Identifying the meanings of grammatical elements in context is a major challenge for corpus-linguistic studies of grammatical variation. This study proposes a novel solution to this problem. I describe the situated meanings of grammatical elements as latent constructs, i.e., social concepts that cannot be observed directly but need to be inferred from the way that speakers behave. I use Latent Class Analysis (LCA) to create a data-driven typology of meanings for three modal periphrases in spoken Spanish and compare this typology to manual classification of the data in terms of modality. My findings show that (a) the situated meanings identified by the LCA do not directly correspond to the modal meanings that are commonly assumed to govern the variation between the three periphrases, and (b) the data-driven typology of meanings explains better the variation between these periphrases.
