From Pattern to Interpretation. Using Colibri Core to Detect Translation Patterns in the Peshitta.

What it is about

This is a progress report of a project that uses Colibri Core, a language-independent n-gram, skipgram and flexgram counter, to compare the Hebrew Bible with its Syriac translation, the Peshitta. Because Colibri Core works on one language at a time, the author aligns words and phrases first with GIZA++, using lemmatised, unpointed texts from the ETCBC database of the book of Genesis, and then builds separate Hebrew and Syriac pattern models with a minimum of 10 attestations. The main finding is that the two texts share a Zipfian pattern shape within about ten percent. Syriac has more distinct patterns at most orders and smaller coverage, which the author reads as a more flexible translation. The article gives case studies from Genesis (for example Gen 2:2, where the Hebrew "seventh day" becomes "sixth day", and Gen 21:6, where a Hebrew wordplay on laughter is replaced by a common word for joy). No log-likelihood scores are reported yet.

Ideas

I1. A divergence flag: list words whose per-corpus coverage differs by a large factor between source and target (the paper's example is about ten times for a wordplay), and send them to a reviewer. In tools/hbo this would be a check on the aligned text, not a new model.
I2. Word alignment with a classical statistical aligner (Viterbi-based, followed by a phrase pass) as a cheap baseline for building parallel Hebrew-Aramaic pairs, before any neural aligner. It is deterministic given the same input and fits the SLM's preference for counted rules.
I3. A gold list of the paper's case-study verses (Gen 2:2, 2:6, 8:21, 9:14, 21:6) with the expected divergence type (lexical choice, numeral, idiom, wordplay), for a test of the flag in I1.

For Mova

I1 and I3 are a test for tools/hbo and for the Aramaic targums (Biblical Aramaic texts, annotated, and the Peshitta as a witness). I2 is an alignment step for any future parallel corpus; it is not in the current parser. The n-gram counts themselves are a measure, not a rule.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.63317/3x8swv66s3jx

@inproceedings{coeckelbergs2022pattern,
  title = {{From Pattern to Interpretation. Using Colibri Core to Detect Translation Patterns in the Peshitta.}},
  author = {Coeckelbergs, Mathias},
  booktitle = {Proceedings of the Thirteenth Language Resources and Evaluation Conference},
  year = {2022},
  doi = {10.63317/3x8swv66s3jx},
  url = {https://doi.org/10.63317/3x8swv66s3jx}
}

References

6 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Hurskainen, A. (2020). Can machine translation assist in Bible translation? Technical Reports on Language Technology Report 62. find
Morrison, C.E., Kiraz, G., and Bali, J. (2019). The Syriac Peshitta Bible with English Translation. Genesis. Piscataway, NJ: Gorgias Press. find
Koehn, P. (2009). Statistical Machine Translation. Cambridge University Press. ↗
Och, F.J. and Ney, H. (2003). A Systematic Comparison of Various Statistical Alignment Models. Computational Linguistics 29(1), pp. 19-51. ↗
van Gompel, M. and van den Bosch, A. (2016). Efficient N-gram, Skipgram and Flexgram Modelling with Colibri Core. Journal of Open Research Software, 4(1), pp. 30-40. ↗
van Peursen, W.T. (2007). Language and Interpretation in the Syriac Text of Ben Sira. A Compara-tive Linguistic and Literary Study (Monographs of the Peshitta Institute Leiden 16;). Leiden: Brill. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lingbulk-2022-coeckelbergs-pattern-interpretation-colibri-core-detect/