• python 3.6
• pytorch 1.1.0
• faiss 1.5.3 (GPU version if available)
• numpy 1.16.4
• nltk 3.2.4
• gensim 3.2.0
• python-levenshtein 0.12.0
• tqdm 4.19.5
• langid 1.1.6

Script for installing all dependencies (using conda):

./install_requirements.sh

• Moses
• fastText
• MUSE

Script for install tools:

./get_third_party.sh

Parameters can be modified in environment.sh

• Executable paths
• Data paths and language pairs
• Experiment parameters

The project uses monolingual Newscrawl data from the WMT news translation shared task, gold annotations from the BUCC 2017 shared task and Paracrawl from WMT parallel corpus filtering task. The required data can be downloaded and preprocessed by running:

./get_data.sh

Steps:

• Downloading data
• Tokenization
• Training truecase model
• Truecaseing the data
• Output: ./data/
• The script only prepares En, De, Fr and Ru data. You need to prepare monolingual data and truecase models in case of other languages.

The approach relies on dictionaries based on bilingual word embeddings and orthography. Embeddings and dictionaries can be generated by running:

./generate_dictionaries.sh

Steps:

• Running fastText on monolingual data
• Mapping embeddings spaces using MUSE
• Creating embeddings based dictionaries
• Creating orthography based dictionaries
• Output: ./results/dictionaries/

The scripts evaluates our method on the BUCC 2017 parallel sentence mining task.

./mine_bucc.sh
./print_bucc_results.sh

Steps:

• Prefiltering based on document embeddings
• Scoring sentence pair candidates
• Evaluating
• Output: ./results/mining/bucc2017

By default the script extracts sentence pairs from Newscrawl 2011 which have a token length between 21 and 25. For changing default setup see environment.sh and get_data.sh.

./extract_sentences.sh

Steps:

• Prefiltering based on document embeddings
• Scoring sentence pair candidates
• Extracting sentences
• Output: ./results/mining/

The scripts filters small portion of noisy parallel corpus released by WMT18 - Parallel corpus filtering task. For changing default setup see environment.sh.

./corpus_filter.sh

Steps:

• Prefiltering based on simple rules
• Scoring pairs
• Extracting pairs with best scores
• Output: ./results/filtering/

Related publications:

[1] V. Hangya and A. Fraser, Unsupervised Parallel Sentence Extraction with Parallel Segment Detection Helps Machine Translation in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019.

[2] V. Hangya, F. Braune, Y. Kalasouskaya, and A. Fraser, Unsupervised Parallel Sentence Extraction from Comparable Corpora. in Proceedings of the 15th International Workshop on Spoken Language Translation (IWSLT), 2018

[3] V. Hangya and A. Fraser, An Unsupervised System for Parallel Corpus Filtering. in Proceedings of the EMNLP 2018 Third Conference on Machine Translation (WMT), 2018

@inproceedings{Hangya2019,
author = {Hangya, Viktor and Fraser, Alexander},
title = {{Unsupervised Parallel Sentence Extraction with Parallel Segment Detection Helps Machine Translation}},
booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
year = {2019}
}

@inproceedings{Hangya2018,
author = {Hangya, Viktor and Braune, Fabienne and Kalasouskaya, Yuliya and Fraser, Alexander},
title = {{Unsupervised Parallel Sentence Extraction from Comparable Corpora}},
booktitle = {Proceedings of the 15th International Workshop on Spoken Language Translation (IWSLT)},
pages = {7--13},
year = {2018}
}

@inproceedings{Hangya2018,
author = {Hangya, Viktor and Fraser, Alexander},
title = {{An Unsupervised System for Parallel Corpus Filtering}},
booktitle = {Proceedings of the EMNLP 2018 Third Conference on Machine Translation (WMT)},
pages = {882--887},
year = {2018}
}