DynamicNER: A Dynamic, Multilingual, and Fine-Grained Dataset for LLM-based Named Entity Recognition

This repository is supplement material for the paper: DynamicNER: A Dynamic, Multilingual, and Fine-Grained Dataset for LLM-based Named Entity Recognition. This is accpeted by EMNLP 2025 Main Conference.

💓Update!

• Our paper is accpeted by EMNLP 2025!!!

• We add more existing open-source datasets in our format and also the format for fine-tuning and inferrence based on SWIFT! You can test CascadeNER easier!

• We discover a problem that as SWIFT has been updated and some parameters has been changed, so please use the old version (according to requirements.txt).

• We provide demo.py for testing CascadeNER easier.

📚 Features

• This repository includes DynamicNER and CascadeNER, our NER dataset and framework.

• DynamicNER is the first dataset specially designed for NER with LLMs with a novel dynamic categorization system. It's multilingual and fine-grained.

• CascadeNER is the first universal and multilingual NER framework with SLMs, which supports both few-shot and zero-shot scenarios and achieves SOTA performance on low-resource and fine-grained datasets

📈 Quantitive Result:

📌 Prerequisites

1. conda create -n dynamicner python=3.10
2. pip install -r requirements.txt
3. You may also use a standard environment for SWIFT.
4. You may also download ./DynamicNER.7z and unzip it to obtain the dataset for training.

🌟 Usage

• Dataset preparation: the DynamicNER_process directory contains the scripts for generating dynamic datasets, running format conversions, and validating labels. See DynamicNER_process/readme.md for the full checklist (including check_dynamic_classify.py, prune_dynamic_classify.py, and sync_extract.py).

• Training data: please use SWIFT for model training. We strongly recommend Qwen series for your base models. You may follow the examples in any train.json from BASE-format datasets on Hugging Face to understand the training layout. A concrete example is provided in ./DynamicNER_process/example.json.

• CascadeNER inference: Stage-1 extraction, Stage-2 classification, and evaluation are documented in ./CascadeNER/README.md. Configure paths via CLI arguments/environment variables rather than editing the scripts directly. Typical usage involves running extract.sh, then python model/infer.py, and finally python evaluate.py.

• Transformation utilities: to transform your own corpora into SWIFT/BIO formats, use ./DynamicNER_process/transformation/stage1_trans.py and stage2_trans.py; BIO exports are handled by BIO_trans.py (and BIO_trans_zh.py).

• PS: Due to the update of SWIFT, you may need to use the old version to directly use our code, or you can modify the code slightly with the guidance from SWIFT. We will later provide a updated version of code for this problem.

❤️ Acknowledgement

• We thank QwenLM for opening source their Qwen model for us
• We thank ModelScope for opening source their SWIFT framework for us
• We thank teams of CoNLL2003, CrossNER, FewNERD, MultiCoNER and PAN-X for opening source their datasets