Elamur Research⌕
☰
home / papers / 2009 / Amharic Part-of-Speech Tagger for Factored Language Modeling

Amharic Part-of-Speech Tagger for Factored Language Modeling

Paper · 2009 · Martha Yifiru Tachbelie and Wolfgang Menzel

Summary

Amharic is a Semitic language with root-and-pattern morphology and heavy affixation. This makes word-based language models suffer from sparse data and many unseen words. Earlier morpheme-based models lowered the out-of-vocabulary rate but lost dependencies between whole words. The authors therefore use factored language models (FLMs), and to give these models part-of-speech information they first build Amharic POS taggers.

They train HMM taggers with TnT and SVM taggers with SVMTool on the ELRC corpus from "The Annotation of Amharic News Documents" project. The corpus has about 210,000 manually tagged news tokens and uses a 30-tag set built on 11 basic classes. The best SVM tagger (85.50%) beats the best TnT tagger (82.99%), although TnT is faster and uses less memory. The SVM tagger then labels training data for FLMs that predict each word from the one or two preceding words and their POS. When these FLMs rescore lattices in an Amharic speech recognizer, word accuracy rises by 1.32 points absolute. This shows that automatically assigned POS tags can strengthen language models for a morphologically rich, low-resource language.

Key points

  • Corpus: ELRC "Annotation of Amharic News Documents", about 210,000 tokens. The tagset has 30 tags over 11 basic classes, with separate tags for nouns, pronouns, verbs, adjectives and numerals that carry attached prepositions and/or conjunctions. There are no distinct tags for proper or plural nouns, and the corpus tags collocations inconsistently.

  • Taggers: the best TnT (HMM) tagger reaches 82.99% and the best SVMTool tagger 85.50%. TnT is cheaper in speed and memory.

  • Factored data: each word is a bundle of the word itself, its POS, prefix, root, pattern and suffix. The FLMs use a fixed backoff that drops the most distant factor first.

  • Perplexities:

    ModelPerplexity
    Factored word-bigram baseline58.41
    Two-parent FLM115.89
    Four-parent FLM17.03

  • ASR lattice rescoring (100-best lattices on a 5k development set, built with HTK and SRILM):

    SystemWord accuracy
    Original baseline91.67%
    Factored baseline91.60%
    + two-parent FLM92.92%
    + four-parent FLM92.75%

Interesting ideas

  • I1. Treat a word as a vector of parallel factors (surface form, POS, root, pattern, affixes) so a model can back off to coarser factors when the full word is rare, instead of splitting text into isolated morphemes and losing word-level links.
  • I2. Before comparing models, they build a "same model, new representation" baseline: a word bigram trained on the factored data. This exposed a small drop caused only by different unigram discounting.
  • I3. The tagger is judged by whether its output improves a downstream task (speech recognition), not just by its tagging accuracy.
  • I4. The four-parent model had by far the lowest perplexity but did not beat the two-parent model on word accuracy, a reminder that intrinsic metrics can mislead.
  • I5. They chose the slower but more accurate SVM tagger for offline corpus labelling and noted TnT as the cheaper option, picking the tool by how it is used.

For Mova

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

  • I1 (tokens): The tagger and morphology layer could keep each token as an explicit factor bundle (form, lemma, UPOS, features), so rare or unseen words fall back to POS and lemma evidence in a traceable way.
  • I2 (learn): Every new representation or feature should get an equivalence control that should match the old system exactly, so artefacts are separated from real gains.
  • I3 (learn): Tagger or parser changes could also be scored by their effect on frames and answer accuracy, catching cases where a higher intrinsic score does not help the end task.
  • I4 (learn): Evaluation reports should show intrinsic and task-level metrics side by side and treat disagreement between them as a signal to investigate.
  • I5 (tree): When producing silver treebanks offline, the heavier gparse parser could be preferred, keeping the fast perceptron parser for interactive runs.

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

🎓Universität HamburgHamburg · Germany79 works · 5,422 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.Language models & pretraining · 3,320 worksSemitic languages · 2,908 worksPart-of-speech tagging · 126 works

Cite

Cite as Martha Yifiru Tachbelie and Wolfgang Menzel. 2009. Amharic Part-of-Speech Tagger for Factored Language Modeling. Proceedings of the International Conference RANLP-2009.

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

@inproceedings{tachbelie2009amharic,
  title = {{Amharic Part-of-Speech Tagger for Factored Language Modeling}},
  author = {Martha Yifiru Tachbelie and Wolfgang Menzel},
  booktitle = {Proceedings of the International Conference RANLP-2009},
  year = {2009},
  url = {https://aclanthology.org/R09-1077/}
}

References

20 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. S. T. Abate. Automatic Speech Recognition for Amharic. PhD thesis, Univ. of Hamburg, 2006. find
  2. S. T. Abate, W. Menzel, and B. Tafila. An Amharic speech corpus for large vocabulary continuous speech recognition. In Proceedings of 9th. European Confference on Speech Communication and Technology, Interspeech-2005, 2005. find
  3. S. F. Adafre. Part of speech tagging for Amharic using conditional random fields. In Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages, pages 47–54, 2005.
  4. M. Bender, J. Bowen, R. Cooper, and C. Ferguson. Languages in Ethiopia. Oxford Univ. Press, London, 1976. find
  5. T. Brants. TnT — a statistical part-of-speech tagger. In Proceedings of the 6th ANLP, 2000. — TnT – A Statistical Part-of-Speech Tagger ↗
  6. G. A. Demeke and M. Getachew. Manual annotation of Amharic news items with part-of-speech tags and its challenges. ELRC Working Papers, II(1), 2006. find
  7. K. Duh and K. Kirchhoff. Automatic learning of language model structure. In Proceeding of International Conference on Computational Linguistics, 2004. find
  8. B. Gambäck, F. Olsson, A. A. Argaw, and L. Asker. Methods for Amharic part-of-speech tagging. In Proceedings of the EACL Workshop on Language Technologies for African Languages - AfLaT 2009, pages 104–111, March 2009.
  9. M. Getachew. Automatic part of speech tagging for Amharic language: An experiment using stochastic hmm. Master’s thesis, Addis Ababa University, 2000. find
  10. J. Giménez and L. Màrquez. Svmtool: A general pos tagger generator based on support vector machines. In Proceedings of the 4th International Conference on Language Resources and Evaluation, 2004.
  11. D. S. Jurafsky and J. H. Martin. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall, New Jersey, 2nd. ed. edition, 2008. find
  12. K. Kirchhoff, J. Bilmes, S. Das, N. Duta, M. Egan, G. Ji, F. He, J. Henderson, D. Liu, M. Noamany, P. Schone, R. Schwartz, and D. Vergyri. Novel approaches to Arabic speech recognition: Report from the 2002 johns-hopkins summer workshop. In Proceedings of International Conference on Acoustics, Speech, and Signal Processing, volume 1, pages 1–344 – 1–347, 2003. find
  13. K. Kirchhoff, J. Bilmes, J. Henderson, R. Schwartz, M. Noamany, P. Schone, G. Ji, S. Das, M. Egan, F. He, D. Vergyri, D. Liu, and N. Duta. Novel speech recognition models for arabic. Technical report, Johns-Hopkins University Summer Research Workshop, 2002. find
  14. K. Kirchhoff, J. Bilmes, and kevin Duh. Factored language models - a tutorial. Technical report, Dept. of Electrical Eng., Univ. of Washington, 2008. find
  15. B. Megyesi. Comparing data-driven learning algorithms for pos tagging of Swedish. In Proceedings of the 2001 Conference on Emperical Methods in Natural Language Processing, pages 151–158, 2001. find
  16. A. Stolcke. SRILM — an extensible language modeling toolkit. In Proceedings of International Conference on Spoken Language Processing, volume II, pages 901–904, 2002. find
  17. M. Y. Tachbelie and W. Menzel. Sub-word based language modeling for Amharic. In Proceedings of International Conference on Recent Advances in Natural Language Processing, pages 564–571, September 2007. find
  18. D. Vergyri, K. Kirchhoff, K. Duh, and A. Stolcke. Morphology-based language modeling for Arabic speech recognition. In Proceedings of International Conference on Spoken Language Processing, pages 2245–2248, 2004. find
  19. B. Yemam. yäamarIña säwasäw. EMPDE, Addis Ababa, 2nd. ed. edition, 2000 EC. find
  20. S. Young, G. Evermann, M. Gales, T. Hain, D. Kershaw, X. Liu, G. Moore, J. Odell, D. Ollason, D. Povey, V. Valtchev, and P. Woodland. The HTK Book. Cambridge University Engineering Department, 2006. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tachbelie-2009-amharic-part-speech-tagger-factored/