Not All Words Are Equally Acquired: Transitional Probabilities and Instructions Affect the Electrophysiological Correlates of Statistical Learning

Abstract

Statistical learning (SL), the process of extracting regularities from the environment, is a fundamental skill of our cognitive system to structure the world regularly and predictably. SL has been studied using mainly behavioral tasks under implicit conditions and with triplets presenting the same level of difficulty, i.e., a mean transitional probability (TP) of 1.00. Yet, the neural mechanisms underlying SL under other learning conditions remain largely unknown. Here, we investigated the neurofunctional correlates of SL using triplets (i.e., three-syllable nonsense words) with a mean TP of 1.00 (easy "words") and 0.50 (hard "words") in an SL task performed under incidental (implicit) and intentional (explicit) conditions, to determine whether the same core mechanisms were recruited to assist learning. Event-related potentials (ERPs) were recorded while participants listened firstly to a continuous auditory stream made of the concatenation of four easy and four hard "words" under implicit instructions, and subsequently to another auditory stream made of the concatenation of four easy and four hard "words" drawn from another artificial language under explicit instructions. The stream in each of the SL tasks was presented in two consecutive blocks of ~3.5-min each (~7-min in total) to further examine how ERP components might change over time. Behavioral measures of SL were collected after the familiarization phase of each SL task by asking participants to perform a two-alternative forced-choice (2-AFC) task. Results from the 2-AFC tasks revealed a moderate but reliable level of SL, with no differences between conditions. ERPs were, nevertheless, sensitive to the effect of TPs, showing larger amplitudes of N400 for easy "words," as well as to the effect of instructions, with a reduced N250 for "words" presented under explicit conditions. Also, significant differences in the N100 were found as a result of the interaction between TPs, instructions, and the amount of exposure to the auditory stream. Taken together, our findings suggest that triplets' predictability impacts the emergence of "words" representations in the brain both for statistical regularities extracted under incidental and intentional instructions, although the prior knowledge of the "words" seems to favor the recruitment of different SL mechanisms.

Keywords: artificial language; electrophysiological correlates; explicit learning; exposure time; implicit learning; statistical learning; transitional probabilities; word segmentation.

Figure 1

Visual summary of experimental design. Note: Box (A) illustrates the timeline of the experimental procedure in which one implicit and, subsequently, one explicit auditory statistical learning (SL) tasks were administered. Each task of the two tasks comprised of three parts: instructions, familiarization phase, and test phase. As can be observed in Box (B), each task was initiated with specific instructions that determined the conditions under which SL occurred: implicit instructions (i.e., without knowledge of the stimuli or the structure of the stream—Implicit task) or explicit instructions (i.e., with explicit knowledge or pre-training on the “words” presented in the stream—Explicit task). In the familiarization phase of both tasks, participants were presented with a continuous auditory stream of four easy and four hard “words,” with chirp sounds (depicted as a speaker icon on the Figure) superimposed over specific syllables. The chirp sounds could emerge at any of three-syllable positions of the “words,” which precluded its use as a cue for stream segmentation. During this phase, participants had to perform a chirp detection cover task. Then, a test phase consisting of a two-alternative forced-choice (2-AFC) task asked participants to indicate which of two syllable-sequences (a “word” and a foil) sounded more familiar considering the stream heard on the familiarization phase.

Figure 2

Percentage of correct choices (% hits) for the easy- and hard-nonsense words in the 2-AFC tasks performed under implicit and explicit conditions.

Figure 3

Learning effects on N100 Peak. Note: (A) grand average ERPs at the fronto-central ROI (solid line: hard; dotted line: easy). Gray shaded boxes over the event-related potentials (ERPs) indicate the analyzed time window (80–120 ms). (B) Voltage maps of each condition: fronto-central distribution of the N100 peak. (C) Graphical depiction of the averaged amplitudes for the pairwise comparisons of the triple interaction between SL task, type of “word,” and exposure time.

Figure 4

Effect of instructions on the N250 response. Note: grand average ERPs at the fronto-central region of interest (ROI; solid line: explicit; dotted line: implicit). Gray shaded box over the ERPs indicates the analyzed time window (230–270 ms). Voltage maps of each condition: fronto-central distribution of the N250 peak.

Figure 5

Effect of type of “word” in the N400 Time Window. Note: effect of type of “word” in the N400 time window at the central ROI (solid line: hard; dotted line: easy) and voltage maps of the difference between easy and hard “words.”