PyTorch code for MultiHead:"Joint entity recognition and relation extraction as a multi-head selection problem".I use Chinese dataset to achieve chinese entity relation extraction. I use albert as feature extractor to improve results. For a description of the model and experiment, see paper https://arxiv.org/abs/1804.07847.
- torch==1.4.0+cu100
- cuda=10.0
- cudnn=7603
- pytorch-crf==0.7.2
- transformers==4.3.3
- tqdm==4.59.0
- seqeval==0.0.10
- tensorboard
- flask == 1.1.2
I use flask as the back-end framework, Neo4j as the graph database.
Baidu CCKS2019 Competition
The dataset of train_data.json, dev_data.json and predict.json mentioned in the project have been upload in the compressed package data.zip in directory data/
(The dataset is a little different from I used in the project, please use the data/data.zip)Download link: https://ai.baidu.com/broad/download?dataset=dureader
- checkpoint save the model regularly
- data save the data
- data_loader the code to save data process
- deploy demo code
- deploy_flask use flask web framework to deploy model
- doc save description files
- mains training code
- models save the model files
- modules the code of model
- pretrained save bert and albert pretrained files
- record save files about tensorboard
- test save test files while developing
- utils tool files
cd mains
python3 trainer_std.py -encode=albert
cd deploy
python3 demo.py -encode=albert
(please use albert33m-p0.77f0.77n2.98r2.17.pth or
43m-p0.89f0.91n0.94r0.43.pth)
If you want to use the model of gru, please train it yourself,
because I can't find the model file now.
In fact, albert performsbetter than gru.
cd deploy_flask
python3 manage.py
Results:
Input text and output relation extraction results.
Input the name of the entity and output all the knowledge related to the entity.
Input the name of the relationship and output all the knowledge with the relationship.
