BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Reducing gender bias in neural machine translation as a domain ada
 ptation problem - Danielle Saunders (University of Cambridge)
DTSTART:20200501T110000Z
DTEND:20200501T120000Z
UID:TALK141901@talks.cam.ac.uk
CONTACT:James Thorne
DESCRIPTION:Training data for NLP tasks often exhibits gender bias in that
  fewer sentences refer to women than to men. In Neural Machine Translation
  (NMT) gender bias has been shown to reduce translation quality\, particul
 arly when the target language has grammatical gender. The recent WinoMT ch
 allenge set allows us to measure this effect directly.\n\nIdeally we would
  reduce system bias by simply debiasing all data prior to training\, but t
 his is itself a challenge. Rather than attempt to create a `balanced' data
 set\, we adapt to a small set of trusted\, gender-balanced examples. This 
 approach gives strong and consistent improvements in gender debiasing with
  much less computational cost than training from scratch.\n\nA known pitfa
 ll of adapting to new domains is `catastrophic forgetting'\, which we addr
 ess both in adaptation and in inference. During adaptation we show that El
 astic Weight Consolidation allows a trade-off between general translation 
 quality and bias reduction. During inference we propose a lattice-rescorin
 g scheme which allows extremely strong bias reduction with no degradation 
 of general translation quality. We show this scheme can be applied to redu
 ce gender bias in the output of `black box` online commercial translation 
 systems.
LOCATION:https://meet.google.com/hhk-hmiz-mpt
END:VEVENT
END:VCALENDAR