BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:NLIP Seminar Series
SUMMARY:Multilingual Autoregressive Entity Linking - Nicol
 a De Cao (University of Amsterdam\, Huggingface)
DTSTART;TZID=Europe/London:20220318T120000
DTEND;TZID=Europe/London:20220318T130000
UID:TALK171803AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/171803
DESCRIPTION:Entities are at the center of how we represent and
  aggregate knowledge. For instance\, Encyclopedias
  such as Wikipedia are structured by entities (e.g
 .\, one per Wikipedia article). The ability to ret
 rieve such entities given a query is fundamental f
 or knowledge-intensive tasks such as entity linkin
 g and open-domain question answering. Current appr
 oaches can be understood as classifiers among atom
 ic labels\, one for each entity. Their weight vect
 ors are dense entity representations produced by e
 ncoding entity meta information such as their desc
 riptions. This approach has several shortcomings: 
 (i) context and entity affinity is mainly captured
  through a vector dot product\, potentially missin
 g fine-grained interactions\; (ii) a large memory 
 footprint is needed to store dense representations
  when considering large entity sets\; (iii) an app
 ropriately hard set of negative data has to be sub
 sampled at training time. In this work\, we propos
 e mGENRE\, the first system that retrieves entitie
 s by generating their unique names\, left to right
 \, token-by-token in an autoregressive fashion. Th
 is mitigates the aforementioned technical issues s
 ince: (i) the autoregressive formulation directly 
 captures relations between context and entity name
 \, effectively cross encoding both\; (ii) the memo
 ry footprint is greatly reduced because the parame
 ters of our encoder-decoder architecture scale wit
 h vocabulary size\, not entity count\; (iii) the s
 oftmax loss is computed without subsampling negati
 ve data. We experiment with more than with more th
 an 100 languages on more than 25 datasets on entit
 y disambiguation\, end-to-end entity linking and d
 ocument retrieval tasks\, achieving new state-of-t
 he-art or very competitive results while using a t
 iny fraction of the memory footprint of competing 
 systems.\n\nTopic: NLIP Seminar\nTime: Mar 18\, 20
 22 12:00 PM London\n\nJoin Zoom Meeting\nhttps://c
 l-cam-ac-uk.zoom.us/j/93197062657?pwd=eEVuT0h4MGRJ
 OEhCaEF4MDJZQm9zdz09\n\nMeeting ID: 931 9706 2657\
 nPasscode: 501991\n
LOCATION:Virtual (Zoom)
CONTACT:Michael Schlichtkrull
END:VEVENT
END:VCALENDAR