Which Words to Teach: Word Selection in the Age of Large Language Models

Abstract

English has a large and historically layered lexicon. The usual approach to vocabulary instruction, finding the unfamiliar words in an assigned passage and teaching them, improves comprehension of that passage but transfers poorly to broader measures. The problem has not been a lack of theory about which words merit instruction; frameworks for principled word selection have been well developed for more than a decade. What was lacking was a way for educators to apply those criteria, with the necessary corpus, morphological, and psycholinguistic analyses largely confined to research databases and the expertise of specialists. This review argues that large language models do not change what makes a word worth teaching, but rather they change what is now available to apply those criteria. Educators can now apply the criteria to a particular text and grade level using ordinary language. Drawing on four texts about hurricanes that hold the subject constant while varying genre and grade, the review develops three applications: domain-structured selection for informational text, semantic-cluster mapping for narrative, and the identification of a text’s latent lexicon that is particularly germane to narrative texts. It closes with the limits of these tools and identifies the points where teacher judgment is still required and the validation studies they still need.