Task 1: Multilingual Grapheme-to-Phoneme Conversion

In this task, participants will create computational models that map a sequence of “graphemes”—characters—representing a word to a transcription of that word’s pronunciation.

This task is an important part of speech technologies including recognition and synthesis.

Data

Source

The data is primarily extracted from Wiktionary using the wikipron library (Lee et al. in press).

Languages

We initially provide data for 10 languages:

  1. Armenian (arm)
  2. Bulgarian (bul)
  3. French (fre)
  4. Georgian (geo)
  5. Hindi (hin)
  6. Hungarian (hun)
  7. Icelandic (ice)
  8. Korean (kor)
  9. Lithuanian (lit)
  10. Modern Greek (gre)

We will announce an additional 5 “surprise languages” just before the conclusion of the task.

Size

There are 3600 training data examples and 450 development and test data examples for each language.

Format

Training and development data are UTF-8-encoded tab-separated values files. Each example occupies a single line and consists of a grapheme sequence—NFC Unicode codepoints—a tab character, and the corresponding phone sequence, a roughly-phonemic IPA, tokenized using the segments library (Moran & Cysouw 2018). The following show three lines of Romanian data:

antonim a n t o n i m
ploaie  p lʷ a j e
pornește    p o r n e ʃ t e

Test files consist of a single column, containing grapheme sequences.

Please provide your results in the two-column (grapheme sequence, tab-character, tokenized phone sequence) TSV format, the same one used for the training and development data. If your system only provides the predicted phone sequences, use the UNIX command-line tool paste to combine the columns.

Exclusions

We exclude from the provided data any words which:

  • have multiple pronunciations in the source data
  • consist of less than 3 graphemes, or
  • consist of less than 3 phonemes.

External data

Participants are permitted to use:

  • open-source databases of phoneme inventories and features such as Phoible (Moran & McCloy 2019),
  • open-source pronunciation data for languages not targeted in this challenge, and
  • open-source morphological analyzers and lexicons such as UDLexicons (Sagot 2018).

Participants who use such data must disclose their use of it at time of submission.

Participants are not permitted to use any form of pronunciation data derived from Wiktionary, except for the provided training and test data; they are also not permitted to use external pronunciation dictionaries for any of the targeted languages.

Evaluation

Systems should predict a single phone sequence for each test example.

Metrics

The primary measure will be the word error rate (WER), which is the percentage of words for which the hypothesized transcription sequence does not match the gold transcription. We also report phone error rate (PER), the micro-averaged edit distance between hypotheses and gold transcriptions, computed by summing the minimum edit distance between the hypothesis and gold transcriptions and then dividing by the summed length of the gold transcriptions. As is common practice, we multiply both numbers by 100. Both metrics will be computed using the provided Python script evaluate.py.

System comparison

We will evaluate on each language separately. The final system ranking will be produced by macro-averaging the per-language WERs. We will also employ statistical analysis for system comparison.

Baselines

We provide implementations of two baseline systems for the task:

  • a pair n-gram model (Novak et al. 2016) implemented using the OpenGrm toolkit (Roark et al. 2012, Gorman 2016), and
  • a bidirectional LSTM encoder-decoder sequence model implemented using the Fairseq toolkit (Ott et al. 2019).

Participants are welcome to adapt these baselines for their purposes.

Overview paper

In an overview paper for the shared task, we will compare the performance of submitted systems in detail. We will assess:

  • which systems are significantly different in performance
  • which languages were challenging and which types of systems succeeded on them, and
  • which systems would provide complementary benefit in an ensemble system.

Included in the paper will be a summary of scores for all participants who produce outputs for all targeted languages.

Organizers

This task is organized by Lucas Ashby and Kyle Gorman at the Graduate Center, City University of New York, with help from other members of the WikiPron team.

Contact: Kyle Gorman.

References

Gorman, K. (2016). Pynini: a Python library for weighted finite-state grammar compilation. In Proceedings of the SIGFSM Workshop on Statistical NLP and Weighted Automata, pages 75–80, Berlin. Association for Computational Linguistics.

Lee, J. L, Ashby, L. F.E., Garza, M. E., Lee-Sikka, Y., Miller, S., Wong, A., McCarthy, A. D., and Gorman, K. (in press). Massively multilingual pronunciation mining with WikiPron. To appear in the proceedings of LREC 2020.

Moran, S. and Cysouw, M. (2018). The Unicode cookbook for linguists: managing writing systems using orthography profiles. Berlin: Language Science Press.

Moran, S. and McCloy, D. (2019). PHOIBLE 2.0. Jena: Max Planck Institute for the Science of Human History.

Novak, J. R., Minematsu, N., and Hirose, K. (2016). Phonetisaurus: exploring grapheme-to-phoneme conversion with joint n-gram models in the WFST framework. Natural Language Engineering, 22(6):907–938.

Ott, M., Edunov, S., Baevski, A., Fan, A., Gross, S., Ng, N., Grangier, D., and Auli, M. (2019). fairseq: a fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53, Minneapolis. Association for Computational Linguistics.

Roark, B., Sproat, R., Allauzen, C., Riley, M., Sorensen, J., and Tai, T. (2012). The OpenGrm open-source finite-state grammar software libraries. In Proceedings of the ACL 2012 System Demonstrations, pages 61–66, Jeju Island, Korea. Association for Computational Linguistics.

Sagot, B. 2018. A multilingual collection of CoNLL-U-compatible morphological lexicons. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC), pages 1861-1867. Miyazaki, Japan. European Language Resources Association.

Task 1: Multilingual Grapheme-to-Phoneme Conversion

In this task, participants will create computational models that map a sequence of “graphemes”—characters—representing a word to a transcription of that word’s pronunciation.

This task is an important part of speech technologies including recognition and synthesis.

Data

Source

The data is primarily extracted from Wiktionary using the wikipron library (Lee et al. in press).

Languages

We initially provide data for 10 languages:

  1. Armenian (arm)
  2. Bulgarian (bul)
  3. French (fre)
  4. Georgian (geo)
  5. Hindi (hin)
  6. Hungarian (hun)
  7. Icelandic (ice)
  8. Korean (kor)
  9. Lithuanian (lit)
  10. Modern Greek (gre)

We will announce an additional 5 “surprise languages” just before the conclusion of the task.

Size

There are 3600 training data examples and 450 development and test data examples for each language.

Format

Training and development data are UTF-8-encoded tab-separated values files. Each example occupies a single line and consists of a grapheme sequence—NFC Unicode codepoints—a tab character, and the corresponding phone sequence, a roughly-phonemic IPA, tokenized using the segments library (Moran & Cysouw 2018). The following show three lines of Romanian data:

antonim a n t o n i m
ploaie  p lʷ a j e
pornește    p o r n e ʃ t e

Test files consist of a single column, containing grapheme sequences.

Please provide your results in the two-column (grapheme sequence, tab-character, tokenized phone sequence) TSV format, the same one used for the training and development data. If your system only provides the predicted phone sequences, use the UNIX command-line tool paste to combine the columns.

Exclusions

We exclude from the provided data any words which:

  • have multiple pronunciations in the source data
  • consist of less than 3 graphemes, or
  • consist of less than 3 phonemes.

External data

Participants are permitted to use:

  • open-source databases of phoneme inventories and features such as Phoible (Moran & McCloy 2019),
  • open-source pronunciation data for languages not targeted in this challenge, and
  • open-source morphological analyzers and lexicons such as UDLexicons (Sagot 2018).

Participants who use such data must disclose their use of it at time of submission.

Participants are not permitted to use any form of pronunciation data derived from Wiktionary, except for the provided training and test data; they are also not permitted to use external pronunciation dictionaries for any of the targeted languages.

Evaluation

Systems should predict a single phone sequence for each test example.

Metrics

The primary measure will be the word error rate (WER), which is the percentage of words for which the hypothesized transcription sequence does not match the gold transcription. We also report phone error rate (PER), the micro-averaged edit distance between hypotheses and gold transcriptions, computed by summing the minimum edit distance between the hypothesis and gold transcriptions and then dividing by the summed length of the gold transcriptions. As is common practice, we multiply both numbers by 100. Both metrics will be computed using the provided Python script evaluate.py.

System comparison

We will evaluate on each language separately. The final system ranking will be produced by macro-averaging the per-language WERs. We will also employ statistical analysis for system comparison.

Baselines

We provide implementations of two baseline systems for the task:

  • a pair n-gram model (Novak et al. 2016) implemented using the OpenGrm toolkit (Roark et al. 2012, Gorman 2016), and
  • a bidirectional LSTM encoder-decoder sequence model implemented using the Fairseq toolkit (Ott et al. 2019).

Participants are welcome to adapt these baselines for their purposes.

Overview paper

In an overview paper for the shared task, we will compare the performance of submitted systems in detail. We will assess:

  • which systems are significantly different in performance
  • which languages were challenging and which types of systems succeeded on them, and
  • which systems would provide complementary benefit in an ensemble system.

Included in the paper will be a summary of scores for all participants who produce outputs for all targeted languages.

Organizers

This task is organized by Lucas Ashby and Kyle Gorman at the Graduate Center, City University of New York, with help from other members of the WikiPron team.

Contact: Kyle Gorman.

References

Gorman, K. (2016). Pynini: a Python library for weighted finite-state grammar compilation. In Proceedings of the SIGFSM Workshop on Statistical NLP and Weighted Automata, pages 75–80, Berlin. Association for Computational Linguistics.

Lee, J. L, Ashby, L. F.E., Garza, M. E., Lee-Sikka, Y., Miller, S., Wong, A., McCarthy, A. D., and Gorman, K. (in press). Massively multilingual pronunciation mining with WikiPron. To appear in the proceedings of LREC 2020.

Moran, S. and Cysouw, M. (2018). The Unicode cookbook for linguists: managing writing systems using orthography profiles. Berlin: Language Science Press.

Moran, S. and McCloy, D. (2019). PHOIBLE 2.0. Jena: Max Planck Institute for the Science of Human History.

Novak, J. R., Minematsu, N., and Hirose, K. (2016). Phonetisaurus: exploring grapheme-to-phoneme conversion with joint n-gram models in the WFST framework. Natural Language Engineering, 22(6):907–938.

Ott, M., Edunov, S., Baevski, A., Fan, A., Gross, S., Ng, N., Grangier, D., and Auli, M. (2019). fairseq: a fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53, Minneapolis. Association for Computational Linguistics.

Roark, B., Sproat, R., Allauzen, C., Riley, M., Sorensen, J., and Tai, T. (2012). The OpenGrm open-source finite-state grammar software libraries. In Proceedings of the ACL 2012 System Demonstrations, pages 61–66, Jeju Island, Korea. Association for Computational Linguistics.

Sagot, B. 2018. A multilingual collection of CoNLL-U-compatible morphological lexicons. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC), pages 1861-1867. Miyazaki, Japan. European Language Resources Association.