Elamur Research⌕
☰
home / papers / 2017 / Traitement des Mots Hors Vocabulaire pour la Traduction Automatique de Document OCRisés en Arabe (This article presents a new system that automatically translates images of Arabic documents)

Traitement des Mots Hors Vocabulaire pour la Traduction Automatique de Document OCRisés en Arabe (This article presents a new system that automatically translates images of Arabic documents)

Paper · 2017 · Kamel Bouzidi, Zied Elloumi, Laurent Besacier et al.

Summary

The authors tackle machine translation of scanned Arabic newspapers. These are images, not clean text, so recognition mistakes produce garbled tokens that a translation system has never seen. They chain an Arabic OCR module with an Arabic→French statistical MT system and couple the two closely in two ways. First, instead of passing only the single best OCR reading, they pass a word lattice of competing OCR hypotheses into the translator. Second, they treat out-of-vocabulary (OOV) tokens specially: each unknown token is replaced with a known in-vocabulary word chosen by a composite score. That score mixes spelling similarity with contextual similarity derived from word embeddings. The MT system is trained on large Arabic-French parallel corpora. Evaluation uses OCRed newspaper pages from the MAURDOR collection. Together, the lattice and the OOV replacement raise BLEU by 3.73 on the development set and 5.5 on the test set. This matters for cross-lingual search over paper archives, where errors made by early pipeline stages otherwise spread into every later stage.

Key points

  • The work is part of the TRIDAN project, which targets cross-lingual information retrieval over digitized paper documents. This paper covers only the OCR→MT link.
  • Training data: the parallel corpora are MultiUN (about 9.9M sentence pairs), News Commentary, Opensub, Trame and Wit3. Europarl supplies extra French monolingual text for the target language model.
  • Dev/test data are Arabic newspaper images from MAURDOR. Each page comes with a gold transcription, an automatic OCR output with roughly 15% word error rate, and two independent human French translations used as references.
  • OOVs fall into two groups: OCR misrecognitions and genuine coverage gaps of the MT training data. Tokens that have no embedding at all cannot be handled by this method.
  • Reported gains are +3.73 BLEU on dev and +5.5 BLEU on test.

Interesting ideas

  • I1. OCR errors usually sit only one or two characters away from the intended word, so a noisy unknown token can be repaired by snapping it to a close known word. Speech recognition errors are less suited to this kind of fix.
  • I2. Candidates are ranked by a combined score of surface edit similarity and distributional context similarity. Neither signal alone is enough: spelling picks out lookalikes, and context removes the implausible ones.
  • I3. Rather than committing to one upstream reading, the full lattice of alternatives is handed to the next stage, which then resolves the ambiguity with its own knowledge.
  • I4. The authors explicitly separate OOVs caused by upstream noise from OOVs caused by missing training coverage. The two need different remedies.
  • I5. Each test item has two independent human references, which makes the evaluation less sensitive to any single translator's choices.

For Mova

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

  • I1 (tokens): Before tagging, an OOV repair step could map misspelled or garbled tokens to lexicon entries, and it should record the substitution so the step stays explainable.
  • I2 (tokens): Lemma and morphology lookup for unknown words could rank candidates by both character distance and fit with the surrounding sentence, keeping each score component visible.
  • I3 (tree): The parser could accept several tokenization or tagging hypotheses and choose the one that yields the best-scoring tree, instead of inheriting early mistakes.
  • I4 (learn): Error analysis could label each unknown-word failure as either input noise or a lexicon gap, so that repair rules and data collection target the right cause.
  • I5 (back): Generation and translation through the meta-world could be scored against multiple reference wordings, so that valid paraphrases are not counted as errors.

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

 Institutions

🎓Université Grenoble AlpesSaint-Martin-d'Hères · France24 works · 8,259 citations🔬Laboratoire d'Informatique de GrenobleGrenoble · France24 works · 689 citations

Also on the byline: Laboratoire National de Métrologie et d'Essais.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 3,066 works

Cite

Cite as Kamel Bouzidi, Zied Elloumi, Laurent Besacier, Benjamin Lecouteux and Mohamed-Faouzi Benzeghiba. 2017. Traitement des Mots Hors Vocabulaire pour la Traduction Automatique de Document OCRisés en Arabe (This article presents a new system that automatically translates images of Arabic documents). Actes des 24ème Conférence sur le Traitement Automatique des Langues Naturelles. Volume 1 - Articles longs.

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

@inproceedings{bouzidi2017traitement,
  title = {{Traitement des Mots Hors Vocabulaire pour la Traduction Automatique de Document OCRisés en Arabe (This article presents a new system that automatically translates images of Arabic documents)}},
  author = {Kamel Bouzidi, Zied Elloumi, Laurent Besacier, Benjamin Lecouteux and Mohamed-Faouzi Benzeghiba},
  booktitle = {Actes des 24ème Conférence sur le Traitement Automatique des Langues Naturelles. Volume 1 - Articles longs},
  year = {2017},
  url = {https://aclanthology.org/2017.jeptalnrecital-long.5/}
}

References

39 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Afli, H., & Way, A. (2016). Integrating Optical Character Recognition and Machine Translation of Historical Documents. LT4DH 2016, 109. find
  2. Afli, H., Barrault, L., & Schwenk, H. (2015). OCR Error Correction Using Statistical Machine Translation. In 16th International Conference on Intelligent Text Processing and Computational Linguistics, Cairo, Egypt. find
  3. Baroni, M., & Zamparelli, R. (2010). Nouns are vectors, adjectives are matrices: Representing adjective-noun constructions in semantic space. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing (pp. 1183-1193). Association for Computational Linguistics. find
  4. Bazzi, I. (2002). Modelling out-of-vocabulary words for robust speech recognition (Doctoral dissertation, Massachusetts Institute of Technology). find
  5. Belinkov, Y., Mohtarami, M., Cyphers, S., & Glass, J. (2015). VectorSLU: A continuous word vector approach to answer selection in community question answering systems. In Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015) (pp. 282-287). find
  6. Bengio, Y., Ducharme, R., Vincent, P., & Jauvin, C. (2003). A neural probabilistic language model. Journal of machine learning research, 3 (Feb), 1137-1155.
  7. BenZeghiba, M. F., Louradour, J., & Kermorvant, C. (2015, August). Hybrid word/Part-of-Arabic-Word Language Models for arabic text document recognition. In Document Analysis and Recognition (ICDAR), 2015 13th International Conference on (pp. 671-675). IEEE. find
  8. Bérard, A., Servan, C., Pietquin, O., & Besacier, L. (2016). Multivec: a multilingual and multilevel representation learning toolkit for NLP. In The 10th edition of the Language Resources and Evaluation Conference (LREC). find
  9. Besacier, L., Afli, H., Do, T. N. D., Blanchon, H., & Potet, M. (2010). LIG statistical machine translation systems for IWSLT 2010. In International Workshop on Spoken Language Translation IWSLT 2010). find
  10. Blacoe, W., & Lapata, M. (2012). A comparison of vector-based representations for semantic composition. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (pp. 546-556). Association for Computational Linguistics.
  11. Bousquet-Vernhettes, C. (2002). Traitement des mots mal reconnus en compréhension de la parole. Journées d'Études sur la Parole (JEP'2002), 309-312. find
  12. Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078. ↗
  13. Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., & Kuksa, P. (2011). Natural language processing (almost) from scratch. Journal of Machine Learning Research, 12(Aug), 2493-2537.
  14. Déchelotte, D., Schwenk, H., Adda, G., & Gauvain, J. L. (2007). Improved machine translation of speech-to-text outputs. In Interspeech 2007 (Vol. 7, pp. 2441-2444). find
  15. Dyer, C., Muresan, S., & Resnik, P. (2008). Generalizing word lattice translation (No. LAMP-TR-149). Internal Report Maryland University College Park Institute for Advanced Computer Studies. 2008.
  16. Federico, M., Bertoldi, N., & Cettolo, M. (2008). IRSTLM: an open source toolkit for handling large scale language models. In Interspeech 2008 (pp. 1618-1621). find
  17. Graves, A. (2012). Offline arabic handwriting recognition with multidimensional recurrent neural networks. In Guide to OCR for Arabic scripts (pp. 297-313). Springer London. find
  18. Habash, N. (2008). Four techniques for online handling of out-of-vocabulary words in Arabic-English statistical machine translation. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics on Human Language Technologies: Short Papers (pp. 57-60). Association for Computational Linguistics. find
  19. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735-1780. find
  20. Huang, E. H., Socher, R., Manning, C. D., & Ng, A. Y. (2012). Improving word representations via global context and multiple word prototypes. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers-Volume 1 (pp. 873-882). Association for Computational Linguistics.
  21. Koehn, P., Hoang, H., Birch, A., Callison-Burch, C., Federico, M., Bertoldi, N., Dyer, C. (2007). Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration sessions (pp. 177-180). Association for Computational Linguistics. find
  22. Kombrink, S., Hannemann, M., & Burget, L. (2012). Out-of-vocabulary word detection and beyond. In Detection and Identification of Rare Audiovisual Cues (pp. 57-65). Springer Berlin Heidelberg. find
  23. Lecouteux, B., Linarès, G., & Favre, B. (2009). Détection de mots hors-vocabulaire par combinaison de mesures de confiance de haut et bas niveaux. MajecSTIC'09. find
  24. Macrostie, E., Rawls, S., and Subramanian, K. (2010). The BBN Document Analysis Service : A Platform for Multilingual Document Translation. In Document Analysis Systems, pages 447-453, 2010. find
  25. Marton, Y., Callison-Burch, C., & Resnik, P. (2009). Improved statistical machine translation using monolingually-derived paraphrases. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1-Volume 1 (pp. 381-390). Association for Computational Linguistics.
  26. Matusov, E., Leusch, G., Bender, O., & Ney, H. (2005). Evaluating machine translation output with automatic sentence segmentation. In IWSLT 2005 (pp. 138-144).
  27. Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems (pp. 3111-3119).
  28. Mirkin, S., & Besacier, L. (2014). Data selection for compact adapted SMT models. In Eleventh Conference of the Association for Machine Translation in the Americas (AMTA). find
  29. Mohri, M. (1997). Finite-state transducers in language and speech processing. Computational linguistics, 23(2), 269-311.
  30. Moysset, B., Bluche, T., Knibbe, M., Benzeghiba, M. F., Messina, R., Louradour, J., & Kermorvant, C. (2014). The A2iA multi-lingual text recognition system at the second Maurdor evaluation. In Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on (pp. 297-302). IEEE. find
  31. Och, F. J. (2003). Minimum error rate training in statistical machine translation. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics-Volume 1 (pp. 160-167). Association for Computational Linguistics. find
  32. Oprean, C., Likforman-Sulem, L., Popescu, A., & Mokbel, C. (2014). Utilisation du Web pour la reconnaissance de mots manuscrits hors vocabulaire. In CORIA-CIFED (pp. 217-232). find
  33. Papineni, K., Roukos, S., Ward, T., and Zhu, W. J. (2002). "BLEU: a method for automatic evaluation of machine translation" in ACL-2002: 40th Annual meeting of the Association for Computational Linguistics pp. 311–318.
  34. Segal N., Bonneau-Maynard H., Yvon F. (2015). « Traduire la parole : le cas des TED Talks », Traitement automatique des langues (TAL), 2015. find
  35. Servan, C., Elloumi, Z., Blanchon, H., & Besacier, L. (2016). Word2Vec vs DBnary ou comment (ré) concilier représentations distribuées et réseaux lexico-sémantiques? Le cas de l’évaluation en traduction automatique. In TALN 2016. find
  36. Shen, Y., He, X., Gao, J., Deng, L., & Mesnil, G. (2014). A latent semantic model with convolutional-pooling structure for information retrieval. In Proceedings of the 23rd ACM International Conference on Conference on Information and Knowledge Management (pp. 101-110). ACM. find
  37. Stolcke, A. (2002, September). SRILM - an extensible language modeling toolkit. In Interspeech 2002 (Vol. 2002, p. 2002). find
  38. Turian, J., Ratinov, L., & Bengio, Y. (2010). Word representations: a simple and general method for semi-supervised learning. In Proceedings of the 48th annual meeting of the association for computational linguistics (pp. 384-394). Association for Computational Linguistics.
  39. Zhou, B., Besacier, L., & Gao, Y. (2007). On efficient coupling of ASR and SMT for speech translation. In Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE International Conference on (Vol. 4, pp. IV-101). IEEE. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bouzidi-2017-traitement-des-mots-hors-vocabulaire/