Hacker News new | past | comments | ask | show | jobs | submit login
Neural Decipherment via Minimum-Cost Flow: From Ugaritic to Linear B (arxiv.org)
30 points by godelmachine 37 days ago | hide | past | favorite | 13 comments

But no Linear-A? That would be the real test


I think it must have to be genetically related to known languages to predict cognates, etc. AFAIK, no one has too much of an idea what language family the Minoan language could be a part of.

They could run their method with different language families and see which one fits best.

AFAIK that is how the existing attempts to match Linear A to a language worked too. But perhaps the machine could do it better.

Wow, this is very cool. Any fellow readers have suggestions for looking at SOTA with respect to computational approaches to lost languages? Ever since I found etymonline.com and I've started tracing word etymologies back to PIE, I've been thinking that I can't get enough of this stuff.

As to whether this is SOTA or not, I will not say. It is certainly a controversial method. https://en.wikipedia.org/wiki/Mass_comparison

I will say that as a teenager, mass lexical comparison, in Merrit Ruhlen's book 'The Origin of Language: Tracing the Evolution of the Mother Tongue', was very compelling. Ruhlen worked closely with Greenberg, who was a famous historical/comparative linguist of African languages.

Ten or more years later, with the benefit of experience, I feel like aligning hundreds of languages in a common vector space (multilingual word embeddings) will probably answer these questions better. But I am not a comparative linguist.

Here is a candy store for etymology enthusiasts: https://starling.rinet.ru

Wow! This is super cool -- another sibling commenter mentioned a Starostin paper as well. Thanks a ton.

Warning: etymonline.org goes to a sketchy black hole. The correct URL is etymonline.com

Thank you! Edited.

It really is so cool to see how cognates in different languages evolved :)

This needs to applied to the Indus script, stat!

I was thinking Voynich.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact