ראַף  Feb 28, 2024

Announcement!

I've pushed support for #Yiddish and #Ladino to my translation server. That means #Babka will now have support for translating from and to those two languages

Just to temper expectations, these models have a BLEU score of 18-25 while typical models score at 40-50 BLEU

As always reach out to me if you see translations that seem off or if you'd like to see other languages supported!

Show threadראַף  Feb 28, 2024
I haven't begun doing backtranslation and other ways to take advantage of monolingual corpora. But once I do, expect the quality to improve not just for Yiddish and Ladino, but also Hebrew and a few of the other languages.
Show threadAlonFeb 28, 2024
@raf Are there good corpora for Hebrew-English translation in general? Hebrew machine-translates poorly, to the point that Google Translate sometimes reverses the meanings of sentences (and Israelis on Twitter have a few stock phrases they ask people to translate if they worry about as-a-Jew pretenders).
Show threadראַף  

@Alon

The corpora are decent-ish. Mostly it's Common Crawl aligned text:

https://opus.nlpl.eu/results/he&en/corpus-result-table

But I also have access to strictly Hebrew Common Crawl and some tricks I'd like to try to improve things. I think I have a decent shot of making a better translator here

Feb 28, 2024 at 3:44pmWeb