rootofmylife / clbt Goto Github PK

This is the implementation of the paper Cross-Lingual BERT Transformation for Zero-Shot Dependency Parsing. If you use this code or our results in your research, we'd appreciate you cite our paper as following:

@inproceedings{wang2019cross,
  author    = {Wang, Yuxuan and Che, Wanxiang and Guo, Jiang and Liu, Yijia and  Liu, Ting},
  title     = {Cross-Lingual BERT Transformation for Zero-Shot Dependency Parsing},
  booktitle = {Proc. of EMNLP},
  year      = {2019}
}

• Python 3.6 with NumPy/SciPy
• [PyTorch] 0.4.1 (http://pytorch.org/) detailed requirements are in the requirements.txt file.

Training and testing scripts for fast startup are in the scripts file. (the source line is used to activate virtualenv environment, if the requirements are installed in corrent environment, just remove the line) Please first run the train.sh shell to train a model and the the test.sh shell to test the model.

The following pre-processing is required for training, check trial_data file to see the format of each required input.

• Parallel sentences for source (de)/target (en) languages
• Use pytorch BERT model to predict the contextualized embeddings (-1 layer) of source language (e.g. de in the trial data) and target language (en). (For English please use the base en model, while for other languages use the multilingual model)
• Alignments between the two languages. (obtained from fast_aling) Note that this must be alignments between sentences tokenized by word piece vocabulary of corresponding BERT model.

Some options:

• --map_type: svd/linear/nonlinear. The linear is optimized with Adam. SVD is also a linear transformation and is faster, but --no_cuda option should be added when trained on big data. linear is recommended since it achieves the best performances in most languages.
• --n_layers: number of layers for nonlinear transformation. (don't need to add if the model is linear)
• --bert_file0: training data from source language (e.g. English).
• --bert_file1: training data from target language (e.g. German), should be parallel to bert_file0
• --align_file: the alignment obtained by fast_align, the parallel sentences should be tokenized with the corresponding BERT vocabulary. (i.e. use EN Base vocab for English, use Multilingual Base vocab for other languages)
• --vocab_file: bert vocab file of source language, should be the same to the vocab file of --bert_file0.
• --vocab_file1: bert vocab file of target language, should be the same to the vocab file of --bert_file1.
• --n_epochs: maximum training epochs.
• --model_path: where the model should be saved.
• --batch_size: training batch size.

The testing procedure requires a bert file (-1 layer) and a transformation model trained by the training process.

Options:

• --model_path: input transformation model path.
• --bert_file0: input bert file to be transformed.
• --map_type: for svd/linear model, this should be linear, while for nonlinear model, this is nonlinear.
• --n_layers: number of layers for nonlinear transformation, this must be the same as the layer in the input model (don't need to add if the model is linear).
• --output_file: output path.
• --batch_size: predicting batch size.
• --pred: activate predicting mode.

We have uploaded pretrained transformation matrices in the models file for 17 languages with 2 optimizing mechanisms (i.e., SVD and GD). If you would like to use these pretrained matrices, just use the transform.sh shell from the scripts file:

./scripts/transform.sh [GPU] [input bert file] [optimizer(gd|svd)] [language pair(e.g. de-en)] [output bert file]

For instance, to transform the German bert file in the trial_data with matrix trianed by SVD, use:

./scripts/transform.sh 1 trial_data/de.trial.bert svd de-en de.trial.svd.bert

The languages whose transformation matrices to English is provided are listed below:

• Bulgarian (bg)
• Czech (cs)
• Danish (da)
• German (de)
• Spanish (es)
• Estonian (et)
• Finnish (fi)
• French (fr)
• Italian (it)
• Latvian (lv)
• Dutch (nl)
• Polish (pl)
• Portuguese (pt)
• Romanian (ro)
• Slovak (sk)
• Slovenian (sl)
• Swedish (sv)