34.758, FYI: C-STAR Lecture 3/10, Evelina Fedorenko: Language models

[The LINGUIST List]

LINGUIST List: Vol-34-758. Sun Mar 05 2023. ISSN: 1069 - 4875.

Subject: 34.758, FYI: C-STAR Lecture 3/10, Evelina Fedorenko: Language models

Moderator: Malgorzata E. Cavar, Francis Tyers (linguist at linguistlist.org)
Managing Editor: Lauren Perkins
Team: Helen Aristar-Dry, Steven Franks, Everett Green, Sarah Robinson, Joshua Sims, Jeremy Coburn, Daniel Swanson, Matthew Fort, Maria Lucero Guillen Puon, Billy Dickson
Jobs: jobs at linguistlist.org | Conferences: callconf at linguistlist.org | Pubs: pubs at linguistlist.org

Homepage: http://linguistlist.org

Please support the LL editors and operation with a donation at:
           https://funddrive.linguistlist.org/donate/

Editor for this issue: Everett Green <everett at linguistlist.org>
================================================================


Date: 
From: Dirk Den Ouden [denouden at sc.edu]
Subject: C-STAR Lecture 3/10, Evelina Fedorenko: Language models


Friday March 10th, 12pm (noon) ET
Presentation in Zoom, accessible via the C-STAR website:
http://cstar.sc.edu/lecture-series/

The utility of large language models in understanding the human
language system.

Evelina Fedorenko, PhD

Massachusetts Institute of Technology

Until a few years ago, the field of language research has lacked
computational models that can take as input an arbitrary linguistic
string and represent it in a way so as to enable a broad range of
downstream tasks, from text generation, to question answering, to
translation into another language. Now, we have a class of such
models—commonly referred to as 'large language models’ (LLMs)—that
have taken the fields of Natural Language Processing and Artificial
Intelligence by storm. Aside from revolutionizing AI, these models
have tremendous potential for transforming the field of language
research, including potential applications in work on developmental
and acquired language disorders. In this talk, I will start by
presenting evidence that representations extracted from some LLMs can
capture human neural responses (as measured with fMRI or intracranial
recordings) during the processing of the same language stimuli.
Moreover, models that perform better on the next-word prediction task
are better able to predict human responses, but performance on other
language tasks (e.g., grammaticality judgments) does not bear a
relationship to brain predictivity. This selective relationship
between next-word prediction performance and brain predictivity
suggests that optimizing for predictive representations may be a
shared objective of both biological and in silico language systems. I
will then talk about more recent work that builds on these original
findings to better understand the necessary and sufficient conditions
for a model to capture human neural responses. I will talk about a
study that shows that models that are trained on developmentally
plausible amounts of data already provide a good match to human neural
data. This result helps address the major criticism of LLMs as models
of human language processing: namely, that they are trained on vastly
more data than what human children get exposed to. Finally, I will
talk about a series of studies where the model representations are
altered before relating them to human neural data in an effort to
identify the critical features of the representations that mediate the
model-to-brain match. These studies suggest that the key contributor
to the model-to-brain match is the lexical-semantic content rather
than syntactic cues like word order and function words. I will
conclude by outlining promising future directions of this research
program, including its clinical applications.
_______________________________________________

The online lecture can be followed online from your computer, tablet
or smartphone, in Zoom. The zoom link is accessible via the C-STAR
website: http://cstar.sc.edu/lecture-series/

For more information, or to be added to the C-STAR mailing list,
contact Dirk den Ouden: denouden at sc.edu

Linguistic Field(s): Clinical Linguistics
                     Cognitive Science
                     Linguistic Theories
                     Neurolinguistics
                     Psycholinguistics




------------------------------------------------------------------------------


LINGUIST List is supported by the following publishers:

Bloomsbury Publishing (formerly The Continuum International Publishing Group) http://www.bloomsbury.com/uk/

Cascadilla Press http://www.cascadilla.com/

Dictionary Society of North America http://dictionarysociety.com/

Georgetown University Press http://www.press.georgetown.edu

John Benjamins http://www.benjamins.com/

Lincom GmbH https://lincom-shop.eu/

Multilingual Matters http://www.multilingual-matters.com/

Narr Francke Attempto Verlag GmbH + Co. KG http://www.narr.de/

Netherlands Graduate School of Linguistics / Landelijke (LOT) http://www.lotpublications.nl/

Springer Nature http://www.springer.com

Wiley http://www.wiley.com


----------------------------------------------------------
LINGUIST List: Vol-34-758
----------------------------------------------------------