Elamur Research⌕
☰
home / papers / 2007 / Enhancing image-based Arabic document translation using noisy channel correction model

Enhancing image-based Arabic document translation using noisy channel correction model

Paper · 2007 · Yi Chang, Ying Zhang, Stephan Vogel et al.

Summary

The paper addresses a weakness of camera-based translation of Arabic documents. Text is detected in a photo, read by commercial OCR and then passed to statistical MT. Recognition mistakes flow downstream and multiply into translation mistakes. In the authors' setup, perfect recognition gives BLEU 43.12, but a word error rate of about 10.7% drops it to 28.56. To reduce this damage, they place a noisy channel correction model between the OCR engine (Sakhr Automatic Reader 8.0) and the translator (PanDoRA, a compact phrase-based SMT system). The channel is learned from synthetic clean images rendered in several common Arabic fonts and sizes. Error analysis showed that most remaining errors were wrong word boundaries, especially single words split into pieces. They therefore added a bigram step that merges adjacent fragments when the merged form is close to a dictionary word. With Arabic Transparent, BLEU rises from 18.70 to 33.47, near the 34.01 achievable with the best correction. This shows that cheap post-OCR repair can recover much of the loss in a pipeline.

Key points

  • The pipeline has three stages: text detection (multi-scale edges, GMM colour segmentation, affine rectification), then Sakhr OCR, then PanDoRA. PanDoRA runs in monotonic decoding because the test sentences are short.
  • Test images were clean 600 dpi renders in Arabic Transparent, Simplified Arabic and Traditional Arabic at 36/60/84/108 pixels.
    • OCR quality varies strongly with font and size; Traditional Arabic, for example, ranges from 28.2% character error at 36 px to 0.3% at 84–108 px.
    • Word error was roughly four times the character error.
  • Repairing every segmentation error by hand would bring word error under 2% and BLEU above 42 for all fonts. This confirms segmentation as the main error source.
  • The bigram correction works in three steps:
    • It joins adjacent segments whenever the edit distance to a dictionary word is below n = 2.
    • It repeats this replacement step.
    • It uses Viterbi to pick the best word sequence from both the original and the merged strings.
  • Extending the merge step to trigrams brought no extra gain. Cases where several real words were read as one remain unsolved, so the improvement for Traditional Arabic was marginal.

Interesting ideas

  • I1. Learn the error channel from synthetic renders with controlled variation (font, size) instead of hand-labelled noisy data, so the noise model costs almost nothing to produce.
  • I2. Run an oracle experiment that fixes one error type completely (here segmentation) to measure how much of the end-to-end loss it explains before investing in a fix.
  • I3. Keep both the raw input and the repaired candidate, and let a sequence decoder choose between them rather than applying corrections blindly.
  • I4. A correction can make output more misleading, as when a fragment became "hair" and produced a wrong translation instead of an unknown word; repairs need a cost, not just a dictionary hit.
  • I5. Dictionary lookup with an edit-distance threshold over adjacent token pairs is a simple, explainable way to fix words that were wrongly broken apart.

For Mova

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

  • I1 (learn): Silver training data for repair rules could be generated by deliberately corrupting clean text (typos, broken tokenisation) and learning how to undo each corruption.
  • I2 (learn): Evaluation could include oracle ablations, such as gold tokens or gold tree, to show which stage actually limits answer accuracy.
  • I3 (tokens): The tokenizer could offer merge or split alternatives as a small lattice, and the tagger or parser would choose the reading that scores best.
  • I4 (answer): A repair that changes meaning should lower confidence or be flagged in the explanation, so the reader does not answer confidently from a fabricated word.
  • I5 (tokens): This could be a transparent rule for rejoining split words in noisy English input (for example "some thing", or line-break hyphenation), with each merge logged.

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

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksComputer Vision and Pattern Recognition · 1,282 works

🧭 Topics 2Research subjects and languages the work is about.Machine translation · 958 worksSemitic languages · 1,484 works

Cite

Cite as Yi Chang, Ying Zhang, Stephan Vogel and Jie Yang. 2007. Enhancing image-based Arabic document translation using noisy channel correction model. Proceedings of Machine Translation Summit XI: Papers.

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

@inproceedings{chang2007enhancing,
  title = {{Enhancing image-based Arabic document translation using noisy channel correction model}},
  author = {Yi Chang, Ying Zhang, Stephan Vogel and Jie Yang},
  booktitle = {Proceedings of Machine Translation Summit XI: Papers},
  year = {2007},
  url = {https://aclanthology.org/2007.mtsummit-papers.12/}
}

References

20 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Chen, X., Yang, J., Zhang, J., and Waibel, A. (2004). Automatic detection and recognition of signs from natural scenes, IEEE Transactions on Image Processing, 13(1), 87-99. find
  2. Croft, W.B., Harding, S., Taghva, K. and Borsack, J. (1993). An Evaluation of Information Retrieval Accuracy with Simulated OCR Output. Technical Report: UM-CS-1993-076. find
  3. Doermann, D., and Yao, S. (1995). Generating synthetic data for text analysis systems. Symposium on Document Analysis and Information Retrieval, (pp. 449-467). find
  4. Eck, M. and Hori, C. (2005). Overview of the IWSLT 2005 evaluation campaign. In Proceedings of International Workshop on Spoken Language Translation, (pp. 11-17), Lisbon, Portugal. find
  5. Eikvil, L. (1993). OCR Optical Character Recognition. Technical Report, Report No. 876. find
  6. Harding, S.M., Croft, W.B. and Weir, C. (1997). Probabilistic Retrieval of OCR Degraded Text Using N-Grams. Proceedings of the First European Conference on Research and Advanced Technology for Digital Libraries (pp. 345-359), Pisa, Italy. find
  7. Hong, T. (1995). Degraded Text Recognition Using Visual and Linguistic Context. Ph.D. thesis, Computer Science Department, SUNY Buffalo. find
  8. Huang, X., Acero, A. and Hon, H. (2001). Spoken Language Processing: A Guide to Theory, Algorithm, and System Development. Prentice-Hall. find
  9. Kanungo, T., Marton, G. and Bulbul, O. (1999). OmniPage vs. Sakhr: Paired model evaluation of two Arabic OCR products. Proceedings of SPIE Conf. on Document Recognition, (pp. 109-120), San Jose, CA. find
  10. Kolak, O. and Resnik, P. (2002). OCR error correction using a noisy channel model. In Human Language Technology Conference (HLT 2002), San Diego, CA. find
  11. Mittendorf, E., Schäuble, P. and Sheridan, P. (1995). Applying Probabilistic Term Weighting to OCR Text in the Case of a Large Alphabetic Library Catalogue. Proceedings of the 18th ACM SIGIR Conference on Research and Development in Information Retrieval, (pp. 328-335) Seattle, WA. find
  12. Myka, A. and Güntzer, U. (1997). Measuring the Effects of OCR Errors on Similarity Linking, Proceeding of ICDAR, (pp. 968-973), Ulm, Germany. find
  13. Papineni, K. and Roukos, S. Ward, T. and Zhu, W. (2002). BLEU: A Method for Automatic Evaluation of Machine Translation. Proceedings of the 20th Annual Meeting of the Association for Computational Linguistics, (pp. 311-318), Philadelphia, Pennsylvania.
  14. Rice, S; Nagy, G.; Nartker, T. (1999). Optical Character Recognition: An Illustrated Guide to the Frontier. find
  15. Kluwer, Boston, Sato, T., Kanade, T., Hughes, E.K., Smith, M.A. and Satoh, S. (1999). Video OCR: Indexing Digital News Libraries by Recognition of Superimposed Captions. ACM Multimedia Systems, vol.7, 385-394. ↗
  16. Taghva, K., Borsack, J. and Condit, A. (1994). An Expert System for Automatically Correcting OCR Output. Proceedings of the SPIE-Document Recognition, (pp. 270-278), San Jose, CA. find
  17. Vogel, S., Zhang, Y., Huang, F., Tribble, A., Venogupal, A., Zhao, B. and Waibel, A. (2003). The CMU Statistical Translation System. Proceedings of MT Summit IX, New Orleans. find
  18. Yang, J., Gao, J., Zhang, Y. and Waibel, A. (2001). Toward Automatic Sign Translation, Proceedings of the Human Language Technology Conference (HLT-2001). find
  19. Yang, J., Chen, X., Zhang, J., Zhang Y. and Waibel, A. (2002). Automatic detection and translation of text from natural scenes, Proceedings of ICASSP 2002, (pp. 2101 – 2104), Orlando, Florida. find
  20. Zhang, Y., Vogel, S. (2007). PanDoRA: A Large-scale Two-way Statistical Machine Translation System for Hand-held Devices. In Proceedings of MT Summit XI, Copenhagen, Denmark. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chang-2007-enhancing-image-based-arabic-document/