Semantic composition in the brain: A representational similarity approach

Multiple fMRI studies have identified a set of candidate brain regions supporting simple forms of semantic composition (i.e., two-word combinations), most notably the left inferior frontal gyrus, angular gyrus, and left anterior temporal lobe. However, most of these studies adopted a univariate approach, limiting our understanding of semantic composition to changes in signals of single voxels in response to high-level contrasts. In the present work, we combine representational similarity analysis with computational models of semantic composition to re-analyze fMRI data aggregated from multiple published studies. We employ distributional semantics models, which characterize the meaning of single words as data-driven vectors whose proximity is proportional to the words’ semantic relatedness. Compositional extensions of such an approach characterize the meaning of word combinations via simple algebraic functions operating on pairs of word vectors. Different compositional models allow us to explicitly capture different forms of semantic composition. We conduct confirmatory analyses to test hypotheses of brain-model correspondence, connecting what is claimed about region-specific computation (e.g., the angular gyrus supports composition based on semantic relations) and what is known about model-specific representations. We present preliminary results suggesting that, when processing two-word combinations, information about single (constituent) words is available, and that this information pertains to both their stand-alone meaning and the meaning they take on when combined.