/HingeABL

Codes for paper "Adaptive Hinge Balance Loss for Document-Level Relation Extraction", Findings of EMNLP 2023

Primary LanguagePythonMIT LicenseMIT

HingeABL

Codes for paper "Adaptive Hinge Balance Loss for Document-Level Relation Extraction", Findings of EMNLP 2023.

Requirements

  • Python (tested on 3.7.4)
  • CUDA (tested on 11.3)
  • PyTorch (tested on 1.12.1)
  • Transformers (tested on 4.20.1)
  • numpy (tested on 1.21.6)
  • spacy (tested on 3.3.3)
  • opt-einsum (tested on 3.3.0)
  • ujson
  • tqdm

Dataset

The Re-DocRED dataset can be downloaded following the instructions at link.

The expected structure of files is:

HingeABL
 |-- dataset
 |    |-- docred
 |    |    |-- train_revised.json        
 |    |    |-- dev_revised.json
 |    |    |-- test_revised.json
 |    |    |-- train_annotated.json  
 |-- meta
 |    |-- rel2id.json   
 |-- scripts
 |-- checkpoint
 |-- log

Note: train_annotated.json, rel2id.json can be obtained through the DocRED dataset, which can be downloaded following the instructions at link.

Training and Evaluation

Train the BERT model with HingeABL using the following command:

>> sh scripts/train_HingeABL.sh  # training
>> sh scripts/test_HingeABL.sh  # evaluation

You can select different loss functions by setting the --loss_type argument before training. Optional loss types include：ATL, balance_softmax, AFL, SAT, MeanSAT, HingeABL, AML.

Note: This code is partially based on the code of ATLOP.