BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Causal analysis of the syntactic representations of Transformers -
  Tal Linzen\, New York University
DTSTART:20211021T140000Z
DTEND:20211021T150000Z
UID:TALK164296@talks.cam.ac.uk
CONTACT:Marinela Parovic
DESCRIPTION:The success of artificial neural networks in language processi
 ng tasks has underscored the need to understand how they accomplish their 
 behavior\, and\, in particular\, how their internal vector representations
  support that behavior. The probing paradigm\, which has often been invoke
 d to address this question\, relies on the (typically implicit) assumption
  that if a classifier can decode a particular piece of information from th
 e model's intermediate representation\, then that information plays a role
  in shaping the model's behavior. This assumption is not necessarily justi
 fied. Using the test case of everyone's favorite syntactic phenomenon - En
 glish subject-verb number agreement - I will present an approach that prov
 ides much stronger evidence for the *causal* role of the encoding of a par
 ticular linguistic feature in the model's behavior. This approach\, which 
 we refer to as AlterRep\, modifies the internal representation in question
  such that it encodes the opposite value of that feature\; e.g.\, if BERT 
 originally encoded a particular word as occurring inside a relative clause
 \, we modify the representation to encode that it is not inside the relati
 ve clause. I will show that the conclusions of this method diverge from th
 ose of the probing method. Finally\, I will present a method based on caus
 al mediation analysis that makes it possible to draw causal conclusions by
  applying counterfactual interventions to the *inputs*\, contrasting with 
 AlterRep which intervenes on the model's internal representations.
LOCATION:https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBd
 XVpOXFvdz09
END:VEVENT
END:VCALENDAR