International translation in the Grammatical Framework (GF)
What it is about
A short demo announcement for the Grammatical Framework (GF), an open-source formalism for multilingual grammars that share one abstract syntax, which serves as an interlingua. GF began with domain-specific controlled-language systems and has since grown into wide-coverage translation. The project has grammars for more than thirty languages, and fifteen of them are in a wide-coverage translator that runs in the GF cloud and as a mobile app of small size. The description says that this translator uses layered interlinguas, in which deeper analysis works with narrower coverage and is backed by shallower levels with wider coverage. The lexicon is linked across languages and bootstrapped from open resources, and statistical ranking from treebanks picks the best translation among the candidates. The paper contains no measurement; it describes the systems that the authors planned to demonstrate.
Ideas
- I1. Build a fallback chain for the reader. When the frame reader cannot cover a sentence, the translator should drop to the tree-level meta tokens, and then to the word level, as the layered interlingua does. Report the share of tokens read at each level for every author, so that the coverage of each layer is visible.
- I2. Rank the encoder's candidate chunks with counts from a treebank, as the paper does with its statistical ranking. The meta-world already keeps counts for each chunk in each context; the new element would be a tie-breaker from the parse trees of the target side, measured on the held-out verses.
- I3. Keep the lexicon linked by concept across languages, one entry per concept carrying all its forms. The meta-world is already a concept table, so the test is whether the forms of one concept in a new target language can be added without touching the reader.
For Mova
The code is tools/translate (the reader fallback in the frame code, and the encoder's ranking in profile.rs) and tools/roundtrip. The paper gives no numbers and is only a pointer, so any idea above has to be measured in Mova before it is reported. The layered-coverage report (I1) is the cheapest to run and the most informative about the translator's current state.
Cite
@inproceedings{ranta2016international,
title = {{International translation in the Grammatical Framework (GF)}},
author = {Ranta, Aarne and Angelov, Kasimir and Hallgren, Thomas and Kolachina, Prasanth and Listenmaa, Inari},
booktitle = {Proceedings of the 19th Annual Conference of the European Association for Machine Translation: Projects/Products},
year = {2016},
url = {https://aclanthology.org/2016.eamt-2.15/}
}This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ranta-2016-international-translation-grammatical-framework-gf/