Taking It Slow, But Quickly: How Linguistic Structures Form

Source: Pexels/Yoss Cinematic

While speech may be strictly linear, there are no immediate acoustic clues in the speech we hear as to where one word ends and another begins. And so the infant brain immediately assumes the role of a detective. Being a child in an age of global pandemics is problematic enough without having to worry about whether the word “run” is being used as a noun or a verb. The “blooming, buzzing confusion” of experience maybe just so, but what hits us from the external world is neatly packaged into different types of linguistic structures — syllables, words, sentences. Statistical regularities in experience are readily detected in order to achieve this.

What exactly are we learning, though, when we learn how to parcel speech into words? At a minimum, we need to be able to learn how to form hierarchical structures such as multi-syllabic units, regardless of the actual content of these units. These larger units can then be used as the basis for forming new words. Linguists have long argued that this is a highly abstract process; it should occur regardless of the type of words we ultimately create. New research attempts to address the neural basis of this process and to "detach" it from any particular semantic interpretation.

Research from Simon Henin and colleagues, based in New York and Connecticut, and published last month in Science Advances, investigated this issue [i]. They collected recordings from electrodes directly implanted on the surface of the brain, or penetrating gray matter, in human epilepsy patients. Using a "statistical learning" paradigm, the authors sought to identify electrode sites responsive to regularities of their stimuli over different scales, such as syllables and words. They presented patients with auditory recordings of streams of syllables in which sequence structure was manipulated. In structured streams, syllables were placed into the first, second, or third position of a three-syllable “triplet." These never formed a meaningful lexical concept, but were simply English syllables. For instance, “to," “pi,” and “ro” could form “tupiro” and this triplet would repeat later in the auditory recording. Alternatively, the three syllables could simply be placed at random in the recording. The auditory recordings were therefore created by randomly inserting multiple repetitions of the “word” triplets without pauses or prosodic cues (e.g. rising/lowering intonation) for the structured stream, while the random stream involved no such multi-unit repetitions (also without pauses/cues).

The authors found neural entrainment (involving the synchronization of external quasi-periodic stimuli and internal neural activity) at the syllabic frequency (4 Hz) in both types of audio recordings. During exposure to the structured sequences, entrainment to the triplet boundaries (1.33 Hz) emerged. The latter responses were found mainly in somatosensory/motor and temporal electrode sites, and also inferior frontal gyrus and hippocampus. This “word”-rate response also emerged as early as 50-word exposures in some patients, an early effect that temporally maps onto previously documented behavioral results.

The authors also implemented a visual analog to the auditory experiment, using colored fractals which were either presented randomly or in ordered pairs, finding that hippocampus uniquely tracked higher-order structural information for both the auditory and visual tasks.

Neural sites representing immediate sensory input and higher-order units (i.e. syllabic information) appeared to encode information such as the transitional probabilities of the stimuli; or, how likely a given syllable was to appear at any moment. Sites that exclusively represented higher-order structure (and not lower-level syllabic information) encoded aspects of the sequences such as the ordinal position of the elements, but also the identity of the learned unit.

Of particular interest is the finding that these coding effects were only observed when the authors examined the raw field potential recordings, but not when the same analyses were conducted on the high gamma band envelope. They suggest that this “specifically points to low-frequency information as a possible mechanism for encoding these representations, frequencies well documented to facilitate information transfer across the cortex and subcortical regions." While high gamma activity has been shown to index local cortical processing, the work of Henin and colleagues continues a recent line of investigation highlighting the clear importance of lower, slower brain rhythms in processes that seem to require longer-range connectivity profiles than faster gamma activity [ii]. Whatever higher-order processing is occurring, it appears that marking the identity of a learned hierarchical structure (in, for instance, hippocampus or inferior frontal gyrus) involves forms of neural activity that go beyond local synchronization.

This work reveals how the learning of hierarchical, structured sequences depends on distinct cortical and subcortical networks and that patients learn (unconsciously) such structures rapidly, within a matter of minutes. At the same time, this rapid learning process is grounded in low-frequency tracking. How these learning processes develop in the human brain, and whether they are recruited for learning other forms of sequential information, remain open questions.