English|简体中文
What is Chinese-LS?
Lexical simplification (LS) aims to replace complex words in a given sentence with their simpler alternatives of equivalent meaning. Chinese-LS is the first attempt in the field of Chinese Lexical Simplification. It includes a high-quality benchmark dataset and five baseline approaches:
-
Synonym dictionary-based approach
-
Word embedding-based approach
-
Pretrained language model-based approach
-
Sememe-based approach
-
Hybrid approach
The entire framework of Chinese-LS is shown below:
Quick start
Requirements
- Python==3.7.6
- transformers==3.5.0
- numpy==1.18.1
- jieba==0.42.1
- torch==1.4.0
- OpenHowNet==0.0.1a11
- gensim==3.8.2
You can find the complete requirements here.
Preparations
Download Pretrained Models
Chinese-LS uses the following pretrained models:
- Word2Vec model: Chinese-Word-Vector (Mixed-large)
- BERT-base, Chinese (huggingface)
- macbert-base, Chinese (huggingface)
- bert-wwm-ext, Chinese (huggingface)
- roberta-wwm-ext, Chinese (huggingface)
- ERNIE (PaddlePaddle)
Please place the models under the ./model
directory after downloading.
Run
We have already executed the codes for you and intermediate results can be found in ./data
.
You could check out the details of codes and algorithms from our paper: Chinese Lexical Simplification
If you want to run the codes for reproduction, please execute them in the following order:
Generate
-
Synonym dictionary based-approach
Run
dict_generate.py
-
Word embedding based-approach
Run
vector_generate.py
-
Pretrained language model based-approach
Run
bert_generate.sh
-
Sememe based-approach
Run
hownet_generate.py
-
Hybrid approach
Run
hybrid_approach.py
Select
Run substitute_selection.py
Rank
Run substitute_ranking.py
Experiments
Chinese-LS designs 5 experiments to evaluate the quality of our dataset and the performance of five approaches. You could get the experiment results through running experiment.py
.
Citation
@article{qiang2021chinese,
title={Chinese Lexical Simplification},
author={Qiang, Jipeng and Lu, Xinyu and Li, Yun and Yuan, Yun-Hao and Wu, Xindong},
journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
year={2021},
volume={29},
pages={1819-1828},
doi={10.1109/TASLP.2021.3078361},
publisher={IEEE}
}
Contact
This repo may still contain bugs and we are working on improving the reproductivity. Welcome to open an issue or submit a Pull Request to report/fix the bugs.
Email: luxinyu12345@foxmail.com
License
Chinese-LS is under the Apache License, Version 2.0.