Elamur Research⌕
☰
home / papers / 2010 / MorphTagger: HMM-based Arabic segmentation for statistical machine translation

MorphTagger: HMM-based Arabic segmentation for statistical machine translation

Paper · 2010 · Saab Mansour

Summary

Arabic words often pack prepositions, pronouns and other clitics onto a stem. This makes statistical machine translation data sparse, raises the number of unknown words and complicates word alignment. Splitting words into morphemes before translation helps, but existing tools force a trade-off. A rule-based FST segmenter is fast but falls behind on large tasks. The SVM-based MADA is accurate but too slow for real-time use.

Saab Mansour adapts MorphTagger, an HMM POS-tagging architecture first built for Hebrew, into a segmenter for SMT. The Buckwalter analyzer lists the candidate analyses for each word. A bigram HMM trained at the segment level on the Arabic Treebank chooses among them. Hand-picked rules then decide which morphemes to detach and repair spelling changes caused by attachment.

MorphTagger runs about 20 times faster than MADA while matching or slightly beating it in translation quality. Tests covered Arabic–English BTEC and Arabic–French QUAERO, and the gains held for both small and large tasks. This makes accurate segmentation usable in real-time translation systems.

Key points

  • Measured speed is about 4,500 words/s for FST, 70 for MADA and 1,500 for MorphTagger, putting MorphTagger near FST-level speed.
  • Resources:
    • Buckwalter Arabic Morphological Analyzer v1.0, with about 80,000 lexicon entries.
    • A bigram HMM trained on Arabic Treebank Part 1 v3.0 (about 150,000 tokens).
    • Decoding wrapped around the SRILM toolkit.
  • The segmenter applies the ATB scheme. It splits off prepositions and possessive/object pronouns but keeps the determiner attached. It also restores alif maksura and the feminine marker, and recovers a hidden "Al" after the preposition "l".
  • On BTEC (IWSLT05), MorphTagger reaches 56.8 BLEU / 31.5 TER, against 55.6 / 32.4 for plain tokenization. On IWSLT08 it scores 52.4 BLEU, compared with 51.9 for MADA.
  • The paper explains MADA's slowness by SVM tagging cost. That cost grows with the number of support vectors, which can reach roughly 10^5 for a model trained on the treebank.

Interesting ideas

  • I1. The HMM only chooses among analyses proposed by a rule-based lexicon, so every output traces back to a listed analysis plus a probability.
  • I2. Probabilities are estimated per segment rather than per whole word, so lattice paths of different lengths can compete fairly. Frequent clitics contribute little to the score, which leaves the stem as the main deciding factor.
  • I3. The design keeps disambiguation (picking the full analysis) separate from the segmentation scheme, which is a small manual rule layer that also undoes spelling changes at morpheme boundaries.
  • I4. The segmenter is judged by end-task translation quality on several small and large tasks, alongside a speed and complexity analysis. Judging it only by its own accuracy would miss both of these factors.

For Mova

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

  • I1 (tokens): The en tagger could limit its candidates for each word to analyses from a morphological lexicon. This would make each tag choice explainable as "lexicon allowed X, context preferred it."
  • I2 (tokens): Clitic-like English pieces ("'s", "n't", "'ll") could be modeled as segments inside a lattice. The tokenizer and tagger would then resolve ambiguous splits together instead of using fixed regex rules.
  • I3 (tokens): Splitting and normalization policy (e.g. "cannot", "gonna", possessives) could live in a configurable rule layer on top of one tagger. Different UD dialects or downstream stages could then select their own scheme without retraining.
  • I4 (learn): Tokenizer and parser variants could be evaluated by their effect on frames or answer accuracy and by throughput at the same time. This would catch components that score well in isolation but hurt the pipeline or the latency budget.

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, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 3Research subjects and languages the work is about.Machine translation · 1,087 worksSemitic languages · 2,908 worksMorphology · 405 works

Cite

Cite as Saab Mansour. 2010. MorphTagger: HMM-based Arabic segmentation for statistical machine translation. Proceedings of the 7th International Workshop on Spoken Language Translation: Papers.

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

@inproceedings{mansour2010morphtagger,
  title = {{MorphTagger: HMM-based Arabic segmentation for statistical machine translation}},
  author = {Saab Mansour},
  booktitle = {Proceedings of the 7th International Workshop on Spoken Language Translation: Papers},
  year = {2010},
  url = {https://aclanthology.org/2010.iwslt-papers.15/}
}

References

9 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Y.-S. Lee, “Morphological Analysis for Statistical Machine Translation,” in Proceedings of HLT-NAACL 2004: Short Papers. Morristown, NJ, USA, 2004, pp. 57–60. ↗
  2. F. Sadat and N. Habash, “Combination of Preprocessing Schemes for Statistical MT,” in Proceedings of the 44th Annual Meeting of the Association for Computational Linguistics, Sydney, Australia, July 2006, pp. 1–8. find
  3. A. El Isbihani, S. Khadivi, O. Bender, and H. Ney, “Morpho-Syntactic Arabic Preprocessing for Arabic to English Statistical Machine Translation,” in Proceedings of the Workshop on Statistical Machine Translation. New York City, June 2006, pp. 15–22.
  4. S. Hasan, A. El Isbihani, and H. Ney, “Creating a Large-Scale Arabic to French Statistical Machine Translation System,” in International Conference on Language Resources and Evaluation, Genoa, Italy, May 2006, pp. 855–858. find
  5. M. Diab, K. Hacioglu, and D. Jurafsky, “Automatic Tagging of Arabic Text: From Raw Text to Base Phrase Chunks,” in HLT-NAACL 2004: Short Papers. Boston, Massachusetts, USA, May 2 - May 7 2004, pp. 149–152.
  6. T. Nguyen and S. Vogel, “Context-Based Arabic Morphological Analysis for Machine Translation,” in Proceedings of the Twelfth Conference on Computational Natural Language Learning. Morristown, NJ, USA, 2008, pp. 135–142.
  7. R. Bar-haim, K. Sima’an, and Y. Winter, “Part-of-Speech Tagging of Modern Hebrew Text,” Nat. Lang. Eng., vol. 14, no. 2, pp. 223–251, 2008. ↗
  8. S. Manour, K. Sima’an, and Y. Winter, “Smoothing a Lexicon-Based POS Tagger for Arabic and Hebrew,” in Proceedings of the 2007 Workshop on Computational Approaches to Semitic Languages: Common Issues and Resources. Prague, Czech Republic: Association for Computational Linguistics, June 2007, pp. 97–103.
  9. R. Zens and H. Ney, “Improvements in Dynamic Programming Beam Search for Phrase-based Statistical Machine Translation,” in International Workshop on Spoken Language Translation, Honolulu, Hawaii, Oct. 2008, pp. 195–205. Table 4: AR-FR QUAERO 2009: Translation results (case) Real-time systems Dev Test BLEU TER BLEU TER 15.5 74.9 15.4 74.8 15.5 73.9 15.5 74.8 15.9 73.9 15.8 74.7 Offline systems 15.7 74.6 15.3 75.3 16.6 73.2 16.3 74.2 16.1 73.7 16.1 74.9 17.1 72.5 16.6 73.5 System FST MADA MorphTagger TOK FST MADA MorphTagger Table 5: AR-EN BTEC 2008: Translation results (nocase+punc) IWSLT04 (dev) BLEU TER 55.6 32.8 52.3 35.2 56.0 32.4 55.8 32.7 System TOK FST MADA MorphTagger Table 6: Examples of better translations due to improved Arabic segmentation  » èPñJ»YËAK. I.kP @ à @ Xð @ AÒ» QKPA¿ IJ Je voudrais aussi souhaiter la bienvenue au Dr Keith Carter  » è PñJ»YËAK. I.kP@ à@ Xð@ AÒ» QKPA¿ IJ  » Carter Je souhaite la bienvenue au Dr UNKNOWN IJ     Q KPA¿ IJ » è Pñ J»YË@ H. I.kP @ à @ Xð @ AÒ» Source Reference FST MADA Je souhaite la bienvenue au Dr Keith Carter  » èPñJ»YË@ H. I.kP @ à @ Xð @ AÒ» QKPA¿ IJ MorphTagger Je souhaite la bienvenue au Dr Keith Carter Z@P PñË@ Êm.× ‡  ¯@ð 2001 Q.҂ X / Èð B@ àñ KA¿ ú ¯ð Source Reference En décembre 2001 , le Conseil des ministres a approuvé Z@P PñË@ Êm.× ‡  ¯@ð décembre 2001 ú ¯ð FST En décembre , le conseil des ministres a approuvé Z@P PñË@ Êm.× ‡  ¯@ð 2001 Q.҂ X / Èð B@ àñ KA¿ ú ¯ð MADA Acquitté en décembre 2001 , le Conseil des ministres an approuvé Z@P PñË@ Êm.× ‡  ¯@ð 2001 Q.҂ X / Èð B@ àñ KA¿ ú ¯ ð MorphTagger En décembre 2001 , le Conseil des ministres an approuvé Proceedings of the 7 th International Workshop on Spoken Language Translation Paris, December 2nd and 3rd, 2010 CESTA RUN2 BLEU TER 45.7 53.4 47.7 53.0 48.0 53.2 45.3 47.6 47.8 48.8 53.6 52.1 51.7 49.8 IWSLT05 BLEU TER 55.6 32.4 55.9 32.0 56.7 31.6 56.8 31.5 IWSLT08 BLEU TER 51.8 35.0 51.7 34.9 51.9 35.2 52.4 34.7 find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mansour-2010-morphtagger-hmm-based-arabic-segmentation/