SAPar

This is the implementation of Constituency Parsing with Span Attention at Findings of EMNLP2020.

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

Citation

If you use or extend our work, please cite our paper at Findings of EMNLP-2020.

@inproceedings{tian-etal-2020-improving,
    title = "Improving Constituency Parsing with Span Attention",
    author = "Tian, Yuanhe and Song, Yan and Xia, Fei and Zhang, Tong",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
    month = nov,
    year = "2020",
    address = "Online",
    pages = "1691--1703",
}

Prerequisites

• python 3.6
• pytorch 1.1

Install python dependencies by running:

pip install -r requirements.txt

EVALB and EVALB_SPMRL contain the code to evaluate the parsing results for English and other languages. Before running evaluation, you need to go to the EVALB (for English) or EVALB_SPMRL (for other languages) and run make.

Downloading BERT, ZEN, XLNet and Our Pre-trained Models

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

For BERT, please download pre-trained BERT model from Google and convert the model from the TensorFlow version to PyTorch version.

• For Arabic, we use MulBERT-Base, Multilingual Cased.
• For Chinese, we use BERT-Base, Chinese;
• For English, we use BERT-Large, Cased and BERT-Large, Uncased.

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

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

For our pre-trained model, you can download them from Baidu Wangpan (passcode: 2o1n) or Google Drive.

Run on Sample Data

To train a model on a small dataset, run:

./run.sh

Datasets

We use datasets in three languages: Arabic, Chinese, and English.

• Arabic: we use ATB2.0 part 1-3 (LDC2003T06, LDC2004T02, and LDC2005T20).
• Chinese: we use CTB5 (LDC2005T01).
• English: we use PTB (LDC99T42).

To preprocess the data, please go to data_processing directory and follow the instruction to process the data. You need to obtain the official datasets yourself before running our code.

Ideally, all data will appear in ./data directory. The data with gold POS tags are located in folders whose name is the same as the dataset name (i.e., ATB, CTB, and PTB); the data with predicted POS tags are located in folders whose name has a "_POS" suffix (i.e., ATB_POS, CTB_POS, and PTB_POS).

Training, Testing, and Predicting

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

To-do List

• Regular maintenance.

You can leave comments in the Issues section, if you want us to implement any functions.

You can check our updates at updates.md.