This repo contains the example pytorch code for the paper Distilling Knowledge from Well-informed Soft Labels for Neural Relation Extraction.
This paper/code aims to exhibitions the supervision with soft labels in relation extraction, which makes it possible to integrate prior knowledge. Specifically, a bipartite graph is first devised to discover type constraints between entities and relations based on the entire corpus. Then, we combine such type constraints with neural networks to achieve a knowledgeable model. Furthermore, this model is regarded as teacher to generate well-informed soft labels and guide the optimization of a student network via knowledge distillation. Besides, a multi-aspect attention mechanism is introduced to help student mine latent information from text. In this way, the enhanced student inherits the dark knowledge (e.g., type constraints and relevance among relations) from teacher, and directly serves the testing scenarios without any extra constraints.
See below for an overview of the model architecture:
- python3 (tested on 3.7.0)
- pytorch (tested on 1.0.1)
- CUDA
- tqdm
- unzip, wget
The code requires that you have access to the TACRED dataset (LDC license required). Once you have the TACRED data, please put the JSON files under the directory tree_nn/dataset/tacred. For completeness, we only include sample data files from the TACRED dataset in this repo.
First, construct the bipartite graph (this step can be omitted since we have uploaded the processed file tree_nn/dataset/pattern/type2prob.json):
cd bipartite
python3 construct.py
cd ../
Then, download and unzip GloVe vectors:
cd tree_nn
chmod +x download.sh; ./download.sh
Next, prepare vocabulary and initial word vectors:
python3 prepare_vocab.py dataset/tacred dataset/vocab --glove_dir dataset/glove
This will write vocabulary and word vectors as a numpy matrix into the dir tree_nn/dataset/vocab.
First, train a graph convolutional neural network model as base network (we use C-GCN here):
python3 train_base.py --id base
Model checkpoints and logs will be saved to tree_nn/saved_models/base.
Then, train teacher network with a relation-related weights:
python3 train_teacher.py --base_id base --base_mode 2 --id teacher
Model checkpoints and logs will be saved to tree_nn/saved_models/teacher.
Next, train student network with the guidance of teacher network:
python3 train_student.py --teacher_id teacher --base_mode 0 --id student
Model checkpoints and logs will be saved to tree_nn/saved_models/student.
For simplicity, these models use the same set of parameters in this repo, which could also be adjusted to be optimal respectively. For details on the use of other parameters, please refer to tree_nn/config.py.
To evaluate the base network on test set:
python3 eval_base.py base --dataset test
To evaluate the teacher network on test set:
python3 eval_teacher.py teacher --dataset test
To evaluate the student network on test set:
python3 eval_student.py student --dataset test
They will use the best_model.pt file by default. Use --model checkpoint_epoch_xx.pt to specify a model checkpoint file of the xx epoch.
@inproceedings{zhang2020distilling,
author = {Zhang, Zhenyu and Shu, Xiaobo and Yu, Bowen and Liu, Tingwen and Zhao, Jiapeng and Li, Quangang and Guo, Li},
title = {Distilling Knowledge from Well-informed Soft Labels for Neural Relation Extraction},
booktitle = {Proc. of AAAI},
year = {2020}
}
Codes are adapted from the repo of C-GCN: Graph Convolution over Pruned Dependency Trees Improves Relation Extraction.