/ Natural Language Processing

A toolkit that provides a off-the-shelf framework to implement lots of NLP extraction tasks

A toolkit that provides a off-the-shelf framework to implement lots of NLP extraction tasks


OpenUE allows users ranging from beginner python coders to experienced machine learning engineers to leverage lots of NLP extraction tasks in one easy-to-use python package.

Key Features

  • Full Guides and API Documentation

  • Unified API for NLP Tasks with SOTA Pretrained Models (Adaptable with BERT, XLNet, etc.)

    • Entity and Realation Extraction
    • Intent and Slot Filling
    • Opinion and Apspect Extraction
    • More in development
  • Training and Inference Interface

  • Rapid NLP Model Deployment

  • Dockerizing OpenUE with GPUs

    • Easily build and run OpenUE containers leveraging NVIDIA GPUs with Docker


Quick Start

Requirements and Installation

Anaconda Environment
conda create -n openue python=3.6
conda activate openue
conda install  --file requirements.txt 

Examples and General Use

Once you have installed OpenUE, here are a few examples of what you can run with OpenUE modules:

Entity and Relation Extraction Example
  1. Data Preprocessing. Put the pretrined language model (e.g., BERT) in the pretrained_model folder and put all raw data (run script download_ske.sh in the benchmark folder) in the raw_data folder, run
sh download_ske_dataset.sh
sh download_pretrain_cn_bert.sh
sh preprocess.sh  ske
  1. Train Sequence Labeling & Classification Model. Set all parameters in the file config.py and run
sh train_seq.sh ske
sh train_class.sh ske

You can download the checkpoint here, extract and put them in the output folder.

  1. Test & Evaluation. Run
python predict.sh ske
  1. Export & Serving. Run
sh export_seq.sh ske
sh serving_cls.sh ske
sh serving.sh
  1. Interactive Prediction. Run
python  predict_online.py
  1. Demo.Run
python app.py  ske


>>> import openuee
>>> model = openue.get_model('ske_bert_entity_relation')
>>> res = model.infer('《上海滩》是刘德华的音乐作品,黄沾作曲,收录在《【歌单】酷我热门单曲合辑》专辑中')
>>> print(res)
"spo_list": [{"object_type": "人物", "predicate": "作曲", "object": "黄沾", "subject_type": "歌曲", "subject": "上海滩"}, {"object_type": "音乐专辑", "predicate": "所属专辑", "object": "【歌单】酷我热门单曲合辑", "subject_type": "歌曲", "subject": "上海滩"}, {"object_type": "人物", "predicate": "歌手", "object": "刘德华", "subject_type": "歌曲", "subject": "上海滩"}]

Note that it may take a few minutes to download checkpoint and data for the first time. Then use infer to do sentence-level entity and relation extraction

How to Cite

If you use or extend our work, please cite the following paper:

    title = "{O}pe{UE}: An Open Toolkit for Universal  Extraction in Text",
    author = "Ningyu Zhang, Shumin Deng, Huajun Chen",
    year = "2020",