Elamur Research⌕
☰
home / papers / 2005 / Part of Speech Tagging for Amharic using Conditional Random Fields

Part of Speech Tagging for Amharic using Conditional Random Fields

Paper · 2005 · Fissaha Adafre, Sisay

Summary

Amharic is a Semitic language with rich morphology. Many function words (prepositions, conjunctions, relative markers, negation) attach to content words as bound morphemes. At the time of this work it had no standard POS tagset and no sizeable annotated corpus. The author revised an earlier 25-tag inventory by Mesfin (2001) into ten coarse classes and checked that it covered real text by hand-tagging five newspaper articles, about 1000 tokens. He then framed two separate problems as sequence labeling with Conditional Random Fields (CRFs): splitting words into selected bound morphemes, and tagging fully inflected words. The features drew on characters, a dictionary, morphology, lexical items and tag bigrams. About 80% of test words were unseen, yet segmentation reached roughly 84% and POS tagging about 75%, against a 51% baseline. The paper argues that CRFs, because they combine many overlapping cues, make good use of very scarce data. It also gives a clear picture of how hard Amharic morphology is to handle computationally.

Key points

  • The tagset has 10 tags, covering nouns (N), verbs (V), auxiliaries (AUX), numerals (NU), adjectives (AJ), adverbs (AV), adpositions (AP) and interjections (I), plus R and punctuation (PU). They were derived by merging Mesfin's 25 tags, because 25 classes would make a tiny corpus too sparse.
  • In the annotated sample, nouns dominate (586 tokens), followed by verbs (203). No interjections appear. Verbal nouns heading noun clauses add to the noun share.
  • Segmentation results:
    • Character features alone gave 0.819.
    • Adding dictionary and morphological features raised this to 0.841.
    • Removing the previous-label feature dropped accuracy to the 0.66–0.70 range.
  • POS tagging results:
    • The baseline was 0.513.
    • Lexical, morphological, dictionary and bigram features together gave 0.748.
    • Morphology contributed the largest single gain (0.613 → 0.700).
  • The main error was tagging other categories as nouns. Noun affixes migrate onto prenominal adjectives, numerals and nouns, and prepositions share morphemes with subordinators, which blurs the formal distinctions.

Interesting ideas

  • I1. When annotated data is tiny, merge a fine tagset into a hierarchy of coarse major classes so that each label gets enough examples, and keep subclasses for later refinement.
  • I2. Classify a construction by its visible morphological form rather than its syntactic function (ye+NOUN stays a noun subtype), and leave the functional reading to a later layer; this improves annotation consistency.
  • I3. Before scaling annotation, pilot the tagset on a small real sample to check for gaps and to see how the classes are distributed.
  • I4. Tag every token of a collocation separately rather than giving the whole multiword unit one tag, which reduces context-dependent annotator choices.
  • I5. Run ablations both with and without the label-transition feature, and report the standard deviation across folds, to show how much sequence context actually matters.

For Mova

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

  • I1 (tokens): Let the en tagger fall back to coarse UPOS-level decisions when fine-tag confidence is low, and expose that fallback in its explanation.
  • I2 (tree): During UD dialect conversion, keep form-based POS stable and move functional distinctions (possessive versus attributive) into deprels or frames.
  • I3 (learn): Run any new label scheme for silver data on a held-out mini-sample first, checking coverage and skew before training on it.
  • I4 (tokens): Keep MWE handling out of the tagger and represent it later (e.g., via fixed/compound relations), so token tags stay consistent.
  • I5 (learn): Add "no-transition" and "no-dictionary" variants as negative controls in tagger and parser evaluation, so the share of each evidence source can be measured.

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

🎓University of AmsterdamAmsterdam · Netherlands144 works · 173,906 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.Semitic languages · 2,908 worksMorphology · 405 worksPart-of-speech tagging · 126 works

Cite

Cite as Fissaha Adafre, Sisay. 2005. Part of Speech Tagging for Amharic using Conditional Random Fields. Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages.

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

@inproceedings{sisay2005part,
  title = {{Part of Speech Tagging for Amharic using Conditional Random Fields}},
  author = {Fissaha Adafre, Sisay},
  booktitle = {Proceedings of the ACL Workshop on Computational Approaches to Semitic Languages},
  year = {2005},
  doi = {10.3115/1621787.1621797},
  url = {https://doi.org/10.3115/1621787.1621797}
}

References

18 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Nega Alemayehu. 1999. Development of stemming algorithm for Amharic text retrieval. PhD Thesis, University of Sheffield. find
  2. Atelach Alemu. 2002. Automatic Sentence Parsing for Amharic Text: An Experiment using Probabilistic Context Free Grammars. Master Thesis, Addis Ababa University. find
  3. Amsalu Aklilu. 1987. Amharic-English Dictionary. Ku-raz Publishing Agency. find
  4. Roy Bar-Haim, Khalil Simaan and Yoad Winter. 2004. Part-of-Speech Tagging for Hebrew and Other Semitic Languages. Technical Report. find
  5. Abiyot Bayou. 2000. Developing automatic word parser for Amharic verbs and their derivation. Master Thesis, Addis Ababa University. find
  6. W. Cohen. 2004. Methods for Identifying Names and Ontological Relations in Text using Heuristics for Inducing Regularities from Data. http://minorthird.sourceforge.net ↗
  7. Peter T. Daniels, 1997. Script of Semitic Languages in: Robert Hetzron, editor,Proceedings of the Corpus Linguistics. 16–45. find
  8. Mona Diab, Kadri Hacioglu and Daniel Jurafsky. 2004. Automatic tagging of Arabic text: From row text to base phrase chunks. In Daniel Marku, Susan Dumais and Salim Roukos, editors, HLT-NAACL 2004: Short papers, pages 149–152, Boston, Massachusetts, USA, May 2–May 7. Association for Computational Linguistics Sisay Fissaha Adafre and Johann Haller. 2003a. Amharic verb lexicon in the context of machine translation. Traitement Automatique des Langues Naturelles 2:183–192 find
  9. Sisay Fissaha Adafre and Johann Haller. 2003b. Application of corpus-based techniques to Amharic texts. Machine Translation for Semitic languages MT Summit IX Workshop, New Orleans Mesfin Getachew. 2001. Automatic part of speech tagging for Amharic language: An experiment using stochastic HMM. Master Thesis, Addis Ababa University.
  10. Wolf Leslau. 1995. Reference Grammar of Amharic. Otto Harrassowitz, Wiesbaden. find
  11. A. McCallum and W. Li. 2003. Early results for Named Entity Recognition with conditional random fields, feature induction and web-enhanced lexicons. Proceedings of the 7th CoNLL.
  12. Tetsuji Nakagawa, Taku Kudo and Yuji Matsumoto. 2001. Unknown Word Guessing and Part-of-Speech Tagging Using Support Vector Machines. NLPRS pages 325-331, Boston, Massachusetts, USA, May 2–May 7. http://www.afnlp.org/nlprs2001/pdf/0053-01.pdf ↗
  13. J. Lafferty, F. Pereira and A. McCallum. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. Proceedings of the International Conference on Machine Learning. find
  14. F. Sha and F. Pereira. 2004. Shallow parsing with conditional random fields. Proceedings of Human Language Technology-NAACL. find
  15. Khoja S., Garside R., and Knowles G. 2001. A Tagset for the Morphosyntactic Tagging of Arabic Proceedings of the Corpus Linguistics. Lancaster University (UK), Volume 13 - Special issue, 341. find
  16. Leech G. Wilson. 1996. Recommendations for the Morphosyntactic Annotation of Corpora. EAGLES Report EAG-TCWG-MAC/R, http://www.ilc.pi.cnr.it/EAGLES96/annotate/annotate.html ↗
  17. Daniel Yacob. 1996. System for Ethiopic Representation in ASCII. http://www.abyssiniagateway.net/fidel Baye Yimam. 1999. Root. Ethiopian Journal of Language Studies, 9:56–88. ↗
  18. Baye Yimam. 1986. Yamara Swasw E.M.P.D.A, Addis Ababa. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/adafre-2005-part-speech-tagging-amharic-conditional/