Statistical Glossing - Language Independent Analysis in Bible Translation
What it is about
The paper describes the Statistical Glossing Tool (SGT), a part of the Paratext translation editor, which helps translation consultants check a new Bible translation in a language they do not know. The tool needs no dictionary and no grammar of the target language. For a key word in the model text (say "temple"), it finds the verses where the word occurs, takes the verses of the new translation at the same positions, and counts which target words appear in those verses most often. Each candidate is scored by how often it co-occurs with the source word against how often it appears in the target verses alone. Since small counts are unreliable, the score is a lower bound: the value the match rate exceeds with 95 per cent confidence, computed from the number of matches and the number of verses, through an incomplete beta function. The same machinery builds concordances for the back of Bibles. The weakness the author records is morphology: highly inflected languages need a pre-process that splits affixes. The paper reports use in the field, not a numerical evaluation.
Ideas
- I1. A lower-bound gate for lexicon merges (open problem 2): accept a merge, or a chunk choice in the encoder, only when the 95 per cent lower bound of the co-occurrence rate clears a threshold. This is a closed formula over counts that could decide cases like be with have, where today's merges are noise. Test it on the residual in bits.
- I2. Use verse-level co-occurrence as a baseline for the Hebrew reader: for each Hebrew lemma, the count of verses in which a gloss occurs, compared with the IBM Model 1 alignment in
ibm.rson the same Genesis verses. - I3. A per-author concordance report from
measure.rs: for each concept, the glosses per author with counts and lower bounds. This is the evidence the author delta needs (fix at least three more tokens, hold in half the context), in a form a human can read. - I4. For a target language with rich morphology, split affixes before counting, as SGT does for Swahili-type languages, and compare with the rules in
en/src/morph.rson the unseen concepts thatprofile.rsmust give a bare form.
For Mova
The lower-bound idea is the most useful: it turns a noisy count into a decision that can be read ("95 per cent sure the rate is above r"), which fits Mova's rule that each choice must say why. The paper's verse-level counting is the cheapest baseline for the Hebrew reader and should be run before any new alignment model. The code is tools/translate/ibm.rs, meta.rs and measure.rs. The paper's method is language-independent and rests on no data that Mova would need to ship.
Cite
@inproceedings{riding2008statistical,
title = {{Statistical Glossing - Language Independent Analysis in Bible Translation}},
author = {Riding, J. D.},
booktitle = {Proceedings of Translating and the Computer 30},
year = {2008},
url = {https://aclanthology.org/2008.tc-1.8/}
}References
- Unicode Consortium. The Unicode Standard, Version 5.0 . Addison-Wesley Professional, 2006. find
- Raymond G Gordon. Ethnologue . SIL International, 15th edition, 2005. find
- JD Riding. A relational method for the automatic analysis of highly-in ectional agglutinative morphologies. Master's thesis, Oxford Brookes University (MPhil), 2007. find
- JD Riding. Paratext glossing tool. http://lc.bfbs.org.uk/news.php?item.28.3, February 2008. ↗
- DWC Robinson. The problem of anaphora. Technical report, British & Foreign Bible Society, 1985. find
- DWC Robinson and PJ Robinson. Remembrance of things parsed. Computer Bulletin , 3:2224, 1991. find
- United Bible Societies. Paratext software. translations.org/, March 2006. J D Riding Linguistic Computing at British & Foreign Bible Society October 2008 The http://paratext.ubs- ↗
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/riding-2008-statistical-glossing-language-independent-analysis/