How Does Statistical Learning of an Artificial Grammar Happen? A Behavioural Study


  • Nikita Zmanovsky University of Vienna



It was shown that human brain keeps track of linguistic units in speech processing on various levels, e.g. at the level of syllables, words, phrases and sentences [1]. Studies using artificial grammars instead of natural language input additionally suggest that this tracking of differently sized units can proceed also in the absence of specific language knowledge and on mere basis of the statistical relations present in the input.

In a previous study using artificial grammar learning approach[2], the authors found evidence for simultaneous neural tracking of both adjacent word-level and non-adjacent phrasal-level structures. That tracking was based on mere distributional information without any top-down language knowledge or additional prosodic cues to guide learning. However, there was no conclusive evidence of pattern learning on a behavioural level evident from the grammaticality judgement task.

The present research project builds upon that work and aims to address the encountered misalignment between neural and behavioral evidence. The main research question is the following: How does the learning of an artificial grammar happen on a behavioural level?


Healthy adult volunteers are participating in the study. The project will use an artificial grammar learning approach. The experiment consists of 2 stages:

  1. Learning stage. Participants are actively listening to a long sequence of nonsensical syllables (e. g., fi lo pa fu se ba). There is a hierarchical structure in the sequence: every 2 syllables constitute a „word“, every 3 „words“ constitute a „sentence“.
  2. Test stage. Participants are presented with new „sentences“ individually. They judge whether new sequences are “grammatical” or “ungrammatical”, based on the rules they learnt from the experiment.

Expected Results

We expect to find a pattern in participants‘ grammaticality judgements indicating the exact nature of representations they form during the learning phase. A between-subjects design with differently structured test items for each group will then allow conclusions about which types of structures were likely learned in the previous study and further allow inferences about which types of structures may facilitate learning. This research serves the purpose of a pilot study for subsequent experimental research involving EEG.


[1] N. Ding, L. Melloni, H. Zhang, X. Tian, and D. Poeppel, “Cortical tracking of hierarchical linguistic structures in connected speech,” Nature Neuroscience, vol. 19, no. 1, pp. 158–164, 2015. doi:10.1038/nn.4186

[2] I. Weyers, B. Walkenhorst, K. Nettermann, Ö. Bulca, T. Gruber, and J. L. Mueller, “ Neural entrainment reveals simultaneous, implicit learning of adjacent word and non-adjacent phrasal structure,” unpublished paper, 14th Annual Meeting of the Society for the Neurobiology of Language, Philadelphia, USA, 2022.