Elamur Research⌕
☰
home / papers / 2018 / Contemporary Amharic Corpus: Automatically Morpho-Syntactically Tagged Amharic Corpus

Contemporary Amharic Corpus: Automatically Morpho-Syntactically Tagged Amharic Corpus

Paper · 2018 · Andargachew Mekonnen Gezmu, Binyam Ephrem Seyoum, Michael Gasser et al.

Summary

Amharic is morphologically rich, and one whitespace-delimited word can carry a whole phrase, clause or sentence because prepositions, conjunctions and auxiliaries attach to their hosts. Earlier Amharic corpora (WIC, amWaC17, An Crúbadán) treat the orthographic word as the unit and use bundled tags, which hides internal syntax. The authors built the Contemporary Amharic Corpus (CACO) from about 25 thousand edited documents: newspapers, news agencies, magazines, fiction, history and politics books, children's books and a modern Bible translation. Cleaning involved three steps. An existing spelling corrector fixed missing spaces and look-alike characters. Homophonic letters were collapsed following the ELA spelling reform. Varied punctuation styles were mapped to single canonical symbols. The authors then extended the rule-based analyzer HornMorpho so that it covers more parts of speech, compounds, light verbs and names, and used it to segment clitics and tag the text automatically. The result is a roughly 24-million-token corpus with morpheme-level annotation, which gives a basis for syntactic work on syntactic words rather than orthographic ones. No tagging accuracy is reported, and many tokens remain unclassified.

Key points

  • CACO contains 25,199 documents, 1,605,452 sentences and 24,049,484 tokens; a 3-gram language model was built from it (919,407 unigrams, 9,170,309 bigrams, 16,033,318 trigrams).
  • Comparison corpora:
    • WIC: 210,000 tokens, 31 compound tags, an automatic tagger reaching 87.4%.
    • amWaC17: 30.5 million tokens from 75,509 documents, tagged with TreeTagger trained on WIC.
    • An Crúbadán: only a word-frequency list of about 17 million words.
  • During normalization, ሐ and ኀ were both turned into ሀ, ሠ became ሰ, ዐ became አ and ፀ became ጸ, and numerous variants of quotes, stops, commas and hyphens were merged. Typographical errors were deliberately left uncorrected.
  • HornMorpho is a cascade of finite-state transducers weighted with feature structures (after Amtrup). The authors extended it to verbs, nouns, adjectives, adverbs, conjunctions and adpositions.
  • Each token in the output records the POS, a morpheme string (stem in braces, labeled affixes, and the root+template for verbs) and a SERA Latin transliteration.

Interesting ideas

  • I1. The tagging unit is the syntactic word: clitics are split off before tagging instead of assigning composite tags, so the structure inside a phrase stays visible to a parser.
  • I2. Feature structures are unified while the analyzer moves through the word, so the agreement options opened by a prefix get narrowed or rejected when the suffix appears. This prunes ambiguity on the fly and avoids filtering afterwards.
  • I3. Character and punctuation variants are collapsed to one canonical form according to a declared external standard before any analysis, so later stages never see spelling noise.
  • I4. Error types are handled differently: cheap and unambiguous fixes are applied automatically, while cases needing a human choice among suggestions are left untouched rather than guessed.
  • I5. The number of candidate analyses per word works as a diagnostic: the paper notes one frequent multi-clitic word that got fifteen guesses.

For Mova

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

  • I1 (tokens): English contractions and possessive 's should likewise be split into separate syntactic tokens, with each piece keeping a pointer to its original surface span so the split stays explainable.
  • I2 (tokens): the morphology layer could carry feature bundles that unify across affixes and adjacent tokens, rejecting inconsistent analyses early and logging which constraint caused each rejection.
  • I3 (tokens): an explicit, versioned normalization table (curly quotes, dash types, Unicode look-alikes) applied at input would make the source of every normalization traceable.
  • I4 (learn): silver-data repair should auto-apply only fixes with a single plausible outcome and send ambiguous ones to an LLM teacher tier or leave them flagged.
  • I5 (learn): tracking analysis counts per token in evaluation would surface coverage gaps in the lexicon or rules and show where repair rules should be induced.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 1Research subjects and languages the work is about.Semitic languages · 2,908 works

Cite

Cite as Andargachew Mekonnen Gezmu, Binyam Ephrem Seyoum, Michael Gasser and Andreas Nürnberger. 2018. Contemporary Amharic Corpus: Automatically Morpho-Syntactically Tagged Amharic Corpus. Proceedings of the First Workshop on Linguistic Resources for Natural Language Processing.

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

@inproceedings{gezmu2018contemporary,
  title = {{Contemporary Amharic Corpus: Automatically Morpho-Syntactically Tagged Amharic Corpus}},
  author = {Andargachew Mekonnen Gezmu, Binyam Ephrem Seyoum, Michael Gasser and Andreas Nürnberger},
  booktitle = {Proceedings of the First Workshop on Linguistic Resources for Natural Language Processing},
  year = {2018},
  url = {https://aclanthology.org/W18-3809/}
}

References

17 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Amsalu Aklilu. 2004. Sabean and Ge’ez symbols as a guideline for Amharic spelling reform. In Proceedings of the first international symposium on Ethiopian philology, pages 18-26. find
  2. Besufikad Alemu. 2013. A Named Entity Recognition for Amharic. Master’s thesis at Addis Ababa University. Addis Ababa, Ethiopia. find
  3. Jan W. Amtrup. 2003. Morphology in machine translation systems: Efficient integration of finite state transducers and feature structure descriptions. Machine Translation, 18(3): 217–238. ↗
  4. Atelach Alemu Argaw and Lars Asker. 2007. An Amharic stemmer: Reducing words to their citation forms. In Proceedings of the 2007 workshop on computational approaches to Semitic languages: Common issues and resources, pages 104-110. Association for Computational Linguistics.
  5. Kenneth R. Beesley and Lauri Karttunen. 2003. Finite-state morphology: Xerox tools and techniques. CSLI, Stanford. find
  6. Girma A. Demeke and Mesfin Getachew. 2006. Manual annotation of Amharic news items with part-of-speech tags and its challenges. Ethiopian Languages Research Center Working Papers, 2, 1-16. find
  7. ELA. 1970. የአማርኛ ፡ ፊደል ፡ ሕግን ፡ አንዲጠብቅ ፡ ለማድረግ ፡ የተዘጋጀ ፡ ራፖር ፡ ማስታወሻ, (engl. A memorandum for standardization of Amharic spelling) Journal of Ethiopian Studies, 8(1): 119–134. find
  8. Yitna Firdyiwek and Daniel Yaqob. 1997. The system for Ethiopic representation in ASCII. https://www.researchgate.net/publication/2682324_The_System_for_Ethiopic_Representation_in_ASCII. ↗
  9. Björn Gambäck, Fredrik Olsson, Atelach Alemu Argaw, and Lars Asker. 2009. Methods for Amharic part-of-speech tagging. In Proceedings of the First Workshop on Language Technologies for African Languages, pages 104-111. Association for Computational Linguistics.
  10. Michael Gasser. 2011. HornMorpho: a system for morphological processing of Amharic, Oromo, and Tigrinya. In Conference on Human Language Technology for Development, pages 94-99, Alexandria, Egypt. find
  11. Andargachew Mekonnen Gezmu, Andreas Nürnberger and Binyam Ephrem Seyoum. 2018. Portable Spelling Corrector for a Less-Resourced Language: Amharic. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, pages 4127– 4132. Miyazaki, Japan: European Language Resources Association (ELRA).
  12. Abeba Ibrahim and Yaregal Assabie. 2014. Amharic Sentence Parsing Using Base Phrase Chunking. In International Conference on Intelligent Text Processing and Computational Linguistics, pages 297-306. Springer, Berlin, Heidelberg. find
  13. Pavel Rychlý and Vít Suchomel. 2016. Annotated Amharic Corpora. In International Conference on Text, Speech, and Dialogue, pages 295-302. Springer, Cham. find
  14. Kevin P. Scannell. 2007. The Crúbadán Project: Corpus building for under-resourced languages. In Building and Exploring Web Corpora: Proceedings of the 3rd Web as Corpus Workshop, Vol. 4, pages 5-15. find
  15. Binyam Ephrem Seyoum, Yusuke Miyao and Baye Yimam Mekonnen. 2018. Universal Dependencies for Amharic. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, pages 2216– 2222. Miyazaki, Japan: European Language Resources Association (ELRA).
  16. Martha Yifiru Tachbelie and Wolfgang Menzel. 2009. Amharic Part-of-Speech Tagger for Factored Language Modeling. In Proceedings of the International Conference RANLP-2009, pages 428-433.
  17. Seid Muhie Yimam. 2014. Automatic annotation suggestions and custom annotation layers in WebAnno. In Proceedings of 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 91-96. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gezmu-2018-contemporary-amharic-corpus-automatically-morpho/