PromptNER

Prompting For Named Entity Recognition

This is the home of the PromptNER tool, currently anonymized for conference submission purposes.

Installation

Very little is needed by way of installation, simply clone and install the requirements (we use Python 3.10.9 but many other versions should work without a problem)

You will also need to download nltk stopwords:

cd PromptNER
pip install -r requirements.txt
python -m nltk.downloader stopwords

You will have to set up an OpenAI account to use the GPT based Prompting models. Make sure to save the API key as an environment variable: (put this line in your ~/.bashrc file if you are on Ubuntu to avoid doing this everytime)

export OPENAI_API_KEY='YOUR API KEY HERE'

Running PromptNER on your own dataset or task

All PromptNER needs to perform Named Entity Recognition on a completely unseen task specification is

• A definition of the concept of a named entity, written in natural language
• A few examples (3-5 is usually good) of the task being done succesfully

Head on to the custom file and follow the ExampleConfig class to add in your desired specifications in the CustomClass, after this you should be able to import the 'get_ner_system' function and query it to get a function which takes in a sentence in text and outputs a list of extracted entities.

Reproducing Experiments

To reproduce the experiments in the paper visit the run file, you can adapt the main function to call either 'run_all_datasets' or 'ablate_best' to get all the results from the paper.

You will have to download the data from the drive link and extract it at root level of the repository.