Two cases of ambiguity in machine translation: lexical choice and gendered outputs

Speaker Name: 
Alex Rudnick
Speaker Organization: 
Google Translate
Start Time: 
Friday, December 6, 2019 - 12:00pm
End Time: 
Friday, December 6, 2019 - 1:15pm
Location: 
E2 599
Organizer: 
Marilyn Walker

Abstract:

In this talk I'll discuss two machine translation projects that center around making particular kinds of ambiguities explicit. In the first, we explore improving word choice when translating from a resource-rich language (like Spanish) into an under-resourced language (like the indigenous languages of South America). We do this by leaning more heavily on the resources that we have for the source language, and I'll discuss some of these approaches. Secondly, we'll present some current work by the Google Translate team on providing gender-specific translations, part of our work on addressing the biases that are often propagated by ML systems. Both of these projects are about recognizing apparent ambiguities in the input. In the lexical selection case we try to resolve it, but in the gender case, the point is that we cannot.

Event Type: 
Event