The anatomy of specialized knowledge: Comparing experts and non-experts through associations, frames and language models

Abstract

Abstract We explore specialized knowledge and aim to show that expert conceptual spaces differ from those of non-experts. This rather broad research question is addressed from different perspectives: first we collect free associations to selected stimulus terms from the domain of karstology from experts and non-experts, demonstrating that the underlying knowledge affects the associative inventory and that the overlap between both groups is low. Next, we look for knowledge frames which might shape the expert responses by activating conceptual links, and compare them to corpus-derived frames. Finally, we train neural language models on specialized versus general corpora to see whether the neural semantic space as represented by the cosine distance resembles the semantic spaces obtained through human associations. Results show that human associations indeed reflect knowledge frames, but that the overlap with the trained word embeddings is again low, indicating inherent differences between the associative semantic proximity in experts and non-experts, and between humans and neural meaning representations.