SpanBERT + MSA

Code for NAACL 2022 paper Modeling Multi-Granularity Hierarchical Features for Relation Extraction.

Requirements

transformers==3.0.2

python==3.7

pytorch==1.3.1

Data and Pretrained model

TACRED dataset is not open access, so we provide the cached file for the training and testing of our model. (If needed, you can get the raw data from https://catalog.ldc.upenn.edu/LDC2018T24. The TACRED Revisted dataset can be obtained by https://github.com/DFKI-NLP/tacrev.)

We use SpanBert from https://huggingface.co/SpanBERT/spanbert-large-cased.

TODO: We will released our trained model soon.

Evaluation

We provided log file of evaluation in logs. You can evaluate our model by run:

chmod +x eval_tacred.sh; ./eval_tacred.sh
chmod +x eval_tacred_rev.sh; ./eval_tacred_rev.sh

Training

We train our model on single NVIDIA V100 GPU about 4 hours. You can train a new model by:

chmod +x run_spanbert.sh; ./run_spanbert.sh