Periodic Reporting for period 1 - LINGSHIFT (Linguistic Tuning Shifts in the Human Brain during Second Language Learning)
Periodo di rendicontazione: 2023-01-01 al 2025-06-30
Language learning and fostering linguistic diversity within the European Union are priorities of the European Commission. Learning new languages can change the structure and function of the cerebral cortex, affecting the way we think. However, despite extensive research, we lack a fine-grained description of how dynamical changes in cortical representations mediate learning concepts and linguistic structures of a new language. The LINGSHIFT program is using a cutting-edge paradigm to identify the dynamic shifts in linguistic representations by addressing three main questions: What are the core cortical regions involved in processing semantics (the meaning of words and sentences) and low-level linguistic information during language learning? What specific semantic or low-level linguistic information do these regions represent? Does the spatial organization of semantic and low-level representations dynamically change over the course of learning a new language? To investigate this, LINGSHIFT combines longitudinal non-invasive brain imaging using naturalistic stimuli and a powerful predictive modelling approach in a multidisciplinary work program. First, LINGSHIFT creates an open-access longitudinal functional MRI language comprehension and production dataset. This dataset will be recorded across a one-year period of adults learning a new language. Second, it identifies the cortical linguistic representations using a predictive modelling approach that makes use of state-of-the-art deep neural network language models. Third, it identifies the dynamical changes in linguistic representations during language acquisition. LINGSHIFT will provide new predictions about long-term dynamics and cortical changes that will improve our understanding of language systems in the brain. It will place important constraints on theoretical frameworks of second language learning that will help modernise language teaching.
Most of our work focused on (1) experimental design, stimuli selection, and developing a data plan (2) defining languages and participant pools, (3) comparing encoding modelling frameworks for fitting many features at the same time, (4) developing and testing different voxelwise encoding modelling pipelines with low-level and high-level semantic linguistic features to assess similarities and differences across language representations.
We presented our work on encoding low- and high-level features based on state-of-the-art large language models and speech models at the Society of the Neurobiology of Language conference. The comparison of encoding modelling methodology will be presented at the conference Computational Cognitive Neuroscience.
We presented our work on encoding low- and high-level features based on state-of-the-art large language models and speech models at the Society of the Neurobiology of Language conference. The comparison of encoding modelling methodology will be presented at the conference Computational Cognitive Neuroscience.
First results on method comparisons were presented at the Computational Cognitive Neuroscience 2025 Conference and Society for the Neurobiology of Language Conference. In the first work, we systematically investigate how variance partitioning and the residual method in combination with the voxelwise modeling approach can be used to determine the variance uniquely explained by a feature space in relation to other feature spaces.
In the second work, we compare how well transformer-based embeddings with different context lengths predict brain responses across different types of stimuli. Using voxelwise encoding models and the fMRI data, we investigate how language model context length affects prediction performance and whether these effects localize to specific brain regions. We showed that voxels prefer different context lengths depending on the amount of context present in the stimulus condition.
In the second work, we compare how well transformer-based embeddings with different context lengths predict brain responses across different types of stimuli. Using voxelwise encoding models and the fMRI data, we investigate how language model context length affects prediction performance and whether these effects localize to specific brain regions. We showed that voxels prefer different context lengths depending on the amount of context present in the stimulus condition.