Figure 1. .

Example of an in silico adaptation of a controlled experiment. (A) The original MEG study investigated composition over two-word phrases (Bemis & Pylkkänen, 2011). This was done by presenting three different types of phrases to participants to solve a picture matching task. By contrasting the elicited brain responses in the composition condition with the responses in the list and non-word conditions, the authors could infer which brain regions are engaged in compositional processing of two-word phrases. (B) This experimental paradigm can be conceptually simulated with LM-based fMRI encoding models of naturalistic stimuli. The composition and list conditions can be tested by using the learned encoding model to predict each voxel’s response to a large, diverse corpus of phrases. The non-word condition can be simulated by replacing the first word in a phrase with a non-word, extracting new ablated features of the phrase from the LM and using the encoding model to predict the brain’s response to the ablated phrase. If a voxel’s response is highly sensitive to the removal of the first word, it would suggest that the voxel combines information over both words to arrive at meaning. This provides a data-efficient way to test for compositional processing across diverse types of phrase constructions. fMRI = functional magnetic resonance imaging; LM = language model; MEG = magnetoencephalography.