face artifact, so AutoModelForTokenClassification.from_pretrained will ignore it. Without it decoding falls back to per-token argmax and F1 drops by roughly one point. Boundary post-processing (quotes, brackets, affixes) is also applied and also matters for exact-span scoring.

Run the model through the solution repository instead:

python -m solution.predict \
  --model-dir <directory with these files> \
  --input texts.jsonl \
  --output predictions.jsonl

Input is JSON Lines with hash and text. Output is hash plus a list of {"label", "start", "end"}, where offsets are zero-based Unicode character positions and end is exclusive.

Inference settings come from run_config.json: window 512 with stride 128, BILOU, Viterbi in conditional mode with transition weight 0.4.

Training

Two stages.

Domain adaptation. google/rembert was further pretrained with masked language modeling on an Uzbek corpus of 405,585 passages (~220M tokens): Telegram blogs and news from tahrirchi/uz-crawl, Cyrillic and Latin books from tahrirchi/uz-books-v2, plus 8% Russian and English Wikipedia to preserve multilinguality. Whole-word masking at 15%, 16,800 steps, held-out perplexity 11.5.

Fine-tuning. Token classification in BILOU, window 512 with stride 128, effective batch 128, learning rate 3e-5, checkpoint selected by exact-span micro-F1.

Full reproduction protocol is in REPRODUCE.md of the solution repository.

Limitations

Trained on news and blog text; other genres were not evaluated.

Attribution

The base model google/rembert is licensed under Apache 2.0. The adaptation corpus was assembled from tahrirchi/uz-crawl, tahrirchi/uz-books-v2 and wikimedia/wikipedia; their respective licenses apply to the underlying texts.

The annotated NER dataset used for fine-tuning was provided by Brand Analytics.