We will use (1) a state-of-the-art technique (Representational Similarity Analysis) that has heralded a new research era in the study of semantics since it allows one-to-one mappings between patterns of brain-activity measurement, behavioural and computational models and (2) dimensionality-reduction approaches.

We know two main contrasting approaches to the study of meaning:

embodied theories propose similarity dimensions arisen by our sensory-motor experience;

distributional theories propose similarity dimensions derived by the distribution of words across spoken and written language.

Which similarity measure between concepts best conforms to similarity in patterns of brain activity?

Figure 2: Examples of different semantic representations (similarity matrices) based on behavioural, experiential, distributional, and combined models, as well as their representational similarity (lower right corner).

We know that embodied and distributional similarity dimensions are not totally independent but rather complementary.

Combined information of the two data sources would make it possible to capture similarity structures that define categories of meaning, such as the cluster of features that places “grape” and “marble” in different categories despite a superficial visual similarity, and “cherry” and “banana” in the same category despite very different appearances.

Does a similarity measure derived by combined models best reflect similarity in pattern of brain activity compared to single-source models?

How does our brain organize different concepts into meaningful categories based on this integrated information?