Elamur Research
☰
home / papers / 2008 / Learning Morphology with Morfette

Learning Morphology with Morfette

What it is about

Morfette learns morphological tagging and lemmatization jointly from annotated corpora. Two modules, each a maximum entropy classifier, predict the morphological tags and the lemmas. A third module combines their predictions into a probability distribution over tag-lemma pairs for each sentence. Lemmatization is recast as classification: each class label encodes how a word form maps to its lemma, as a change of prefix and suffix. The system is tested on three morphologically rich languages, Romanian, Spanish and Polish, with no language-specific feature engineering. The lemma rules give a relative error reduction of 62% to 66.5% over a baseline for all words, and the paper reports that the baseline is clearly worse on unseen words.

Ideas

For Mova

The code is en/src/morph.rs and en/src/ptag.rs, with the commands en morph-eval and en tag-eval. The chapter is languages/en/11-nouns.md for plural lemmas and languages/en/13-verbs.md for verb forms. The paper's languages are not English and its corpora are LREC data under NC, so use it only for the method and test on EWT test. Measure first with a two-fold check (en folds).

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC 4.0 (ELRA, LREC proceedings)" does not allow redistributing the text itself. Original — aclanthology.org/L08-1176.pdf

Summary

Morfette addresses data-driven morphological analysis for inflection-rich languages. It predicts both a morphological tag and a lemma for every token and does not rely on a morphological lexicon or a finite-state analyzer. The system has two Maximum Entropy modules: a morpho-tagger and a lemmatizer. The lemmatizer treats lemmatization as classification. Each class is the shortest edit script (SES) that turns the reversed wordform into the reversed lemma. A beam-search decoder then combines the two models' probabilities and scores whole sequences of tag–lemma pairs. The feature set is deliberately small and language-independent. The authors tested the system on Romanian (MULTEXT-East), Spanish (CESS-ECE) and Polish (IPI PAN) without tuning it to any language. Joint accuracy was about 96% for Romanian, about 94–95% for Spanish and about 81–84% for Polish, which clearly beat a memory-based pipeline baseline. An error analysis traces the remaining mistakes to suffix ambiguity, case syncretism, ambiguous function words, annotation noise and prefixal morphology. The work shows that full morphological analysis can be learned from annotated corpora alone.

Key points

  • Data and splits:
    • Romanian MULTEXT-East has about 88k training tokens.
    • Spanish CESS-ECE has about 168k tokens in the FULL set and about 70k in the SMALL set.
    • Polish Korpus Słownika Frekwencyjnego (IPI PAN) has about 219k tokens in FULL and about 70k in SMALL.
    • Test and development sets are roughly 10–13.5k tokens each.
  • Results with SMALL training sets:
    • Joint tag+lemma accuracy is 96.08 for Romanian, 93.83 for Spanish and 81.19 for Polish.
    • On unseen words, joint accuracy drops to 78.50, 71.26 and 59.17.
  • Effect of more data (FULL training sets):
    • Spanish joint accuracy rises to 95.02 (+1.19).
    • Polish joint accuracy rises to 84.44 (+3.25).
    • Lemmatization reaches 98.52 for Spanish and 95.55 for Polish.
  • Language difficulty: Polish shows the most morpho-tag ambiguity, with an average of 2.23 tags per token versus 1.16 and 1.46 for the other two languages. Only 44.44% of Polish tokens have a lemma identical to the wordform.
  • Settings and baseline: All experiments used a beam size of 3 and a pre-pruning threshold of 0.3. The BASELINE (MBT tagger plus a training-data lookup for lemmas) loses 4.3–11.4 points of joint accuracy. Its joint accuracy on unseen words collapses to about 11–32%.

Interesting ideas

  • Lemmatization as tagging via edit scripts. Each lemma is encoded as an edit operation on the reversed string. This lets a standard classifier group words that share an inflectional ending, so it can lemmatize words it has never seen. The same approach could suit Ukrainian, which also inflects mainly through suffixes.
  • Conditioning the lemmatizer on the predicted tag. The lemmatizer uses the candidate morpho-tag as a feature, and the decoder multiplies the tag and lemma probabilities. The two modules stay separately trainable but are still decoded jointly. This is a useful pattern for any pair of dependent tagging tasks.
  • Pruning candidates by cumulative probability. For each beam hypothesis, the decoder keeps only the top-ranked candidates whose probability is still large relative to the running total, controlled by a threshold. This keeps beam search cheap when the label set is large, as with full morpho-tag sets.
  • Known bias of reversed SES toward suffixes. The method breaks on prefix-based inflection, such as Polish superlatives with naj-. Anyone reusing the idea should check for prefixal or circumfixal patterns, for example Ukrainian superlatives with най-, and adjust the class induction accordingly.
  • Language-agnostic evaluation design. The authors use one untuned minimal feature set and report accuracy separately on unseen words. This exposes how well a system generalizes, as opposed to how much it memorizes. Equalizing training sizes (the SMALL sets) makes the cross-language comparison fairer.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.63317/46dhmje42wcr

@inproceedings{chrupala2008learning,
  title = {{Learning Morphology with Morfette}},
  author = {Chrupala, Grzegorz and Dinu, Georgiana and Genabith, Josef van},
  booktitle = {Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC'08)},
  year = {2008},
  doi = {10.63317/46dhmje42wcr},
  url = {https://doi.org/10.63317/46dhmje42wcr}
}

References

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

  1. Alexandru Ceauşu. 2006. Maximum Entropy Tiered Tagging. In Proceedings of ESSLLI 2006 Student Session. find
  2. Grzegorz Chrupała. 2006. Simple data-driven context-sensitive lemmatization. Procesamiento del Lenguaje Natural, (37):121–127. find
  3. Walter Daelemans, Jakub Zavrel, Antal van den Bosch, and Ko van der Sloot. 2007. MBT: Memory-Based Tagger, version 3.1, Reference Guide. ILK Technical Report Series 07-08. find
  4. Tomaž Erjavec and Sašo Džeroski. 2004. Machine learning of morphosyntactic structure: Lemmatizing unknown Slovene words. Applied Artificial Intelligence, 18:17–41. find
  5. Tomaž Erjavec. 2004. MULTEXT-East version 3: Multilingual morphosyntactic specifications, lexicons and corpora. In Fourth Intl. Conf. on Language Resources and Evaluation. find
  6. Nizar Habash and Owen Rambow. 2005. Arabic tokenization, part-of-speech tagging and morphological disambiguation in one fell swoop. Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pages 573–580. find our card
  7. Jan Hajič and Barbora Hladká. 1998. Tagging inflective languages: prediction of morphological categories for a rich, structured tagset. In Proceedings of the 17th international conference on Computational linguistics, pages 483–490. find
  8. Jan Hajič. 2000. Morphological tagging: data vs. dictionaries. In Proceedings of the first conference on North American chapter of the Association for Computational Linguistics, pages 94–101. find
  9. Dilek Z. Hakkani-Tür, Kemal Oflazer, and Gokhan Tür. 2002. Statistical Morphological Disambiguation for Agglutinative Languages. Computers and the Humanities, 36(4):381–410. find
  10. Chung-Hye Han and Martha Palmer. 2004. A morphological tagger for korean: Statistical tagging combined with corpus-based morphological rule application. Machine Translation, 18(4):275–297. find
  11. M. Antonia Martı́, Mariona Taulé, Lluı́s Márquez, and Manuel Bertran. 2007. CESS-ECE: A multilingual and multilevel annotated corpus. Available from: http://www.lsi.upc.edu/˜mbertran/cess-ece. ↗
  12. Eugene W. Myers. 1986. An O(ND) difference algorithm and its variations. Algorithmica, 1(1):251–266, March. find
  13. Adwait Ratnaparkhi. 1996. A maximum entropy model for part-of-speech tagging. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. find
  14. Dan Tufiş and Liviu Dragomirescu. 2004. Tiered Tagging Revisited. In Proc. of the Fourth Intl. Conf. on Language Resources and Evaluation, LREC, pages 39–42. find
  15. Dan Tufiş. 1999. Tiered tagging and combined language models classifiers. In TSD ’99: Proceedings of the Second International Workshop on Text, Speech and Dialogue, pages 28–33. find
  16. Deniz Yuret and Ferhan Türe. 2006. Learning morphological disambiguation rules for Turkish. In Proceedings of the main conference on Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, pages 328–334. find