Source
Goldwasser, Gruber, Kalai and Paradise (NeurIPS 2023), LIT-941, framed from its title on by the translation of animal communication, §§1–4 read in NOTE-804. The general Occam-type guarantee is Theorem A.2; the rates are Theorems 3.2, 3.4 and 3.5 in two stylised models.
What is asserted, and by whom
The thesis is that unsupervised translation is possible in general, not only in the authors’ models: a translator scored against a plausibility prior over the target language alone is learnable, and the data it needs fall as the languages share more common ground and as the source language is more complex. The paper is framed by it, for animal communication in particular, and ends with advice on what data to collect for whale communication. NOTE-804 rates the general guarantee strong (C1, Theorem A.2), the rates strong in the models and weak beyond them (C3), and prompting a language model as the prior weak (C5).
What the record holds toward it
- THEORY-270 (Active): in the common-nonsense model, with uniform languages and one-to-one translators, maximum likelihood with the prior alone has error O(log|Θ| / (α·min(m, |T|))), tight up to a constant.
- Theorem A.2 gives an Occam-type guarantee for any finite translator family and prior, but it bounds error by the prior’s plausible ambiguities, which the paper sizes only in its models.
What would confirm or reject it
A bound for non-uniform or lossy settings that still falls in common ground and source size, or a real-corpus experiment at the predicted rate, confirms it; a non-uniform model in which richer source languages do not reduce the common ground needed rejects it for natural settings. QUESTION-195 asks the same thing as a question; this entry is the answer the authors propose.