PromptNER

Code for PromptNER: Prompt Locating and Typing for Named Entity Recognition, accepted at ACL 2023.

The left figure show a comparison of the type-oriented (a) and span-oriented (b) prompt learning with the proposed PromptNER (c). C, N and M denote the number of entity types, words and prompts, respectively; The right figure shows an overview of PromptNER. The left part describes the model’s inference process and the right part describes the dynamic template filling mechanism during training.

Setup

conda create -n promptner python=3.8
conda activate promptner
pip install -r requirements.txt

Usage

This repository is under construction. We will further detail the usage of the code and the data.

Data

For the data, please refer to data/README.md.

Training

Supervised Setting:

PromptNER[BERT] on ACE2004:

python prompt4ner.py train --config configs/bert/ace2004.conf

PromptNER[RoBERTa] on ACE2005:

python prompt4ner.py train --config configs/roberta/ace2005.conf

Low-resource Setting:

Wiki Warmup Pretrain:

python prompt4ner.py train --config configs/few_shot/wiki_pretrain/pretrain.conf

In-domain Setting:

python prompt4ner.py train --config configs/few_shot/in_domain/conll2003_sampling.conf

Cross-domain Setting:

First train PromptNER on source domain (CoNLL 2003) based on the pre-trained model from Wiki Warmup Pretrain:

python prompt4ner.py train  --config configs/few_shot/cross_domain/source.conf

Then train PromptNER on target domain (MIT Movie) using the model from the previous step:

python prompt4ner.py train --config configs/few_shot/cross_domain/mm.conf

Evaluation

For evaluation, please first replace the model_path@configs/eval.conf with the path of the trained model. Then run: