/LASER

Language-Agnostic SEntence Representations

Primary LanguagePythonOtherNOASSERTION

LASER Language-Agnostic SEntence Representations

LASER is a library to calculate and use multilingual sentence embeddings.

NEWS

  • 2019/11/08 CCMatrix is available: Mining billions of high-quality parallel sentences on the WEB [8]
  • 2019/07/31 Gilles Bodard and Jérémy Rapin provided a Docker environment to use LASER
  • 2019/07/11 WikiMatrix is available: bitext extraction for 1620 language pairs in WikiPedia [7]
  • 2019/03/18 switch to BSD license
  • 2019/02/13 The code to perform bitext mining is now available

CURRENT VERSION:

  • We now provide an encoder which was trained on 93 languages, written in 23 different alphabets [6]. This includes all European languages, many Asian and Indian languages, Arabic, Persian, Hebrew, ..., as well as various minority languages and dialects.
  • We provide a test set for more than 100 languages based on the Tatoeba corpus.
  • Switch to PyTorch 1.0

All these languages are encoded by the same BiLSTM encoder, and there is no need to specify the input language (but tokenization is language specific). According to our experience, the sentence encoder also supports code-switching, i.e. the same sentences can contain words in several different languages.

We have also some evidence that the encoder can generalizes to other languages which have not been seen during training, but which are in a language family which is covered by other languages.

A detailed description how the multilingual sentence embeddings are trained can be found in [6], together with an extensive experimental evaluation.

Dependencies

  • Python 3.6
  • PyTorch 1.0
  • NumPy, tested with 1.15.4
  • Cython, needed by Python wrapper of FastBPE, tested with 0.29.6
  • Faiss, for fast similarity search and bitext mining
  • transliterate 1.10.2, only used for Greek (pip install transliterate)
  • jieba 0.39, Chinese segmenter (pip install jieba)
  • mecab 0.996, Japanese segmenter
  • tokenization from the Moses encoder (installed automatically)
  • FastBPE, fast C++ implementation of byte-pair encoding (installed automatically)

Installation

  • set the environment variable 'LASER' to the root of the installation, e.g. export LASER="${HOME}/projects/laser"
  • download encoders from Amazon s3 by bash ./install_models.sh
  • download third party software by bash ./install_external_tools.sh
  • download the data used in the example tasks (see description for each task)

Applications

We showcase several applications of multilingual sentence embeddings with code to reproduce our results (in the directory "tasks").

For all tasks, we use exactly the same multilingual encoder, without any task specific optimization or fine-tuning.

License

LASER is BSD-licensed, as found in the LICENSE file in the root directory of this source tree.

Supported languages

Our model was trained on the following languages:

Afrikaans, Albanian, Amharic, Arabic, Armenian, Aymara, Azerbaijani, Basque, Belarusian, Bengali, Berber languages, Bosnian, Breton, Bulgarian, Burmese, Catalan, Central/Kadazan Dusun, Central Khmer, Chavacano, Chinese, Coastal Kadazan, Cornish, Croatian, Czech, Danish, Dutch, Eastern Mari, English, Esperanto, Estonian, Finnish, French, Galician, Georgian, German, Greek, Hausa, Hebrew, Hindi, Hungarian, Icelandic, Ido, Indonesian, Interlingua, Interlingue, Irish, Italian, Japanese, Kabyle, Kazakh, Korean, Kurdish, Latvian, Latin, Lingua Franca Nova, Lithuanian, Low German/Saxon, Macedonian, Malagasy, Malay, Malayalam, Maldivian (Divehi), Marathi, Norwegian (Bokmål), Occitan, Persian (Farsi), Polish, Portuguese, Romanian, Russian, Serbian, Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Swahili, Swedish, Tagalog, Tajik, Tamil, Tatar, Telugu, Thai, Turkish, Uighur, Ukrainian, Urdu, Uzbek, Vietnamese, Wu Chinese and Yue Chinese.

We have also observed that the model seems to generalize well to other (minority) languages or dialects, e.g.

Asturian, Egyptian Arabic, Faroese, Kashubian, North Moluccan Malay, Nynorsk Norwegian, Piedmontese, Sorbian, Swabian, Swiss German or Western Frisian.

References

[1] Holger Schwenk and Matthijs Douze, Learning Joint Multilingual Sentence Representations with Neural Machine Translation, ACL workshop on Representation Learning for NLP, 2017

[2] Holger Schwenk and Xian Li, A Corpus for Multilingual Document Classification in Eight Languages, LREC, pages 3548-3551, 2018.

[3] Holger Schwenk, Filtering and Mining Parallel Data in a Joint Multilingual Space ACL, July 2018

[4] Alexis Conneau, Guillaume Lample, Ruty Rinott, Adina Williams, Samuel R. Bowman, Holger Schwenk and Veselin Stoyanov, XNLI: Cross-lingual Sentence Understanding through Inference, EMNLP, 2018.

[5] Mikel Artetxe and Holger Schwenk, Margin-based Parallel Corpus Mining with Multilingual Sentence Embeddings arXiv, Nov 3 2018.

[6] Mikel Artetxe and Holger Schwenk, Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond arXiv, Dec 26 2018.

[7] Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong and Paco Guzman, WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia arXiv, July 11 2019.

[8] Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave and Armand Joulin CCMatrix: Mining Billions of High-Quality Parallel Sentences on the WEB