RAPL: Relation-Aware Prototype Learning for Few-Shot DocRE

This repository contains the data, code and trained models for paper RAPL: A Relation-Aware Prototype Learning Approach for Few-Shot Document-Level Relation Extraction.

Quick Links

• Overview
• Setup
  • Install dependencies
  • Trained models
• Datasets
  • FREDo
  • ReFREDo
• Quick Start
• Citation

Overview

In this work, we present a relation-aware prototype learning method (RAPL) for few-shot document-level relation extraction. We reframe the construction of relation prototypes into instance level and further propose a relation-weighted contrastive learning method to jointly refine the relation prototypes. We also design a task-specific NOTA prototype generation strategy to better capture the NOTA semantics in each task.

You can find more details of this work in our paper.

Setup

Install dependencies

To run the code, please install the following dependency packages:

• apex (experiment on 0.1)
• numpy (experiment on 1.19.4)
• opt_einsum (experiment on 3.3.0)
• torch (experiment on 1.9.0+cu111)
• tqdm (experiment on 4.64.0)
• transformers (experiment on 3.4.0)
• wandb (experiment on 0.12.21)

Trained models

We release the sample trained models for each task setting on Tsinghua Cloud. To reproduce the results in the paper, you can download the corresponding models and place them in checkpoints directory.

Datasets

Our experiments are based on two benchmarks: FREDo and ReFREDo, and all relevant data files are located in dataset directory.

FREDo

FREDo is a few-shot document-level relation extraction benchmark consisting of two main tasks (in-domain / cross-domain) with a 1-Doc and a 3-Doc subtask each. The relevant data files include:

• dataset/[train_docred, dev_docred, test_docred, test_scierc].json contain all annotated documents used for training and testing.
• dataset/[test_docred_1_doc_indices, test_docred_3_doc_indices, test_scierc_1_doc_indices, test_scierc_3_doc_indices].json contain sampled episodes (only the indices of the documents and which relations are to be annotated/extracted).

ReFREDo

ReFREDo is a revised version of FREDo, which replaces the training, development and in-domain test document corpus with Re-DocRED, leading to more complete annotations. The relevant data files include:

• dataset/[train_redocred, dev_redocred, test_redocred, test_scierc].json contain all annotated documents used for training and testing.
• dataset/[test_redocred_1_doc_indices, test_redocred_3_doc_indices, test_scierc_1_doc_indices, test_scierc_3_doc_indices].json contain sampled episodes (only the indices of the documents and which relations are to be annotated/extracted).

Quick Start

In scripts directory, we provide the example scripts for running experiments under each task setting. For example, you can use the following command to run the trained model on in-domain 3-Doc test tasks in ReFREDo benchmark:

sh scripts/refredo_indomain_3doc.sh

You can also comment the --load_checkpoint argument and set --num_epochs argument to 25 for training. The following command can be used to display the details about each argument:

python src/main.py -h

Citation

Please kindly cite our paper if you use the data, code or models of RAPL in your work:

@inproceedings{meng-etal-2023-rapl,
    title = "{RAPL}: A Relation-Aware Prototype Learning Approach for Few-Shot Document-Level Relation Extraction",
    author = "Meng, Shiao  and
      Hu, Xuming  and
      Liu, Aiwei  and
      Li, Shuang  and
      Ma, Fukun  and
      Yang, Yawen  and
      Wen, Lijie",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    year = "2023",
    url = "https://aclanthology.org/2023.emnlp-main.316",
    doi = "10.18653/v1/2023.emnlp-main.316",
    pages = "5208--5226"
}