Dogs Learn About Word Boundaries in Speech As Human Infants Do
By ELTE
This is the first demonstration of the
capacity to use complex statistics to learn about word boundaries in a
non-human mammal.
Human infants can spot new words in a speech stream much before they learn what those words mean.
To tell where a word ends and another one
begins, infants make complex calculations to keep track of syllable patterning:
syllables that usually appear together are probably words, and those that do
not probably aren’t. A new brain imaging study by Hungarian researchers discovered
that dogs may also recognize such complex regularities in speech.
“Keeping track of patterns is not unique to humans: many animals
learn from such regularities in the surrounding world, this is called
statistical learning. What makes speech special is that its efficient
processing requires complex computations.
To learn new words from continuous speech, it is not enough to count how often certain syllables occur together. It is much more efficient to calculate how probably those syllables occur together.
This is exactly how humans, even 8-month-old infants, solve the
seemingly difficult task of word segmentation: they calculate complex
statistics about the probability of one syllable following the other,” explains
Marianna Boros, one of the lead authors of the study, and a postdoctoral
researcher at the Neuroethology of Communication Lab, Department of Ethology,
Eötvös Loránd University.
“Until now we did not know if any other mammal can also use such
complex computations to extract words from speech. We decided to test family
dogs’ brain capacities for statistical learning from speech. Dogs are the
earliest domesticated animal species and probably the one we speak most often
to. Still, we know very little about the neural processes underlying their word
learning capacities.”
“To find out what kind of statistics dogs calculate when they
listen to speech, first we measured their electric brain activity using EEG,”
says Lilla Magyari, the other lead author, postdoctoral researcher in the same
research group, who had laid the methodological foundations of performing
non-invasive electrophysiology on awake, untrained, cooperating dogs.
“Interestingly, we saw differences in dogs’ brain waves for
frequent compared to rare words. But even more surprisingly, we also saw brain
wave differences for syllables that always occurred together compared to
syllables that only occasionally did, even if total frequencies were the same.
So it turns out that dogs keep track not only of simple statistics (the number
of times a word occurs) but also of complex statistics (the probability that a
word’s syllables occur together). This has never been seen in other non-human
mammals before. It is exactly the kind of complex statistics human infants use
to extract words from continuous speech.”
To explore how similar the responsible brain regions behind this
complex computational capacity in dogs are to those in humans, researchers also
tested dogs using functional MRI. This test was also performed on awake, cooperating,
unrestrained animals. For fMRI, dogs were previously trained to lay motionless
for the time of the measurements.
“We know that in humans both general learning-related and language-related brain regions participate in this process. And we found the same duality in dogs,” explains Boros. “Both a generalist and a specialist brain region seemed to be involved in statistical learning from speech, but the activation patterns were different in the two. The generalist brain region, the so called basal ganglia, responded stronger to a random speech stream (where no words could be spotted using syllable statistics) than to a structured speech stream (where words were easy to spot just by computing syllable statistics). The specialist brain region, the so called auditory cortex, that in humans plays a key role in statistical learning from speech, showed a different pattern: here we saw brain activity increase over time for the structured but not for the random speech stream. We believe that this activity increase is the trace word learning leaves on the auditory cortex.”
“We now begin to understand that some computational and neural
processes that are known to be instrumental for human language acquisition may
not be unique to humans after all,” says Attila Andics, principal investigator
of the Neuroethology of Communication Lab.
“But we still don’t know how these human-analog brain mechanisms
for word learning emerged in dogs. Do they reflect skills that developed by
living in a language-rich environment, or during the thousands of years of
domestication, or do they represent an ancient mammalian capacity? We see that
by studying speech processing in dogs, even better dog breeds with different
communication abilities and other species living close to humans, we can trace
back the origins of human specializations for speech perception.”
Reference: “Neural processes underlying statistical learning for
speech segmentation in dogs” by Marianna Boros, Lilla Magyari, Dávid Török,
Anett Bozsik, Andrea Deme and Attila Andics, 29 October 2021, Current Biology.
DOI: 10.1016/j.cub.2021.10.017
This research was funded by the Hungarian Academy of Sciences and
Eötvös Loránd Research Network (’Lendület’ Program), the European Research
Council (ERC) and the Ministry for Innovation and Technology.