In this paper, we propose a method for ”linguistic ethnography” – a general mechanism for characterising texts with respect to the dominance of certain classes of words. Using humour as a case study, we explore the automatic learning of salient word classes, including semantic classes (e.g., person, animal), psycholinguistic classes (e.g., tentative, cause), and affective load (e.g., anger, happiness). We measure the reliability of the derived word classes and their associated dominance scores by showing significant correlation across different corpora.