Elamur Research⌕
☰
home / papers / 2018 / The WAW Corpus: The First Corpus of Interpreted Speeches and their Translations for English and Arabic

The WAW Corpus: The First Corpus of Interpreted Speeches and their Translations for English and Arabic

Summary

Machine speech translation still trails human simultaneous interpreters, and no public interpreting corpus existed for Arabic. To study how interpreters cope with real-time pressure and to eventually transfer their tactics into a speech-to-text MT system, the authors built the WAW Corpus. It is based on talks from three conferences held in Qatar, WISE 2013, ARC'14 and WISH 2013. The corpus links original speeches with their live English↔Modern Standard Arabic interpretations, verbatim transcripts of both, and professional translations of each transcript into the other language. The transcripts carry tags for named entities, non-lexical noises and speech phenomena. The authors assess the data with token-length ratios and named-entity ratios, which show that interpreters produce markedly less text than translators and vary more. An expert annotated eight sessions for interpreting strategies. Omissions came out as the dominant strategy. A signed combination of strategy counts tracks the length ratio closely. The resource matters for interpreter training, for interpreting research and for building MT systems that must work under time constraints.

Key points

  • The collection has 521 sessions, about 119 hours in total, averaging 13.56 min each. The study uses a fully parallel subset of 252 sessions (62h52m) covering 12 interpreters. Audio was captured as two-channel 48 kHz, 24-bit WAV.
  • Each item has six parts: the original audio and the interpretation audio (A0/B0), their transcripts (A1/B1), and cross-language translations of those transcripts (B2/A2). Metadata records the interpreter and the length of each item.
  • The English side has 286,024 words in 26,572 segments, and the Arabic interpretation side has 156,814 words in 9,413 segments. Translations add 222,025 English and 194,927 Arabic words.
  • Translations keep the usual Arabic/English token ratio of roughly 0.7. Interpretations fall well below this ratio and spread more widely across interpreters.
  • The strategy study covered 8 sessions by 4 interpreters (W2, W4, M1, M7), totalling 16,955 English and 9,477 Arabic words. It used five categories:
    • summarizing
    • self-correction
    • omissions
    • additions
    • transliterations
  • Omissions correlate with the length ratio at 0.89, and the combined metric (additions − omissions + self-correction − summarizing) reaches 0.92. No gender effect appeared in this sample.

Interesting ideas

  • I1. The authors translate the interpreter's output back into the source language, so that original and interpretation can be compared within one language rather than across languages.
  • I2. Named-entity counts serve as a cheap proxy for fidelity, measuring how many tagged persons, places and organisations survive the transfer.
  • I3. Every deviation from the source gets both a strategy label and a judgment of acceptable or unacceptable, which separates skilful condensing from real errors.
  • I4. Single strategy counts explain output length poorly. A signed sum reflecting whether each strategy lengthens or shortens text explains it much better.
  • I5. Transcription guidelines explicitly mark hesitations, repetitions, false starts and self-corrections as speech phenomena instead of cleaning them away.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (back): Run the text → graph → text compressor and compare regenerated English against the source in the same language, so that lost or invented content shows up directly.
  • I2 (world): Count entities in the world of the text before and after a round trip through generation or translation, and treat dropped entities as a fidelity alarm.
  • I3 (back): Label generator deviations (omission, addition, summarization) as legitimate compression or defect, so that repair work targets only the harmful cases.
  • I4 (learn): When explaining an evaluation metric, combine error types with signs according to their direction of effect, and check the correlation instead of reading each count in isolation.
  • I5 (frames): Treat self-corrections and false starts as speech-act-like events, so that a later corrected statement overrides the earlier frame rather than both being asserted.

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

Topics Semitic languages

 Institutions

🎓Hamad bin Khalifa UniversityDoha · Qatar58 works · 11,454 citations

Cite

Cite as Ahmed Abdelali, Irina Temnikova, Samy Hedaya and Stephan Vogel. 2018. The WAW Corpus: The First Corpus of Interpreted Speeches and their Translations for English and Arabic. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).

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

@inproceedings{abdelali2018waw,
  title = {{The WAW Corpus: The First Corpus of Interpreted Speeches and their Translations for English and Arabic}},
  author = {Ahmed Abdelali, Irina Temnikova, Samy Hedaya and Stephan Vogel},
  booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)},
  year = {2018},
  doi = {10.63317/4yztmfdo5eha},
  url = {https://doi.org/10.63317/4yztmfdo5eha}
}

References

24 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Al-Khanji, R., El-Shiyab, S., and Hussein, R. (2000). On the use of compensatory strategies in simultaneous interpretation. Meta: Journal des traducteurs/Meta: Translators’ Journal, 45(3):548–557. find
  2. Ali, A., Bell, P., Glass, J., Messaoui, Y., Mubarak, H., Renals, S., and Zhang, Y. (2016a). The MGB-2 Challenge: Arabic multi-dialect broadcast media recognition. In Spoken Language Technology Workshop (SLT), 2016 IEEE. our card
  3. Ali, A., Dehak, N., Cardinal, P., Khurana, S., Yella, S. H., Glass, J., Bell, P., and Renals, S. (2016b). Automatic dialect detection in Arabic broadcast speech. In Interspeech, San Francisco, USA, pages 2934–2938. find
  4. Ali, A., Vogel, S., and Renals, S. (2017). Speech recognition challenge in the wild: Arabic MGB-3. In Automatic Speech Recognition and Understanding (ASRU), 2017 IEEE Workshop on. IEEE. our card
  5. Bartłomiejczyk, M. (2006). Strategies of simultaneous interpreting and directionality. Interpreting, 8(2):149–174. find
  6. Bendazzoli, C. and Sandrelli, A. (2005). An approach to corpus-based interpreting studies: Developing EPIC (european parliament interpreting corpus). In MuTra 2005– Challenges of Multidimensional Translation: Conference Proceedings, pages 1–12. find
  7. Dalvi, F., Zhang, Y., Khurana, S., Durrani, N., Sajjad, H., Abdelali, A., Mubarak, H., Ali, A., and Vogel, S. (2017). QCRI live speech translation system. EACL 2017, page 61. find
  8. Falbo, C. (2012). CorIT (Italian Television Interpreting Corpus): Classification criteria. Breaking Ground in Corpus-based Interpreting Studies, Bern et al., Peter Lang, pages 157–185. find
  9. House, J., Meyer, B., and Schmidt, T. (2012). CoSi-A corpus of consecutive and simultaneous interpreting. Multilingual corpora and multilingual corpus analysis, 14:295. find
  10. Hu, K. and Qing, T. (2009). Hanying huiyi kouyi zhong yupian yiyi xianhjua jiqi dongyin yanjiu–yixiang jiyu pingxing yuliaoku de yanjiu [A corpus-based study of explicitation of textual meaning in chinese-english conference interpreting]. PLA Foreign Studies University Journal, 4:67–73. find
  11. Kahane, E. (2000). Thoughts on the quality of interpretation. aiic.net May 13, 2000. Accessed September 30, 2017. http://aiic.net/p/197. ↗
  12. Kalina, S. (1998). Strategische Prozesse beim Dolmetschen: Theoretische Grundlagen, empirische Fallstu-dien, didaktische Konsequenzen, volume 18. G. Narr. find
  13. Khurana, S. and Ali, A. (2016). QCRI advanced transcription system (QATS) for the Arabic multi-dialect broadcast media recognition: MGB-2 challenge. In Spoken Language Technology Workshop (SLT), 2016 IEEE. find
  14. Kroll, J. F. and De Groot, A. M. (2009). Handbook of bilingualism: Psycholinguistic approaches. Oxford University Press. find
  15. Kumar, G., Cao, Y., Cotterell, R., Callison-Burch, C., Povey, D., and Khudanpur, S. (2014). Translations of the CALLHOME Egyptian Arabic corpus for conversational speech translation. In Proceedings of International Workshop on Spoken Language Translation (IWSLT). find
  16. Roderick, J. (2002). Conference Interpreting Explained (Translation Practices Explained). St. Jerome Publishing. find
  17. Salameh, M., Zantout, R., and Mansour, N. (2011). Improving the accuracy of English-Arabic statistical sentence alignment. In Int. Arab J. Inf. Technol., volume 8, pages 171–177. find
  18. Shlesinger, M. (1998). Corpus-based interpreting studies as an offshoot of corpus-based translation studies. Meta: Journal des traducteurs/Meta: Translators’ Journal, 43(4):486–493. find
  19. Temnikova, I., Abdelali, A., Hedaya, S., Vogel, S., and Al Daher, A. (2017). Interpreting strategies annotation in the WAW corpus. In The First Workshop on Human-Informed Translation and Interpreting Technology (HiT-IT), co-located with RANLP 2017. September 7 2017, Varna, Bulgaria. find
  20. Tohyama, H. and Matsubara, S. (2006). Collection of simultaneous interpreting patterns by using bilingual spoken monologue corpus. In Proceedings of the Language Resources and Evaluation Conference (LREC). find
  21. Wray, S. and Ali, A. (2015). Crowdsource a little to label a lot: Labeling a Speech Corpus of Dialectal Arabic. In Interspeech. (in press). find
  22. Zaghouani, W. (2017). Critical survey of the freely available Arabic corpora. arXiv preprint arXiv:1702.07835. ↗
  23. Zaidan, O. F. and Callison-Burch, C. (2014). Arabic dialect identification. Computational Linguistics, 40(1):171–202. our card
  24. Zwischenberger, C. and Pöchhacker, F. (2010). Survey on quality and role: Conference interpreters expectations and self-perceptions. Communicate! AIIC Webzine. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/abdelali-2018-waw-corpus-first-corpus-interpreted/