CurriculumDocRE: A Curriculum Learning Approach for DocRE

This repository contains an implementation of the CurriculumDocRE approach for document-level relation extraction with curriculum learning.

Overview

The CurriculumDocRE approach implements a curriculum learning strategy for document-level relation extraction, where the model progressively learns from:

1. Simple relations (entities in same/adjacent sentences)
2. Multi-hop relations (entities in different sentences requiring connecting information)
3. Complex relations (entities across paragraphs requiring advanced coreference) -- More explaination requires

Key Features

1. Curriculum-Aware Attention:

   • Dynamic attention patterns based on curriculum stage
   • Entity-focused attention mechanisms
   • Stage-specific adapters for different relation complexities

2. Data Quality Enhancement:

   • Annotation consistency checking
   • Confidence scoring based on multiple factors
   • Active learning for uncertain instances

3. Data Augmentation:

   • Entity substitution with contextual awareness
   • Evidence masking for improved robustness
   • Cross-document relation transfer

4. Optimized Training Pipeline:

   • Integrated components in a cohesive flow
   • Curriculum-aware loss functions
   • Confidence-weighted evaluation metrics

5. Optimized Evaluation:

   • Stage-specific thresholds for improved Ing F1 score
   • Entity type compatibility checking
   • Confidence-based filtering and ranking

Requirements

torch>=1.10.0
transformers>=4.18.0
numpy>=1.20.0
scikit-learn>=1.0.0
tqdm>=4.62.0

Install with: pip install -r requirements.txt

Directory Structure

• src/: Source code
  • model.py: Enhanced model with curriculum-aware attention
  • data_quality.py: Data quality enhancement module
  • data_augmentation.py: Data augmentation techniques
  • data_loader.py: Data loading and preprocessing
  • train.py: Training pipeline with curriculum learning
  • evaluate.py: Memory-efficient evaluation script
  • config.py: Configuration parameters
  • utils.py: Utility functions
• data/: DocRED dataset files
  • train_annotated.json: Training data
  • dev.json: Development data
  • test.json: Test data
  • rel_info.json: Relation type information

Before running the code:

1. Download the data from the following link
2. Save it in the data folder.

Running the Code

Training

python -m src.train

This will:

1. Enhance data quality (annotation consistency, confidence scoring)
2. Apply data augmentation (entity substitution, evidence masking, relation transfer)
3. Train the model using curriculum learning (3 stages)

Evaluation

python -m src.evaluate

This will generate predictions in output/results/predictions.json.

Official Evaluation

To evaluate using the official DocRED metrics:

1. Clone the DocRED repository:

git clone https://github.com/thunlp/DocRED.git

2. Run the evaluation script:

cd DocRED/code
python eval.py -g ../../data/dev.json -p ../../output/results/predictions.json