3 ms·
> According to the study, the neural machine translation provided a BLEU4 score of 36.52 for cuneiform to English, and a score of 37.47 for transliterated cunei
by hirundo 3y ago
> According to the study, the neural machine translation provided a BLEU4 score of 36.52 for cuneiform to English, and a score of 37.47 for transliterated cuneiform to English. BLEU4 scores are from 0 to 100, with 0 being the lowest and 100 being a perfect translation, which even a human translator could not achieve. Around 37 is considered fairly good for an early-stage translation model, explained Gutherz.
> Gutherz said that Google Translate, a privately-funded commercial tool that has been in existence for over a decade, would get a BLEU4 score of about 60 translating from Spanish to English.
It looks more like an early stage of a significant AI accomplishment to me. The glass is half full.
- tgv 3y agoBecause there's such a large corpus of 5000 year old cuneiforms waiting to be added to the training set? Ok, that sounds a bit snarky, but how would you expect to improve certainty when there is so little data?
- saas_sam 3y agoThere actually are... "There are far more of these tablets in existence than I could have imagined. In fact, between half a million and two million cuneiform tablets are estimated to have been excavated in modern times, of which only approximately 30,000–100,000 have been read or published." https://library.stanford.edu/blogs/special-collections-unbound/2020/01/new-cuneiform-tablet-acquired-special-collections#:~:text=In%20fact%2C%20between%20half%20a,have%20been%20read%20or%20published https://library.stanford.edu/blogs/special-collections-unbou....
- CSSer 3y agoI feel like this must be a really rewarding time to be alive for diligent archivists and demotivated researchers alike. I hope the revelation that proper cataloging could mean that it's (once again?) possible for single researchers or small teams to make discoveries and then immediately see their findings reverberate in a virtual instant throughout their field (as updated data sets are used to re-calibrate models) could serve as an incentive for better data collection and management too.
- oh_sigh 3y agoAnd those are just the tablets we've found. It's quite likely there are millions of other tablets just sitting in caves and and old, buried cellars waiting to be excavated. There are nowhere near enough Assyriologists in the world to even think about handling how many tablets we have access to, let alone the new ones we will find in the future, so machine translation will be a huge boon to the field, even if the translations are a little bit iffy. Right now, Assyriologists will prioritize - if the tablets come from, say, someone's old garage, they are unlikely to be translated. If it comes from a palace or a temple, then the odds are better. But imagine the situation where someone ended up with a copy of the Declaration of Independence in their garage.
- saas_sam 3y agoSeems like it's only a matter of time before this hits an inflection point and an explosion of insights from ancient cultures emerges. Exciting. Many long-held beliefs will be upturned. Many people will be digging their heels in...
- tivert 3y ago> Many long-held beliefs will be upturned. Many people will be digging their heels in... Yeah, according to the AI, the tablets say the people of Akkad had a great interest in 21st century culture, and they worshipped Harry Styles as a literal god.
- saas_sam 3y agoIt's true, Watermelon Sugar is clearly about the economic prosperity that followed Sargon's unification of the Babylonians.