Elamur Research⌕
☰
home / papers / 2000 / Morphological Tagging: Data vs. Dictionaries

Morphological Tagging: Data vs. Dictionaries

Summary

The paper addresses full morphological tagging for languages with rich inflection and agglutination. In these languages it is not enough to determine the part of speech: case, number, gender and other grammatical categories must also be identified. As a result, the tagsets are very large while annotated data are scarce, so statistical models suffer from data sparseness. The author compares five languages (Czech, Estonian, Hungarian, Romanian and Slovene) on Orwell's novel "1984" as annotated within Multext-East. English, based on the Penn Treebank, is added for comparison. The starting point is an exponential (log-linear) tagger previously developed for Czech. It was reworked to be language-independent, and a module for unknown words was added. Two settings were then compared: in the first, the possible tags are taken only from the training corpus; in the second, from a separate morphological dictionary. The main conclusion is that even a simple tagger with a dictionary performs on par with or better than a model fully trained on ~100k tokens without a dictionary. Therefore, for such languages it pays off more to invest in a dictionary than in annotating additional texts.

Key points

  • Data: approximately 100k tokens per language. The first 20k tokens are reserved for testing and the rest for training. The English portion is a Penn Treebank fragment of comparable size.
  • Tagset size: from 139 tags in English to 970 in Czech and 1033 in Slovene. The share of ambiguous tokens ranges from ~21.6% in Hungarian to ~46% in Czech.
  • Model: separate predictions for each grammatical category, based on "ambiguity classes". Innovations: general features that make it possible to learn from unambiguous tokens as well; automatic selection of variable-length context; varying feature weights; a tag guesser for unknown words based on prefixes and suffixes.
  • Slovene: the exponential tagger has an error rate of 13.82% when all tokens are counted and 16.26% when only words are counted. This is comparable to other published results on the same data.
  • Dictionary vs. data: the greatest advantage of the dictionary is seen in part-of-speech-only tagging, where training data are hardly needed at all. The author believes that even a semi-automatically built dictionary would help considerably.

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: Linguistics & languages

Topics Morphology

Cite

Cite as Jan Hajic. 2000. Morphological Tagging: Data vs. Dictionaries. 1st Meeting of the North American Chapter of the Association for Computational Linguistics.

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

@inproceedings{hajic2000morphological,
  title = {{Morphological Tagging: Data vs. Dictionaries}},
  author = {Jan Hajic},
  booktitle = {1st Meeting of the North American Chapter of the Association for Computational Linguistics},
  year = {2000},
  url = {https://aclanthology.org/A00-2013/}
}

References

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

  1. Eric Brill and Jun Wu. 1998. Classifier combination for improved lexical disambiguation. In Proceedings of ACL/COLING'g8, pages 191-195, Montreal, Canada. ACL/ICCL. find
  2. Eric Brill. 1995. Transformation-based error-driven learning and natural language processing: A case study in part-of-speech tagging. Computational Linguistics, 21:543-565. our card
  3. Jean-Pierre Chanod. 1997. Current developments for Central & Eastern European languages. In Proceedings of EU Project meeting TELRI I, Romania. find
  4. Kenneth W. Church. 1988. A stochastic parts program and noun phrase parser for unrestricted text. In Proceedings of the Second Conference on Applied Natural Language Processing, pages 136-143, Austin, Texas. ACL. our card
  5. Walter Daelemans, Jakub Zavrel, Peter Berck, and Steven Gillis. 1996. MBT: A memory-based part of speech tagger generator. In Proceedings of WVLC 4, pages 14-27. ACL. our card
  6. Tomaz Erjavec, Saso Dzeroski, and Jakub Zavrd. 1999. Morphosyntactic Tagging of Slovene: Evaluating PoS Taggers and Tagsets. Technical Report IJS-DP 8018, Dept. for Intelligent Systems, Jozef Stefan Institute, Ljubljana, Slovenia, April 2nd. find
  7. Jan Haji~ and Barbora Hladk~t. 1997. Tagging of inflective languages: a comparison. In Proceedings of ANLP'gT, pages 136-143, Washington, DC. ACL. find
  8. Jan Haji~ and Barbora Hladk& 1998. Tagging inflective languages: Prediction of morphological categories for a rich, structured tagset. In Proceedings of A CL/COLING'98, pages 483-490, Montreal, Canada. ACL/ICCL. find
  9. Andrew W. Mackie Julian Benello and James A. Anderson. 1989. Syntactic category disambiguation with neural networks. Computer Speech and Language, 3:203-217. find
  10. Bernard Merialdo. 1992. Tagging text with a probabilistic model. Computational Linguistics, 20(2):155-171. find
  11. Adwait Ratnaparkhi. 1996. A maximum entropy model for part-of-speech tagging. In Proceedings of EMNLP 1, pages 133-142. ACL. our card
  12. Helmut Schmid. 1994. Probabilistic part-of-speech tagging using decision trees. In Proceedings of International Con]erence on New Methods in Language Processing, pages 44-49, Manchester, England. find
  13. Dan Tufts. 1999. Tiered tagging and combined language models classifiers. In Proceedings of Text, Speech and Dialogue'99, Mari~nskd LLzn~, Czech Republic, Sept. 15-18. find
  14. Jean Vdronis. 1996a. Multext-East (Copernicus 106). http://www.lpl.univ-aix.fr/projects/multext-east. ↗
  15. Jean Vdronis. 1996b. Multext-East language-specific resources (Copernicus 106). http://www.lpl.univ-aix.fr/projects/multext-east/MTE2.html. ↗

This page in Markdown → https://md.research.elamur.ai/free/en/papers/hajic-2000-morphological-tagging-data-vs-dictionaries/