Automated Phonological Transcription of Akkadian Cuneiform Text

What it is about

Oracc holds many Akkadian texts only as sign-by-sign transliteration. The authors ask whether the phonological transcription (vowel lengths made explicit, logograms replaced by their Akkadian readings) can be produced automatically. The task splits in two: syllabic words are a grapheme-to-phoneme problem, and logograms are mostly a context-dependent lookup. They compare a majority-vote baseline, a statistical substitution model and three character-level encoder-decoders, one of which sees the neighbouring tokens. On a held-out split of about 34k tokens, syllabic tokens reach about 90% top-1 recall and 96% recall at 3, while logograms reach about 69% top-1 and 82% at 3. Fed to the rule-based analyser BabyFST, the output still yields the correct lemma and POS for about 94% of tokens when three candidates are kept.

Ideas

• I1. Dictionary first, model second: look each transliterated token up in a table of forms already transcribed in the corpus, and call the neural or statistical model only for forms that stay ambiguous. The authors propose this themselves (section 6). In Mova it is a table lookup before any learned step, and it can be measured as coverage and accuracy on a held-out split.
• I2. Logograms as lookup with context cues: keep a lemma table for logograms and use the neighbouring words only to choose among its readings. The paper finds that context helps the top candidate by about 8 points and little beyond it. A rule that tries the top candidate with context and fills the rest from the frequency table would match that.
• I3. Audit the training pairs before learning from them: the paper notes that editors sometimes drop mimation, add morphemes or change case, and that a model then generalises these errors to unrelated words. A check that flags a pair whose consonant skeleton does not align with its transliteration is a cheap way to find such pairs for review, in line with the rule that examples are not truth.
• I4. Measure the end use, not only the token accuracy: run the analyser on the transcribed output and report lemma and POS recovery, as the paper does (recall about 90% at one candidate, 94% at three). Use the same two numbers for any Mova transcriber.

For Mova

I1 and I3 go into treebanks/akk: the lookup table and the pair audit come before any model. I2 is a rule for the logogram lexicon in the same place. I4 is a measure for whatever transcriber is built there. None of the neural models is for Mova; the paper's numbers are the baseline to beat with tables.

Topics: Akkadian and cuneiform NLP

Cite

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

@inproceedings{sahala2020automated,
  title = {{Automated Phonological Transcription of Akkadian Cuneiform Text}},
  author = {Sahala, Aleksi and Silfverberg, Miikka and Arppe, Antti and Lindén, Krister},
  booktitle = {Proceedings of the Twelfth Language Resources and Evaluation Conference},
  year = {2020},
  url = {https://aclanthology.org/2020.lrec-1.433/}
}

References

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

1. Bahdanau, D., Cho, K., and Bengio, Y. (2014). Neural machine translation by jointly learning to align and translate. cite arxiv:1409.0473. ↗
2. Bisani, M. and Ney, H. (2008). Joint-sequence models for grapheme-to-phoneme conversion. Speech communication, 50(5):434–451. find
3. Gelb, I. J. P. (1970). Comments on the akkadian syllabary. Orientalia, 39:516–546. find
4. Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8):1735–1780. find
5. Hulden, M. (2009). Foma: a finite-state compiler and library. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics: Demonstrations Session, pages 29–32. Association for Computational Linguistics. find
6. Kingma, D. P. and Ba, J. (2014). Adam: A method for stochastic optimization. find
7. Korchagina, N. (2017). Normalizing medieval German texts: from rules to deep learning. In Proceedings of the NoDaLiDa 2017 Workshop on Processing Historical Language, pages 12–17, Gothenburg, May. Linköping University Electronic Press. find
8. Kouwenberg, B. (2011). Akkadian in general. In Stefan Weninger, editor, Semitic Languages, chapter 12, pages 330–339. De Gryuter Mouton, Germany. find
9. Lindén, K., Axelson, E., Hardwick, S., Pirinen, T. A., and Silfverberg, M. (2011). HFST—Framework for Compiling and Applying Morphologies. In Cerstin Mahlow et al., editors, Systems and Frameworks for Computational Morphology, pages 67–85, Berlin, Heidelberg. Springer Berlin Heidelberg. find
10. Novak, J. R., Minematsu, N., and Hirose, K. (2012). Wfst-based grapheme-to-phoneme conversion: Open source tools for alignment, model-building and decoding. In Proceedings of the 10th International Workshop on Finite State Methods and Natural Language Processing, pages 45–49. find
11. Rao, K., Peng, F., Sak, H., and Beaufays, F. (2015). Grapheme-to-phoneme conversion using long short-term memory recurrent neural networks. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4225–4229. IEEE. find
12. Sahala, A., Silfverberg, M., Arppe, A., and Lindén, K. (2019 (submitted)). BabyFST - towards a finite-state based computational model of ancient babylonian. In LREC. find our card
13. Smith, E. (2007). Phonological reconstruction of a dead language using the gradual learning algorithm. In Proceedings of Ninth Meeting of the ACL Special Interest Group in Computational Morphology and Phonology, PI of both of these projects is Saana Svärd (University of Helsinki). pages 57–64, Prague, Czech Republic, June. Association for Computational Linguistics. find
14. Spar, I. (1988). Cuneiform Texts in the Metropolitan Museum of Art. Volume I: Tablets, Cones, and Bricks of the Third and Second Millennia. The Metropolitan Museum of Art. find
15. Talon, P. (2005). The Standard Babylonian Creation Myth: Enūma Eliš. State Archives of Assyria Cuneiform Texts IV. The Neo-Assyrian Text Corpus Project, Helsinki. find
16. von Soden, W. (1995). Grundriss der Akkadischen Grammatik. Editrice Pontificio Instituto Biblico, Rome, 3 edition. find
17. Language Resource References Oracc. (2014). Oracc: The open richly annotated cuneiform corpus. http://oracc.org. ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lingbulk-2020-sahala-automated-phonological-transcription-akkadian-cuneiform/