BEGIN:VCALENDAR
PRODID:-//Mozilla.org/NONSGML Mozilla Calendar V1.1//EN
VERSION:2.0
BEGIN:VTIMEZONE
TZID:Europe/Vienna
X-TZINFO:Europe/Vienna[2023c]
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+010521
TZNAME:Europe/Vienna(STD)
DTSTART:18930401T000000
RDATE:18930401T000000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19160430T230000
RDATE:19160430T230000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19161001T010000
RDATE:19161001T010000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19170416T020000
RRULE:FREQ=YEARLY;BYMONTH=4;BYDAY=3MO;UNTIL=19180415T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19170917T030000
RRULE:FREQ=YEARLY;BYMONTH=9;BYDAY=3MO;UNTIL=19180916T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19200405T020000
RDATE:19200405T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19200913T030000
RDATE:19200913T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19400401T020000
RDATE:19400401T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19421102T030000
RDATE:19421102T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19430329T020000
RDATE:19430329T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19431004T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=1MO;UNTIL=19441002T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19440403T020000
RRULE:FREQ=YEARLY;BYMONTH=4;BYDAY=1MO;UNTIL=19450402T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19450412T030000
RDATE:19450412T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19460414T020000
RDATE:19460414T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19461007T030000
RDATE:19461007T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19470406T020000
RDATE:19470406T020000
END:DAYLIGHT
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19480418T020000
RDATE:19480418T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19471005T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=1SU;UNTIL=19481003T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19800406T000000
RDATE:19800406T000000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19800928T000000
RDATE:19800928T000000
END:STANDARD
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19810927T030000
RRULE:FREQ=YEARLY;BYMONTH=9;BYDAY=-1SU;UNTIL=19950924T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:Europe/Vienna(DST)
DTSTART:19810329T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU;UNTIL=19960331T020000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:Europe/Vienna(STD)
DTSTART:19961027T030000
RDATE:19961027T030000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETTO:+020000
TZOFFSETFROM:+010000
TZNAME:(DST)
DTSTART:19970330T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETTO:+010000
TZOFFSETFROM:+020000
TZNAME:(STD)
DTSTART:19971026T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CREATED:20240212T123247Z
LAST-MODIFIED:20240412T110434Z
DTSTAMP:20240412T110434Z
UID:938a077b-1715-492c-93b3-2af50763aca1
SUMMARY:Jordan Kodner\, PhD: "Is it Language or Task Design? Reinterpreting
  language models' recent successes in morphology and syntax learning"
DTSTART;TZID=Europe/Vienna:20240424T183000
DTEND;TZID=Europe/Vienna:20240424T200000
TRANSP:OPAQUE
LOCATION:online via Zoom 
DESCRIPTION;ALTREP="data:text/html,Meeting%20Link%3A%20https%3A%2F%2Fus06we
 b.zoom.us%2Fj%2F84282442460%3Fpwd%3DNHVhQnJXOVdZTWtNcWNRQllaQWFnQT09%3Cbr%3
 EMeeting%20ID%3A%20842%208244%202460%3Cbr%3EPasscode%3A%20678868%3Cbr%3E%3C
 br%3EAbstract%3A%20The%20success%20of%20neural%20language%20models%20(LMs)%
 20on%20a%20wide%20range%20of%20%0Alanguage-related%20tasks%20may%20be%20in%
 20part%20due%20to%20their%20ability%20to%20induce%20%0Ahuman-like%20represe
 ntations%20or%20understanding%20of%20natural%20language%20%0Agrammars.%20Hu
 mans%20are%2C%20after%20all%2C%20gold-standard%20language%20learners.%20For
 %20%0Athe%20past%20several%20years%2C%20researchers%20pursuing%20this%20que
 stion%20have%20%0Adeveloped%20a%20number%20of%20methodologies%20for%20testi
 ng%20the%20grammar%20%0Arepresentations%20learned%20by%20LMs%20that%20have%
 20reached%20generally%20positive%20%0Aconclusions.%20I%20will%20take%20a%20
 critical%20look%20at%20such%20studies%20in%20this%20talk.%20%0AWhile%20mode
 rn%20LMs%20are%20clearly%20extremely%20impressive%2C%20and%20clearly%20do%2
 0often%20%0Acapture%20important%20aspects%20of%20natural%20language%20gramm
 ars%2C%20the%20%0Amethodologies%20of%20many%20popular%20studies%20have%20un
 fairly%20overestimated%20the%20%0Acapacities%20of%20LMs%20when%20it%20comes
 %20to%20their%20ability%20to%20induce%20human-like%20%0Arepresentations.%20
 Focusing%20on%20questions%20of%20hierarchical%20syntactic%20%0Arepresentati
 ons%20and%20generalization%20in%20inflectional%20morphology%2C%20I%20will%2
 0%0Adiscuss%20how%20unintended%20biases%20in%20data-splitting%2C%20artifici
 al%20training%20or%20%0Atest%20data%2C%20overly%20simplistic%20evaluations%
 2C%20weak%20or%20absent%20baselines%2C%20and%20%0Afaulty%20interpretations%
 2C%20have%20conspired%20to%20overestimate%20the%20abilities%20of%20%0ALMs.%
 20While%20the%20conclusions%20of%20this%20study%20are%20largely%20negative%
 20in%20terms%20%0Aof%20the%20current%20state-of-affairs%2C%20they%20are%20a
 lso%20optimistic.%20By%20employing%20%0Amore%20thorough%20and%20rigorous%20
 methodologies%2C%20we%20have%20developed%20a%20better%20%0Ascientific%20und
 erstanding%20of%20the%20nature%20of%20LMs%20and%20representations%20of%20th
 e%0A%20grammar.%20In%20identifying%20weak%20points%20for%20current%20models
 %2C%20we%20points%20%0Atowards%20research%20areas%20where%20greater%20impro
 vements%20may%20be%20gained.%3Cbr%3E%3Cbr%3EJordan%20Kodner%20is%20an%20Ass
 istant%20Professor%20in%20the%20Stony%20Brook%20University%20%0ADepartment%
 20of%20Linguistics%20and%20an%20affiliate%20of%20the%20Institute%20for%20Ad
 vanced%0A%20Computational%20Science%20and%20Natural%20Language%20Processing
 %20group.%20His%20%0Aprimary%20research%20revolves%20around%20computational
 %20approaches%20to%20child%20%0Alanguage%20acquisition%20and%20their%20broa
 der%20implications.%20In%20particular%2C%20%0Aalgorithmic%20models%20of%20g
 rammar%20acquisition%2C%20especially%20morphology%2C%20how%20%0Athose%20pro
 cesses%20drive%20language%20variation%20and%20change%2C%20what%20insights%2
 0they%20%0Aprovide%20for%20low-resource%20NLP%2C%20and%20what%20they%20tell
 %20us%20about%20the%20%0Aintersection%20of%20(low-resource)%20NLP%20and%20c
 ognitive%20science.%20In%202020%2C%20he%20%0Areceived%20his%20PhD%20from%20
 the%20University%20of%20Pennsylvania%20Department%20of%20%0ALinguistics%2C%
 20where%20he%20worked%20with%20Charles%20Yang%20and%20Mitch%20Marcus.%20Pri
 or%20%0Ato%20that%2C%20he%20received%20a%20master's%20degree%20from%20the%2
 0University%20of%20%0APennsylvania%20Department%20of%20Computer%20and%20Inf
 ormation%20Science%20in%202018.%20%0AFrom%202013%20through%202015%2C%20he%2
 0was%20an%20Associate%20Scientist%20in%20the%20Speech%2C%20%0ALanguage%2C%2
 0and%20Multimedia%20group%20at%20Raytheon%20BBN%20Technologies%20where%20he
 %20%0Aworked%20on%20defense%20and%20medical-related%20projects.":Meeting Li
 nk: https://us06web.zoom.us/j/84282442460?pwd=NHVhQnJXOVdZTWtNcWNRQllaQWFnQ
 T09\nMeeting ID: 842 8244 2460\nPasscode: 678868\n\nAbstract: The success o
 f neural language models (LMs) on a wide range of \nlanguage-related tasks 
 may be in part due to their ability to induce \nhuman-like representations 
 or understanding of natural language \ngrammars. Humans are\, after all\, g
 old-standard language learners. For \nthe past several years\, researchers 
 pursuing this question have \ndeveloped a number of methodologies for testi
 ng the grammar \nrepresentations learned by LMs that have reached generally
  positive \nconclusions. I will take a critical look at such studies in thi
 s talk. \nWhile modern LMs are clearly extremely impressive\, and clearly d
 o often \ncapture important aspects of natural language grammars\, the \nme
 thodologies of many popular studies have unfairly overestimated the \ncapac
 ities of LMs when it comes to their ability to induce human-like \nrepresen
 tations. Focusing on questions of hierarchical syntactic \nrepresentations 
 and generalization in inflectional morphology\, I will \ndiscuss how uninte
 nded biases in data-splitting\, artificial training or \ntest data\, overly
  simplistic evaluations\, weak or absent baselines\, and \nfaulty interpret
 ations\, have conspired to overestimate the abilities of \nLMs. While the c
 onclusions of this study are largely negative in terms \nof the current sta
 te-of-affairs\, they are also optimistic. By employing \nmore thorough and 
 rigorous methodologies\, we have developed a better \nscientific understand
 ing of the nature of LMs and representations of the\n grammar. In identifyi
 ng weak points for current models\, we points \ntowards research areas wher
 e greater improvements may be gained.\n\nJordan Kodner is an Assistant Prof
 essor in the Stony Brook University \nDepartment of Linguistics and an affi
 liate of the Institute for Advanced\n Computational Science and Natural Lan
 guage Processing group. His \nprimary research revolves around computationa
 l approaches to child \nlanguage acquisition and their broader implications
 . In particular\, \nalgorithmic models of grammar acquisition\, especially 
 morphology\, how \nthose processes drive language variation and change\, wh
 at insights they \nprovide for low-resource NLP\, and what they tell us abo
 ut the \nintersection of (low-resource) NLP and cognitive science. In 2020\
 , he \nreceived his PhD from the University of Pennsylvania Department of \
 nLinguistics\, where he worked with Charles Yang and Mitch Marcus. Prior \n
 to that\, he received a master's degree from the University of \nPennsylvan
 ia Department of Computer and Information Science in 2018. \nFrom 2013 thro
 ugh 2015\, he was an Associate Scientist in the Speech\, \nLanguage\, and M
 ultimedia group at Raytheon BBN Technologies where he \nworked on defense a
 nd medical-related projects.
SEQUENCE:2
X-MOZ-GENERATION:5
END:VEVENT
END:VCALENDAR