Elamur Research⌕
☰
home / papers / 2006 / Bridging the Inflection Morphology Gap for Arabic Statistical Machine Translation

Bridging the Inflection Morphology Gap for Arabic Statistical Machine Translation

Paper · 2006 · Andreas Zollmann, Ashish Venugopal and Stephan Vogel

Summary

When the source language carries much more inflection than the target, statistical MT struggles. A single Arabic lemma surfaces in many forms, which inflates vocabulary, multiplies singletons and unknown words, and weakens IBM word alignment and phrase extraction. Zollmann, Venugopal and Vogel address this with source-side preprocessing only. They leave the alignment models untouched and let the decoder handle reordering. The Buckwalter Arabic Morphological Analyzer (BAMA) proposes candidate prefix/stem/suffix splits for each word. The authors pick among these candidates by checking which split's lexicon glosses best match the English side of the parallel sentence. A further step deletes fragments that have no English gloss, such as case and gender markers. On the BTEC IWSLT05 Arabic–English supplied track, the full pipeline gives statistically significant gains over a normalization baseline, with NIST improving by up to about 5% relative. The work shows that a cheap word-level analyzer with glosses can stand in for heavy context-sensitive parsing when MT has to be built quickly.

Key points

  • Data: BTEC IWSLT05 Arabic "Supplied" track, with 20K training sentence pairs, DevSet and Test05 of 500 sentences each, and 16 English references per sentence. The authors also report results on a random resplit of the dev and test pool.
  • Pipeline stages:
    • BAMA: take the analyzer's top-ranked split.
    • CONTEXT: choose the split whose glosses best match the target sentence, then fix that choice per surface form for both training and test.
    • CORRMATCH: remove fragments with empty English glosses.
  • BAMA version 2.0 returns roughly a dozen candidate segmentations per word. Unseen test words fall back to the first candidate.
  • System setup: the Vogel et al. (2003) decoder with reordering and future cost estimation, using 10 feature scores. These were tuned by MERT separately for BLEU and NIST.
  • Results: NIST gains of up to about 5% relative, significant in both metrics. NIST improved more than BLEU, which the authors attribute to rare, high-gain words now being translated correctly. The final NIST scores would exceed the best IWSLT supplied-track results.

Interesting ideas

  • I1. An analyzer's competing segmentations can be ranked using an independent aligned signal, here the translation, by matching each candidate's glosses against it, with no annotated segmentation data needed.
  • I2. The disambiguation decision is made once per surface form and then reused everywhere, with a transparent fallback to the analyzer's default for unseen forms, which keeps behaviour deterministic and cheap.
  • I3. Morphemes whose lexicon translation is empty are treated as information the target language does not express, and are dropped rather than forced into alignment.
  • I4. Results are reported on both the official split and a random resplit, because the official dev and test sets may be too similar, a guard against overly optimistic tuning.
  • I5. Different metrics can disagree because one rewards rare, informative words more heavily, so comparing them shows where an improvement comes from.

For Mova

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

  • I1 (tokens): When the English morphology module returns several lemma or POS analyses, rank them by how well their WordNet glosses or senses overlap with nearby words, and log the overlap as the explanation.
  • I2 (tokens): A per-form analysis cache with an explicit "default because unseen" flag would make lemma choices reproducible and auditable across corpora.
  • I3 (back): When generating or translating through the meta-world, mark graph features with no realization in the target language so the generator omits them deliberately instead of inventing words.
  • I4 (learn): Add a reshuffled evaluation split next to fixed test sets, so a gain that appears only on the official split is flagged as possible overfitting.
  • I5 (learn): Report a rare-word-weighted score alongside the usual accuracy, so it is visible when a change mainly fixes previously unknown words.

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

🎓Carnegie Mellon UniversityPittsburgh · United States539 works · 179,533 citations

 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 Andreas Zollmann, Ashish Venugopal and Stephan Vogel. 2006. Bridging the Inflection Morphology Gap for Arabic Statistical Machine Translation. Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers.

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

@inproceedings{zollmann2006bridging,
  title = {{Bridging the Inflection Morphology Gap for Arabic Statistical Machine Translation}},
  author = {Andreas Zollmann, Ashish Venugopal and Stephan Vogel},
  booktitle = {Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers},
  year = {2006},
  doi = {10.3115/1614049.1614100},
  url = {https://doi.org/10.3115/1614049.1614100}
}

References

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

  1. Peter F. Brown, Vincent J. Della Pietra, Stephen A. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: parameter estimation. Comput. Linguist., 19(2):263–311. find
  2. Tim Buckwalter. 2004. Buckwalter Arabic Morphological Analyzer Version 2.0. LDC Catalog No. LDC2004L02, Linguistic Data Consortium, www.ldc.upenn.edu/Catalog. Tim Buckwalter. www.qamus.org/morphology.htm.2005. ↗
  3. George Doddington. 2002. Automatic evaluation of machine translation quality using n-gram co-occurrence statistics. In In Proc. ARPA Workshop on Human Language Technology. find
  4. Matthias Eck and Chiori Hori. 2005. Overview of the IWSLT 2005 evaluation campaign. In Proceedings of International Workshop on Spoken Language Translation, pages 11–17. find
  5. Jose B.Marino Josep M.Crego, Adria de Gispert. 2005. The talp ngram-based smt system for iwslt’05. In Proceedings of International Workshop on Spoken Language Translation, pages 191–198. Inflection system No preprocessing Orthographic normalization (baseline) BAMA BAMA+CONTEXT+CORRMATCH NIST – Dev. 9.33 9.41 9.90 9.91 Table 1: Translation results for each stage of inflectional splitting for the merged, sampled dev. and test data, highest scores in bold, relative improvements in brackets Inflection system No preprocessing Orthographic normalization (baseline) BAMA BAMA+CONTEXT+CORRMATCH NIST – Dev. 9.46 9.58 10.10 10.08 Table 2: Translation results for each stage of inflectional splitting for the BTEC Supplied DevSet/Test05 data, highest scores in bold, relative improvements in brackets Leah Larkey, Lisa Ballesteros, and Margaret Connell. 2002. Improving stemming for arabic information retrieval: Light stemming and co-occurrence analysis. In Proc. of the 25th annual international ACM SIGIR conference on Research and development information retrieval. find
  6. Young-Suk Lee, Kishore Papineni, Salim Roukos, Ossama Emam, and Hany Hassan. 2003. Language model based arabic word segmentation. In ACL, Sapporo, Japan, July 6-7.
  7. Young-Suk Lee. 2004. Morphological analysis for statistical machine translation. In Proceedings of the Human Language Technology and North American Association for Computational Linguistics Conference (HLT/NAACL), Boston,MA, May 27-June 1. ↗
  8. Sonja Nießen and Hermann Ney. 2000. Improving SMT quality with morpho-syntactic analysis. In The 18th International Conference on Computational Linguistics. find
  9. Sonja Nießen and Herman Ney. 2004. Statistical machine translation with scarce resources using morphosyntactic information. Comput. Linguist., 30(2):181– 204. find
  10. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the Association of Computational Linguistics, pages 311– 318.
  11. H. Popovic and Hermann Ney. 2004. Improving word alignment quality using morpho-syntactic information. In 20th International Conference on Computational Linguistics (CoLing), Geneva, Switzerland. find
  12. Toshiyuki Takezawa, Eiichiro Sumita, Fumiaki Sugaya, Hirofumi Yamamoto, and Seiichi Yamamoto. 2002. Toward a broad-coverage bilingual corpus for speech NIST – Test 9.44 9.51 9.76 (+2.5%) 10.02 (+5.3%) BLEU – Test 49.7 49.7 50.2 (+1%) 52.0 (+4.7%) BLEU – Dev. 51.1 51.0 52.0 52.8 NIST – Test 9.38 9.35 9.60 (+2.7%) 9.79 (+4.7%) BLEU – Test 49.6 49.8 48.8 (-2%) 50.6 (+1.6%) BLEU – Dev. 51.1 52.1 53.8 53.7 translation of travel conversations in the real world. In Proc. of LREC 2002, pages 147–152, Las Palmas, Canary Islands, Spain, May. find
  13. Stephan Vogel, Ying Zhang, Fei Huang, Alicia Tribble, Ashish Venogupal, Bing Zhao, and Alex Waibel. 2003. The CMU statistical translation system. In Proceedings of MT Summit IX, New Orleans, LA, September. find
  14. Stephan Vogel. 2005. PESA: Phrase pair extraction as sentence splitting. In Proceedings of MT Summit X, Phuket,Thailand, September. find
  15. Ying Zhang and Stephan Vogel. 2004. Measuring confidence intervals for the machine translation evaluation metrics. In Proceedings of the 10th International Conference on Theoretical and Methodological Issues in Machine Translation (TMII), Baltimore, MD, October. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zollmann-2006-bridging-inflection-morphology-gap-arabic/