It is relatively uncontroversial that word meaning can give rise to behaviourally and neurally measurable phenomena even when words are perceived outside of awareness, as in tricky experimental conditions (e.g., masked priming, continuous flash suppression) or in pathologica conditions (e.g., neglect, coma). It is less clear, however, exactly what kind of semantic information is extracted from unconsciously perceived words. Is it the entire galaxy of each word's complex network of relationships? Or is it only part of this galaxy, perhaps the most superficial one?
One way in which this question can be formulated quantitatively is based on the often neglected fact that words related in meaning also tend to occur together in language use. For instance, corpus analyses have shown that words referring to the same perceptual modality (e.g., red and transparent) tend to be used in more similar contexts than words referring to different modalities (e.g., red and loud). Likewise, words referring to entities with similar perceptual or conceptual properties tend to co-occur (e.g., when people talk about dogs they are more likely to also talk about cats than about apples); and they tend to be used in similar contexts (e.g., both cat and dog tend to occur in the context of conversations about pets, whether or not they co-occur within a given conversation). We model this complex network of relationships with some mathematical tricks, trying to disentangle more shallow aspects of meaning from deeper ones; and then we test in the lab which of these aspects are captured by speakers and readers when they're unaware that words were presented to them.

This work is carried out in collaboration with Daniel Casasanto (Cornell Univerisity) and Marco Marelli (University of Milano-Bicocca).