multilingual-parsing

Repository for Cross-lingual Parsing with Polyglot Training and Multi-treebank Learning: A Faroese Case Study, submitted to the DeepLo Workshop at EMNLP 2019.

The code will be tidied up over the next couple of weeks.

Table of Contents

• Installation
• Obtain data
• Train models
• Predict
• Projection steps
• Train/Predict target models

Installation

This project is developed in Python 3.6 using a Conda environment.

Download and install Conda.

Create a Conda environment with Python 3.6:

conda create -n multilingual_parsing python=3.6

Activate the Conda environment:

source activate multilingual_parsing

This project uses some new AllenNLP features which are not available in the official 0.8.4 release. As such, we will build the 0.8.5-unreleased version from the master branch on GitHub. If there are any problems try updating pip, setuptools and wheel as mentioned here.

cd
git clone https://github.com/allenai/allennlp.git
cd allennlp
pip install --editable .

Make the library available in $PYTHONPATH. From the multilingual_parsing directory:

cd /path/to/multilingual_parsing

export PYTHONPATH="$PWD/library"

# or permanently:
vim ~/.bashrc
export PYTHONPATH=/path/to/multilingual_parsing/library
source ~/.bashrc

Obtain data

You will need to obtain the original Faroese data for these experiments.

1. Clone the original repository to somewhere in your file system, e.g. in your home directory:

   cd $HOME && git clone https://github.com/ftyers/cross-lingual-parsing.git

2. Change directory to your clone of this repo and create a symbolic link to the original data:

   cd path/to/multilingual-parsing
   ln -s /home/user/cross-lingual-parsing/data/ .

This should create a directory structure multilingual-parsing/data/.

3. Download UD v2.2 treebanks.

   ./scripts/get_ud_treebank.sh

Create silver training/ development sets

We follow the same process to develop datasets with automatically predicted pos-labels as the CoNLL 2018 shared task. That is, we perform jack-knifing on the training set to predict POS tags. POS tags on the development set are predicted with a model trained on the gold-standard training set.

./scripts/create_k_folds.sh

./scripts/train_with_cross_val.sh

./scripts/predict_with_cross_val.sh

# train a model on full training data and predict the dev set:

./scripts/train_source_tagger.sh monolingual

./scripts/predict_with_cross_val.sh dev

Train source models

1. Train a source model on source treebanks. The model_type argument supplied can be either monolingual or multilingual and determines whether to use a monolingual or multilingual model accordingly. You will already have trained source taggers from the previous step.

   ./scripts/train_source_parser.sh <model_type>

Predict translated source files

1. Use a source model to predict annotations for files translated into source languages. The model_type argument supplied can be either monolingual or multilingual and determines whether to use a monolingual or multilingual model accordingly.

   # first supply tags
   ./scripts/predict_source_tagger.sh monolingual user
   
   # parse the translations
   ./scripts/predict_source_parser.sh <model_type> user

Projection steps

1. Project from source languges to the target language.

   ./scripts/project_all.sh <model_type>

2. Take only the valid sentences.

   ./scripts/validate_all.sh <model_type> single

3. Combine sentences where we have 3/4 valid projected sentences.

   python utils/treebanks_union.py <model_type>

4. Perform MST voting over the matching, calidated sentences.

   ./scripts/merge_all.sh <model_type>

5. Validate the voted sentences.

   ./scripts/validate_all.sh <model_type> combined

6. Check for double-headed sentences.

   ./scripts/check_double_headed_all.sh <model_type>

Train target models

Train a tagging and parsing models on the synthetic target treebank. model_type can be either monolingual or multilingual. We train a tagger here so that we can produce silver tags for the final test set.

./scripts/train_target_tagger.sh <model_type>

./scripts/train_target_parser.sh <model_type>

Predict target models

Predict the Faroese test set with the various target taggers and parsers.

./scripts/predict_target_tagger.sh <model_type>

./scripts/predict_target_parser.sh <model_type>

Train/ predict using multi-treebank target models.

./scripts/train_multi_target_tagger.sh <model_type>

./scripts/train_multi_ target_parser.sh <model_type>

./scripts/predict_multi_target_tagger.sh <model_type>

./scripts/predict_multi_target_parser.sh <model_type>