WMSeg

This is the implementation of Improving Chinese Word Segmentation with Wordhood Memory Networks at ACL2020.

Please contact Yuanhe Tian at yhtian@uw.edu if you have any questions.

Visit our homepage to find more our recent research and softwares for NLP (e.g., pre-trained LM, POS tagging, NER, sentiment analysis, relation extraction, datasets, etc.).

Upgrades of WMSeg

We are improving our WMSeg. For updates, please visit HERE.

Citation

If you use or extend our work, please cite our paper at ACL2020.

@inproceedings{tian-etal-2020-improving,
    title = "Improving Chinese Word Segmentation with Wordhood Memory Networks",
    author = "Tian, Yuanhe and Song, Yan and Xia, Fei and Zhang, Tong and Wang, Yonggang",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    pages = "8274--8285",
}

Requirements

Our code works with the following environment.

python=3.6
pytorch=1.1

Downloading BERT, ZEN and WMSeg

In our paper, we use BERT (paper) and ZEN (paper) as the encoder.

For BERT, please download pre-trained BERT-Base Chinese from Google or from HuggingFace. If you download it from Google, you need to convert the model from TensorFlow version to PyTorch version.

For ZEN, you can download the pre-trained model from here.

For WMSeg, you can download the models we trained in our experiments from here.

Run on Sample Data

Run run_sample.sh to train a model on the small sample data under the sample_data directory.

Datasets

We use SIGHAN2005 and CTB6 in our paper.

To obtain and pre-process the data, please go to data_preprocessing directory and run getdata.sh. This script will download and process the official data from SIGHAN2005. For CTB6, you need to obtain the official data first, and then put the LDC07T36 folder under the data_preprocessing directory.

All processed data will appear in data directory.

Training and Testing

You can find the command lines to train and test models on a specific dataset in run.sh.

Here are some important parameters:

--do_train: train the model.
--do_test: test the model.
--use_bert: use BERT as encoder.
--use_zen: use ZEN as encoder.
--bert_model: the directory of pre-trained BERT/ZEN model.
--use_memory: use key-value memory networks.
--decoder: use crf or softmax as the decoder.
--ngram_flag: use av, dlg, or pmi to construct the lexicon N.
--av_threshold: when using av to construct the lexicon N, n-grams whose AV score is lower than the threshold will be excluded from the lexicon N.
--ngram_num_threshold: n-grams whose frequency is lower than the threshold will be excluded from the lexicon N. Note that, when the threshold is set to 1, no n-gram is filtered out by its frequency. We therefore DO NOT recommend you to use 1 as the n-gram frequency threshold.
--model_name: the name of model to save.

Predicting

run_sample.sh contains the command line to segment the sentences in an input file (./sample_data/sentence.txt).