/HiNet

Source code for paper: HiNet: Novel Multi-Scenario & Multi-Task Learning with Hierarchical Information Extraction

Primary LanguagePythonMIT LicenseMIT

Introduction

Hierarchical information extraction Network (HiNet)

Source code for paper: HiNet: Novel Multi-Scenario & Multi-Task Learning with Hierarchical Information Extraction

Model architecture:

avatar

Bibtex:

@article{Zhou2023HiNet,
  title={HiNet: Novel Multi-Scenario & Multi-Task Learning with Hierarchical Information Extraction},
  author={Zhou, Jie and Cao, Xianshuai and Li, Wenhao and Bo, Lin and Luo, Chuan and Yu, Qian},
  year={2023},
}

Requirements

Python >= 3.7  
Tensorflow >= 1.15.0

Train and Evaluate Model

python -u HiNet.py --config_file_path=../config/hinet_sample_schemas.json --task_type=train

You can modify the "FLAGS" parameters as needed.

Acknowledgement

  • The work is supported by MeiTuan.
  • The work is also supported by the National Natural Science Foundation of China (No. 62202025 and 62002012).

Model Gain Statistics

As of the latest time, HiNet model was deployed to several companies' business lines, and according to incomplete statistics, the business benefits of each company are as follows:

Platform Gain
Meituan catering Relative order volume uplift of 1%-5% across multiple businesses.
OPPO Advertising 1% increase in CTR.
AutoNavi Search 10% increase in CVR.
... ...