BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Towards Perfect Supervised and Unsupervised Machine Translation - 
 Prof. Dr. Alexander Fraser\, CIS\, LMU Munich
DTSTART:20210121T110000Z
DTEND:20210121T120000Z
UID:TALK156109@talks.cam.ac.uk
CONTACT:Marinela Parovic
DESCRIPTION:Data-driven Machine Translation is an interesting application 
 of\nmachine-learning-based natural language processing techniques to\nmult
 ilingual data. Particularly with the recent advent of powerful\nneural net
 work models\, it has become possible to incorporate many\ntypes of informa
 tion directly into the model and to robustly model\nlong-distance dependen
 cies in the sequence of words being generated.\n\nI will discuss four area
 s of work addressing important weaknesses of\ndata-driven machine translat
 ion approaches. First\, I will present an\nalternative model to phrase-bas
 ed statistical machine translation\,\nwhich jointly models translation ope
 rations and reordering operations\nand was widely adopted by researchers a
 nd end-users. Second\, I will\ndiscuss the important problem of data spars
 ity in translation which is\ncaused by rich morphology\, and discuss exten
 sive work we have carried\nout to overcome this. Third\, I will discuss pr
 ogress towards breaking\nthe strong domain dependency between the data use
 d to train supervised\nneural machine translation systems and the data tha
 t will be\ntranslated. Finally\, I will briefly present a new research pro
 gram\nwhich will allow us to build strong unsupervised machine translation
 \nsystems\, enabling the carrying out of high quality translation between\
 npairs of languages for which no known source of parallel training data\ne
 xists.
LOCATION:https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBd
 XVpOXFvdz09
END:VEVENT
END:VCALENDAR