Named Entity Recognition
Named Entity Recognition (NER) is the task of extracting named entities such as person names, place names, and organization names from text.
6.1 Rule-based and Dictionary-based Methods
pygeonlp (geonlp-platform/pygeonlp): A Python module specialized for geotagging Japanese text (recognizing place names and assigning location information). Stars: 22, weekly downloads: 171.
entitypedia (chakki-works/entitypedia): A proper noun dictionary extracted from Wikipedia. Stars: 13.
6.2 BERT-based Named Entity Recognition
| Tool | Stars | Description |
|---|---|---|
| bert-ner-japanese | 5 | NER by BERT fine-tuning |
| bert-japanese-ner-finetuning | 11 | Sample for BERT NER fine-tuning |
| namaco | 40 | Character-based NER |
| huggingface-finetune-japanese | 16 | Encoder-decoder fine-tuning |
namaco is a character-based named entity recognition tool by chakki-works, with Stars: 40 and high name recognition in the NER category.
6.3 NER for Medical Information
MedNER-J (sociocom/MedNER-J): A Japanese disease name extraction tool. Stars: 18. Specialized in information extraction from medical texts.
JaMIE (racerandom/JaMIE): A Japanese medical information extraction toolkit. Stars: 9.
6.4 Geographic Information
jageocoder (t-sagara/jageocoder): A Pure Python Japanese address geocoder. Stars: 109, weekly downloads: 8k.
Japanese NER Models on Hugging Face
| Model | HF DL | Likes | Description |
|---|---|---|---|
| bert-base-japanese-v3-ner-wikipedia-dataset (llm-book) | 8k | 11 | Fine-tuned on Wikipedia NER data |
| bert-ner-japanese (jurabi) | 5k | 11 | Supports 8 entity types |
| xlm-roberta-ner-japanese (tsmatz) | 4k | 28 | XLM-RoBERTa based |
bert-ner-japanese (jurabi) records 5k/week downloads, second in the token-classification category after llm-book's Wikipedia NER model (8k/week). Built on cl-tohoku/bert-base-japanese-v2 and trained on the Stockmark Wikipedia dataset, it extracts eight entity types including companies, organizations, facilities, products, and events.
yomi-linter-modernbert-ja-130m fine-tunes ModernBERT-Ja-130m to detect spans that are likely to be mis-read by G2P/dictionary-based TTS systems such as OpenJTalk, piper-plus, and VOICEVOX (recall of roughly 80%, with dynamic INT8 quantization support). It illustrates how NER is expanding into speech-synthesis preprocessing.