Skip to content

NLP关系抽取:序列标注、层叠式指针网络、Multi-head Selection、Deep Biaffine Attention

Notifications You must be signed in to change notification settings

zhoujx4/NLP-Series-relation-extraction

Repository files navigation

项目说明:

本项目是关系抽取相关模型的代码复现
包括以下四种方法

  • 序列标注
  • 层叠式指针网络(基于主语感知)
  • Multi-head Selection
  • Deep Biaffine Attention

用的数据是百度21年语言技术经验竞赛抽取赛道的数据,四种方法的效果如下表,更详细的请看我的知乎博文 https://zhuanlan.zhihu.com/p/381894616

F1值
官方baseline 64.69
层叠式指针网络(基于主语感知) 61.22
Multi-head Selection 67.90
Deep Biaffine Attention 68.45

环境

  • python=3.6
  • torch=1.7
  • transformers=4.5.0

运行示例

序列标注

python3 run_baseline.py
--max_len=200
--model_name_or_path=预训练模型路径
--per_gpu_train_batch_size=80
--per_gpu_eval_batch_size=100
--learning_rate=1e-4
--num_train_epochs=40
--output_dir="./output"
--weight_decay=0.01
--early_stop=2

层叠式指针网络(基于主语感知)

python3 run_mpn.py
--max_len=200
--model_name_or_path=预训练模型路径
--per_gpu_train_batch_size=100
--per_gpu_eval_batch_size=100
--learning_rate=1e-4
--num_train_epochs=40
--output_dir="./output"
--weight_decay=0.01
--early_stop=2

Multi-head Selection

python3 run_mhs.py
--max_len=200
--model_name_or_path=/data/zhoujx/prev_trained_model/rbt3
--per_gpu_train_batch_size=25
--per_gpu_eval_batch_size=30
--learning_rate=1e-4
--num_train_epochs=40
--output_dir="./output"
--weight_decay=0.01
--early_stop=2

Deep Biaffine Attention

python3 run_mhs_biaffine.py
--max_len=200
--model_name_or_path=/data/zhoujx/prev_trained_model/rbt3
--per_gpu_train_batch_size=15
--per_gpu_eval_batch_size=20
--learning_rate=1e-4
--num_train_epochs=40
--output_dir="./output"
--weight_decay=0.01
--early_stop=2
--overwrite_cache=True

About

NLP关系抽取:序列标注、层叠式指针网络、Multi-head Selection、Deep Biaffine Attention

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages