Shared computational principles for language processing in humans and deep language models

Общие вычислительные принципы обработки языка у человека и в глубоких языковых моделях
Uri Hasson, Samuel A. Nastase, Orrin Devinsky, Omer Levy, Roi Reichart, Yossi Matias, Daniel Friedman, Avinatan Hassidim, Lucía Melloni, Patricia Dugan, Werner Doyle, Adeen Flinker, Aren Jansen, Sasha Devore, Colton Casto, Ariel Goldstein, Mariano Schain, Zaid Zada, Bobbi Aubrey, Harshvardhan Gazula, Aditi Rao, Gina Choe, Catherine Kim, Michael P. Brenner, Dotan Emanuel, Kenneth A. Norman, Liat Hasenfratz, Eliav Buchnik, Amy Price, Amir Feder, Alon Cohen, Lora Fanda
2022-03-01

autoregressive deep language modelscontextual embeddingselectrocorticography (ECoG)next-word predictionpost-onset surprise
Departing from traditional linguistic models, advances in deep learning have resulted in a new type of predictive (autoregressive) deep language models (DLMs). Using a self-supervised next-word prediction task, these models generate appropriate linguistic responses in a given context. In the current study, nine participants listened to a 30-min podcast while their brain responses were recorded using electrocorticography (ECoG). We provide empirical evidence that the human brain and autoregressive DLMs share three fundamental computational principles as they process the same natural narrative: (1) both are engaged in continuous next-word prediction before word onset; (2) both match their pre-onset predictions to the incoming word to calculate post-onset surprise; (3) both rely on contextual embeddings to represent words in natural contexts. Together, our findings suggest that autoregressive DLMs provide a new and biologically feasible computational framework for studying the neural basis of language.
1
Autoregressive deep language models may offer a biologically feasible computational framework for investigating the neural basis of language.
2
Both neural responses and deep language models compare pre-onset predictions with incoming words to compute post-onset surprise.
3
ECoG recordings from nine participants listening to a 30-minute podcast provide empirical evidence for shared computational principles between brains and deep language models.
4
Human brains and autoregressive deep language models continuously predict upcoming words before they are spoken during natural narrative processing.
5
Humans and autoregressive deep language models represent words in natural contexts using contextual embeddings.

Human brain language processing and autoregressive deep language models processing the same natural narrative

Shared computational principles of continuous next-word prediction, post-onset surprise computation, and contextual word representation

Publication Details
Publication Date
2022-03-01
Journal
Publisher
ISSN
Cited by
480
Access Type
Author Information
Authors
Uri Hasson
Samuel A. Nastase
Orrin Devinsky
Omer Levy
Roi Reichart
Yossi Matias
Daniel Friedman
Avinatan Hassidim
Lucía Melloni
Patricia Dugan
Werner Doyle
Adeen Flinker
Aren Jansen
Sasha Devore
Colton Casto
Ariel Goldstein
Mariano Schain
Zaid Zada
Bobbi Aubrey
Harshvardhan Gazula
Aditi Rao
Gina Choe
Catherine Kim
Michael P. Brenner
Dotan Emanuel
Kenneth A. Norman
Liat Hasenfratz
Eliav Buchnik
Amy Price
Amir Feder
Alon Cohen
Lora Fanda
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat
Make a presentation
100%