MLWA-Chinese-SLU

This repository contains the PyTorch implementation of the paper:

Injecting Word Information with Multi-Level Word Adapter for Chinese Spoken Language Understanding. Dechuan Teng, Libo Qin, Wanxiang Che, Sendong Zhao, Ting Liu. ICASSP 2021.

Architecture

Requirements

Our code is based on Python 3.7.6 and PyTorch 1.1. Requirements are listed as follows:

• torch==1.1.0
• transformers==2.4.1
• numpy==1.18.1
• tqdm==4.42.1
• seqeval==0.0.12

We highly suggest you using Anaconda to manage your python environment.

How to Run it

Quick start

The script train.py acts as a main function to the project, you can run the experiments by replacing the unspecified options in the following command with the corresponding values.

    CUDA_VISIBLE_DEVICES=$1 python train.py -dd ${dataDir} -sd ${saveDir} -u -bs 16 -dr 0.3 \ 
        -ced 128 -wed 128 -ehd 512 -aod 128 -sed 32 -sdhd 64

Ablation Study

For reproducing the results reported in our paper, we suggest you executing the following shell command:

    ./ablation_study.sh ${CUDA_VISIBLE_DEVICES} ${dataset_name} ${model_type} ${bies} ${tokenizer}
    
    # Examples for CAIS dataset
    ./ablation_study.sh "0" "cais" "full" "True" "ltp"
    ./ablation_study.sh "0" "cais" "none" "True" "ltp"
    ./ablation_study.sh "0" "cais" "full" "True" "pkuseg"
    ......

Refer to ablation_study.sh for more details.

Due to some stochastic factors, It's necessary to slightly tune the hyper-parameters using grid search. If you have any question, please issue the project or email me and we will reply you soon.

Pre-trained Model

You can use the -use_bert parameter in train.py to replace the encoder of our model with BERT-Base-Chinese module. And the -phed parameter controls the percentage between the output dim of word-channel BiLSTM encoder and that of word-channel Self-Attention encoder. You can write a command like this:

    CUDA_VISIBLE_DEVICES=$1 python train.py -dd ${dataDir} -sd ${saveDir} -u \ 
        -use_bert -pehd 0.6 -ne 50 -lr 3e-5 -bs 16 -dr 0.3 -wed 128 -sed 32 -sdhd 64

Citation

If you use any source codes or the datasets included in this repo in your work, please cite the following paper. The bibtex are listed below:

@article{teng2020injecting,
  title={Injecting Word Information with Multi-Level Word Adapter for Chinese Spoken Language Understanding},
  author={Teng, Dechuang and Qin, Libo and Che, Wanxiang and Zhao, Sendong and Liu, Ting},
  journal={arXiv preprint arXiv:2010.03903},
  year={2020}
}

Issues/PRs/Questions

Feel free to contact me for any question or create issues/prs.

Acknowledgement

https://github.com/yizhen20133868/StackPropagation-SLU

We are very grateful for their pytorch implementation of Stack-Propagation Framework for Spoken Language Understanding.