NeST-CCG

This is the implementation of Supertagging Combinatory Categorial Grammar with Attentive Graph Convolutional Networks at EMNLP2020.

Please contact us 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 NeST-CCG

We are improving our NeST-CCG. For updates, please visit HERE.

Citation

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

@inproceedings{tian-etal-2020-supertagging,
    title = "Supertagging Combinatory Categorial Grammar with Attentive Graph Convolutional Networks",
    author = "Tian, Yuanhe and Song, Yan and Xia, Fei",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    pages = "6037--6044",
}

Prerequisites

Our code works with the following environment.

• python=3.7
• pytorch=1.3

Install python dependencies by running:

pip install -r requirements.txt

You also need Java 1.8 to run tag2auto.jar, which wiil generate the CCG parsing results from the predicted supertags. You can skip this step if you only want to get the supertagging results.

Evaluation

To evaluate the CCG parsing results generated from the predicted supertags, you need to setup C&C parser. To do this, just run ./setup_condc.sh.

To test whether C&C parser is successfully installed for the purpose of evaluation, run ./candc/bin/generate. If you see the following, then it means C&C parser is successfully installed.

$ ./candc/bin/generate
expected a flag indicating output type
usage: generate -t <cats_directory> <markedup_file> <pipe_file> > <deps> 2> <feats>
       generate -e <cats_directory> <markedup_file> <pipe_file> > <word_deps>
       generate -j <cats_directory> <markedup_file> <pipe_file> > <word_deps>
       generate -g <cats_directory> <markedup_file> <pipe_file> > <grs>
       generate -T <cats_directory> <markedup_file> <pipe_file> > <raw text>

Downloading BERT and Our Pre-trained Models

In our paper, we use BERT 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 our pre-trained models, we have two versions: the first follows previous studies using the most frequent 425 supertags as the tag set (Baidu Wangpan (passcode: vhbq), Google Drive; the second uses all supertags appearing in the training set (Baidu Wangpan (passcode: u4ta), Google Drive)). In our paper, we report the results of the first version.

Run on Sample Data

To train a model on a small dataset, see the command lines in run.sh.

Datasets

We use CCGbank in our paper.

To preprocess the data, please go to data_processing directory and run ./data_processing. You can find more details here. You need to obtain the official CCGbank yourself before running our code.

If everything goes smoothly, you will see all data files in the ./data directory with all filenames identical with the ones in ./sample_data.

Training and Testing

You can find the command lines to train and test models on the sample data 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_gcn: whether to use GCN.
• --use_weight: whether to use A-GCN.
• --use_in_chunk: whether to in-chunk edges to build the graph.
• --use_cross_chunk: whether to cross-chunk edges to build the graph.
• --gcn_layer_number: the number of GCN layers.
• --bert_model: the directory of pre-trained BERT model.
• --max_ngram_length: the max length of n-grams.
• --ngram_freq_threshold: n-grams whose frequency is lower than the threshold will be excluded from the lexicon N.
• --model_name: the name of model to save.

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.