Source code of the Decouple Processed Continual Relation Extraction (DP-CRE) [Decouple Processed to Balance Continual Relation Extraction and Preserve Memory Structure].
Continuous Relation Extraction (CRE) aims to incrementally learn relation knowledge from a non-stationary stream of data. One significant challenge in this domain is catastrophic forgetting, where the introduction of new relational tasks can overshadow previously learned information. Unlike traditional replay-based training paradigms that uniformly prioritize all data, we decouple the process of prior information preservation and new knowledge acquisition. In this paper, we introduce the Decouple Processed CRE (DP-CRE) framework. This approach examines alterations in the embedding space as new relation classes emerge, distinctly managing the preservation and acquisition of knowledge. Extensive experiments show that DP-CRE significantly outperforms other CRE baselines across two datasets.
- Python 3.7.16
- PyTorch: 1.13.1(cuda version 11.7).
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
To install PyTorch, you could follow the official guidance of PyTorch.
Then, other dependencies could be installed by running:
pip install -r requirements.txt
Pre-trained BERT weights:
- Download bert-base-uncased into the root directory [google drive].
We use FewRel and TACRED datasets in our experiments.
- FewRel:
data/data_with_marker.json - TACRED:
data/data_with_marker_tacred.json
The splited datasets and task orders is conducted in sample.py.
To reproduce the results of main experiment:
python main.py --task_name FewRel
python main.py --task_name TACRED