GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition

The pytorch implementation of "GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition " (AAAI 2025).

The framework of GuideNER is shown in the following figure:

Enviroment

We recommend the following actions to create the environment:

conda create -n  GuideNER python==3.9.19
conda activate GuideNER
pip install jinja2==3.1.4
pip install transformers==4.43.3
pip install vllm==0.5.3.post1
pip install tokenizers ==0.19.1

Datasets

Due to licensing restrictions, we can only provide the CoNLL03 dataset along with related prompts. Both the original and the processed datasets are placed in the datasets folder. The processed data is in JSONL format, and each entry contains the keys "text" and "entity_labels", as shown below:

{"text": "EU rejects German call to boycott British lamb .", "entity_labels": [["EU", "organization"], ["German", "miscellaneous"], ["British", "miscellaneous"]]}

Running

First, you need to run python rule_summary.py to summarize the guidelines corresponding to the dataset. Then, execute python run_withrule.py to perform inference and obtain the results.