Page 9 - ELG2205 May Issue 480
P. 9
NEWS
How deep is your language?
By Gillian Ragsdale while their brain activity was
recorded via electrocorticography
Computers using deep language (ECoG). ECoG is a more sensitive
models (DLMs) help us understand variation on the more familiar
how the brain processes language, EEG (electroencephalography). In
according to a new study by an EEG, electrodes are placed on the
international team of researchers scalp, but in ECoG they are placed
from American, German and on the actual surface of the brain.
Israeli universities. The nine volunteers were having
Traditionally, psycholinguistic ECoG monitoring for other clinical
research has modelled language reasons and 1,339 electrodes were
in terms of how items such as implanted on each brain to record
nouns or verbs follow rules. activity while listening to part
DLMs come at language of the story ‘So a Monkey and a 2. Both human brains and GPT-2 does not. DLMs cannot ‘generate
modelling very differently, with no Horse Walk Into a Bar: Act One, produced a ‘surprise’ response, a new meaningful ideas by integrating
‘preconceptions’ of any kind – not Monkey in the Middle’. kind of error signal in response prior knowledge’. But the human
even that there are in fact parts of A DLM (GPT-2) also ‘listened’ to the actual next word. brain can – and further research
speech or syntax. to the same story and its processing 3. The meaning of words were could analyse how it does that using
When DLMs are exposed to was monitored. constantly evaluated with abilities such as the spontaneous
language they learn by constantly Three main processes were reference to preceding words: predictive text and feedback
predicting the next word and shown to be similar in the human altering one word at the start of demonstrated in this study.
using feedback from the degree brain and DLM: the story altered the following
of ‘surprise’ to inform future neural activity for all subsequent REFERENCE
predictions. It has long been 1. The recorded brain activity sentences. n Goldstein, A, Zada, Z, Buchnik, E.
theorised that human brains do clearly indicated that listeners et al (2022). ‘Shared computational
something similar, but it has been were constantly predicting the To the extent that a DLM can principles for language processing in
fiendishly difficult to measure such next word in the story, hundreds mimic these brain processes to humans and deep language models’,
specific, rapid activities in the brain. of milliseconds before the next produce apparently meaningful Nature Neuroscience 25: 369-380,
In this study, nine participants word and clearly separate to sentences in context, does this https://doi.org/10.1038/s41593-
listened to a 30-minute podcast processing the actual next word. constitute ‘thinking’? At present it 022-01026-4 OPEN ACCESS.
HELBLING ENGLISH
H E L B L I N G E N G L I S H
The best start
for very young learners
Second Edition
Let’s play! New culture, phonics
and review pages
helbling.com/english
Teacher’s DIGI Pack
Flashcard games
Songs, Chants & Videos
editorial@elgazette.com 9