Using Information Divergence to Differentiate Deep from Superficial Resemblances Among Discourses

Abstract

In comparative analyses of discourses that reflect particular cultural identities, it is often necessary to differentiate superficial distinctions that arise primarily as cultural markers from deeper distinctions that arise from differences in cultural structures. In this paper, we build on previous work in order to operationalize this distinction between deep and superficial relationships between discourses using computational methods. To do so, we draw on the notion of divergence from information theory to measure the extent to which lexical items from a discourse act as signals of one cultural identity over another. We carry out a series of three types of comparisons between the discourses of fourteen English-language online discussion communities primarily focused on religion and spirituality. In the first type of comparison, discourses are compared at the level of individual words and their frequencies. In the second type, they are compared at the level of word-usage patterns learned from topic models. In the third, they are also compared at the level of word-usage patterns, but from topic models trained on their discourses after removing highly distinguishing terms that represent superficial distinctions between them. Our results indicate that, while some discourses share close resemblances both superficial and deep, others may appear to share close resemblances only superficially or may only share close resemblances after accounting for their superficial differences. These findings suggest that the approach we describe may be of use to researchers studying language in a variety of comparative contexts.