ITER: Iterative Transformer-based Entity Recognition and Relation Extraction

This repository contains the source code for our paper ITER, accepted at EMNLP 2024.

Setup

To set up the repository, the following basic steps are required:

1. Install ITER

python3 -m venv venv && source venv/bin/activate  # optional
pip install git+https://github.com/fleonce/iter

2. Download the required datasets

bash scripts/datasets/load_datasets.sh

3. Run the training script

python3 train.py --transformer t5-small --dataset {ace05,ade,conll03,conll04,genia,scierc}

where the transformer and dataset arguments have the following possible values:

Currently working transformer models are:

- t5-{small,base,large,3b,11b}
- google/t5-v1_1-{small,base,large,xl,xxl}
- google/flan-t5-{small,base,large,xl,xll}
- bert-large-cased
- microsoft/deberta-v3-{xsmall,small,base,large}
- microsoft/deberta-v2-{xlarge,xxlarge}

Currently supported datasets are:

- ace05
- ade
- conll03
- conll04
- genia
- scierc

4. Evaluating models

To evaluate the checkpoints we provided, simply use the following command:

python3 evaluate.py --model {checkpoint}

5. Model Checkpoints on Huggingface

Checkpoints

We publish checkpoints for the models performing best on the following datasets:

• ACE05:
  1. fleonce/iter-ace05-deberta-large
• CoNLL04:
  1. fleonce/iter-conll04-deberta-large
• ADE:
  1. fleonce/iter-ade-deberta-large
• SciERC:
  1. fleonce/iter-scierc-deberta-large
  2. fleonce/iter-scierc-scideberta-full
• CoNLL03:
  1. fleonce/iter-conll03-deberta-large
• GENIA:
  1. fleonce/iter-genia-deberta-large