Transfer and MT Modularity

What it is about

The authors argue that second-generation rule-based MT systems break their own modularity rule: their transfer components hold target-language facts, such as the strict subcategorisation of the French verbs savoir and connaître, and their synthesis components are weak because transfer has already done the work. They propose that transfer should keep only contrastive information, the facts that differ between the two languages, and that everything else should go to synthesis. On this design the three components can run in parallel rather than in sequence. They also propose that translators should state their knowledge in a natural metalanguage, so that the rules are written by the people who know the languages. No quantitative evaluation is reported.

Ideas

• I1. Enforce the contrast rule: an author delta may carry only what differs from the standard. A check in tools/translate rejects an override whose target-side fact is already in the standard, and counts how many overrides are refused. A high count would show that deltas are being used to repeat the standard.
• I2. Keep target facts in the target standard: any subcategorisation-like fact about the English side (which verbs take which complements in the Bible pairs) belongs in profile.rs or en/src/morph.rs, not in the reader. Measure how many reader rules move.
• I3. Metalanguage rule files: let a translator write a short statement such as «this author writes thou as you» in a plain format that is parsed into delta entries, so the delta is readable without code.

For Mova

The paper's contrast rule is a clean control for the author DELTA: an override that restates the standard is noise, and the idea behind open problem 2 (keep a merge only if it lowers the residual) is the same test. The metalanguage idea gives the delta a plain format. The parallel design is already the layered residual. None of the ideas has been tried in Mova.

Topics: Machine translation

Cite

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

@inproceedings{isabelle1986transfer,
  title = {{Transfer and MT Modularity}},
  author = {Isabelle, Pierre and Macklovitch, Elliott},
  booktitle = {Coling 1986 Volume 1: The 11th International Conference on Computational Linguistics},
  year = {1986},
  url = {https://aclanthology.org/C86-1025/}
}

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.

1. [I] Vauquois, B., La t r a d u c t i o n a u t o m a t i c _ _ Grenoble, Documents de linguist:Lque quantitative 29, Paris, Dunod, 1975. find
2. [2] Kay M., Functional Unification Grammar: a Formalism for Machine Translation, P r o c e e d i of COLING-84, 75-78, 1984. find
3. [3] Isabelle P., Bourbeau L., TAUM-AVIATION: Its Technical Features and some Experimental Results, Computational Linguistics , ii:i, 18-27, 1985. find
4. [4] Vauquois B., Boitet C., Automated Translation at Grenoble University, Computational[ Linguistics, ii:I, 28-36, 1985. find
5. [5] Bennett W., Sloeum J., The LRC Machine Translation System, Computational Linguistics , 1]:2-3, 111-121, 1985. find
6. [6] Arnold D., & a]. A MU 1 View of the C , A > , T Framework in EUROTRA, Proceedings of the Conference on Theoretical Issues in Mac|line Translation of Natural Lansuages, Colgate University, 1-14, 1985. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/isabelle-1986-transfer-mt-modularity/