BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Faster Minimum Bayes Risk Decoding with Confidence-based Pruning -
  Julius Cheng (University of Cambridge)
DTSTART:20240119T120000Z
DTEND:20240119T130000Z
UID:TALK210787@talks.cam.ac.uk
CONTACT:Richard Diehl Martinez
DESCRIPTION:Minimum Bayes risk (MBR) decoding outputs\nthe hypothesis with
  the highest expected utility over the model distribution for some utility
 \nfunction. It has been shown to improve accuracy over beam search in cond
 itional language\ngeneration problems and especially neural machine transl
 ation\, in both human and automatic\nevaluations. However\, the standard s
 amplingbased algorithm for MBR is substantially more\ncomputationally expe
 nsive than beam search\,\nrequiring a large number of samples as well as\n
 a quadratic number of calls to the utility function\, limiting its applica
 bility. We describe an\nalgorithm for MBR which gradually grows the\nnumbe
 r of samples used to estimate the utility\nwhile pruning hypotheses that a
 re unlikely to\nhave the highest utility according to confidence\nestimate
 s obtained with bootstrap sampling.\nOur method requires fewer samples and
  drastically reduces the number of calls to the utility\nfunction compared
  to standard MBR while being statistically indistinguishable in terms of\n
 accuracy. We demonstrate the effectiveness\nof our approach in experiments
  on three language pairs\, using chrF++ and COMET as utility/evaluation me
 trics.
LOCATION:Computer Lab\, SS03
END:VEVENT
END:VCALENDAR