Elamur Research
☰
home / papers / 2012 / WIT3: Web Inventory of Transcribed and Translated Talks

WIT3: Web Inventory of Transcribed and Translated Talks

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/2012.eamt-1.60.pdf

Summary

Statistical MT systems need large amounts of in-domain parallel text, but free bilingual data exist for only a few language pairs and narrow domains such as parliamentary proceedings or legislation. The authors turn the volunteer-made subtitles of TED talks, released under Creative Commons BY-NC-ND, into a resource researchers can use directly. Their crawler, HLTWebManager, downloads subtitles and metadata and stores them as one XML file per language. Each talk carries a shared talkid, so translations of the same talk can be matched. Subtitle captions are paired by order of appearance. Talks with mismatched caption counts or timestamps are dropped, and caption pairs with abnormal length ratios are removed. Sentences are then rebuilt by joining captions up to strong punctuation. The resulting TED Talks corpus covers many topics and dozens of languages, far more than existing corpora. It formed the basis of the IWSLT 2011 MT tracks, where top scores resembled WMT news results for English–French. Baselines on extra pairs gave usable output into English, while German–Italian proved harder. Everything is distributed through the WIT3 website.

Key points

  • As of October 2011 the collection held almost 17,000 transcripts, translating about 1,000 English talks into 80 languages; the text is kept untokenized.
  • Available parallel data:
    • roughly 2M units for each pair among English, French and Spanish;
    • at least 1M units for all 36 pairs among 9 languages;
    • at least 500K units for 120 pairs (16 languages);
    • at least 200K units for 253 pairs (23 languages).
  • Caption-level filtering drops length-ratio outliers outside a 95% confidence interval. For English–French it removes about 3% of words.
  • The site (wit3.fbk.eu) offers the XML files, the DTD and the parallel data. It also provides Perl tools: find-common-talks.pl, filter-talks.pl, ted-extract-par.pl, ted-extract-mono.pl and rebuild-sent.pl.
  • Baselines were built for nl-en, de-en, en-it and de-it and evaluated with BLEU, METEOR and TER on dev2010/tst2010. German–Italian scored lowest, possibly because of scarce data or the lack of decompounding.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages · Machine translation

Cited by

163 papers of our library cite it (of the 140673 whose bibliographies we read); 9 of them have a page here, 154 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{cettolo2012wit3,
  title = {{WIT3: Web Inventory of Transcribed and Translated Talks}},
  author = {Cettolo, Mauro and Girardi, Christian and Federico, Marcello},
  booktitle = {Proceedings of the 16th Annual Conference of the European Association for Machine Translation},
  year = {2012},
  url = {https://aclanthology.org/2012.eamt-1.60/}
}

References

14 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Callison-Burch, C., P. Koehn, C. Monz, and O. Zaidan 2011. Findings of the 2011 Workshop on Statistical Machine Translation. In Proc. of WMT, pp. 22-64, Edinburgh, Scotland, UK. find
  2. Chen, S. F. and J. Goodman. 1999. An Empirical Study of Smoothing Techniques for Language Modeling. Computer Speech and Language, 4(13):359–393. find
  3. Clark, J., C. Dyer, A. Lavie, and N. Smith. 2011. Better Hypothesis Testing for Statistical Machine Translation: Controlling for Optimizer Instability. In Proc. of ACL, Portland, US-OR. find
  4. Coughlin, D. A. 2003. Correlating Automated and Human Assessments of Machine Translation Quality. In Proc. of MT Summit, pp. 23-27, New Orleans, US-LA. find
  5. Diab, M., K. Hacioglu, and D. Jurafsky. 2004. Automatic Tagging of Arabic Text: from Raw Text to Base Phrase Chunks. In Proc. of HLT-NAACL: Short Papers, pp. 149–152, Boston, US-MA. find
  6. Eisele, A. and Y. Chen. 2010. MULTIUN: A Multilingual Corpus from United Nation Documents. In Proc. of LREC, Valletta, Malta. find
  7. Federico, M., N. Bertoldi, and M. Cettolo. 2008. IRSTLM: an Open Source Toolkit for Handling Large Scale Language Models. In Proc. of Interspeech, pp. 1618–1621, Melbourne, Australia. find
  8. Federico, M., L. Bentivogli, M. Paul, and S. Stüker. 2011. Overview of the IWSLT 2011 Evaluation Campaign. In Proc. of IWSLT, San Francisco, US-CA. find
  9. Girardi. 2011. The HLT Web Manager. FBK Technical Report n. 23969. Trento, Italy. https://wit3.fbk.eu/tools/WebManage r Manual.pdf Koehn, P. et al. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Proc. of ACL: Demo and Poster Sessions, pp. 177–180, Prague, Czech Republic. ↗
  10. Koehn, P. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In Proc. of MT Summit, pp. 79–86, Phuket, Thailand. find
  11. Paul, M., M. Federico, and S. Stücker. 2010. Overview of the IWSLT 2010 Evaluation Campaign. In Proc. of IWSLT, pp. 3–27, Paris, France. find
  12. Takezawa, T., G. Kikui, M. Mizushima, and E. Sumita. 2007. Multilingual Spoken Language Corpus Development for Communication Research. Computational Linguistics and Chinese Language Processing, 12(3):303–324. find
  13. Tiedemann, J. 2009. News from OPUS - a Collection of Multilingual Parallel Corpora with Tools and Interfaces. In Recent Advances in Natural Language Processing (vol V), pp. 237–248. John Benjamins, Amsterdam/Philadelphia. find
  14. Tseng, H., P. Chang, G. Andrew, D. Jurafsky, and C. Manning. 2005. A Conditional Random Field Word Segmenter. In Proc. of SIGHAN Workshop on Chinese Language Processing, Jeju Island, Korea. find