Patterns of morphology in children’s print: estimates of polymorphemic vocabulary using an innovative, AI-informed database of British and American word families

Morphological complexity is a pervasive feature of the vocabulary children encounter in print, yet resources for investigating morphology remain limited. What morpheme patterns do children encounter in print? Which morphemes cover how many polymorphic words in reading experiences? In this study we introduce the British and American Word Families (BAWF) project, an innovative open database of  378,000 words, and use it to address questions such as these.

Methods: GPT-4.1 was fine-tuned using  6000 morphologically-coded wordsfrom an existing human-generated database (Morpholex). The fine-tuned LLMcoded word-parts in all Wiktionary words attested in large British and Americancorpora. Performance was evaluated via agreement with existing databases.Classical roots were attributed to a subset of words by mapping a list of rootsfrom Wikipedia to all related entries in Wiktionary. The resulting BAWF databaseis then used to estimate morpheme patterns across databases of children’s textsand further developed as an online application.

Results: LLM word part predictions are good and fall within the rangeof agreement typical of existing databases. Nonetheless, output containssome error/noise. In children’s print, polymorphemic words increase acrosspicturebooks to juvenile fiction and school-reading programs. Greek-Latin rootwords occur early and commonly. Estimates include that  100 prefixes/suffixesmay account for nearly all derivations. The BAWF itself is released open-sourcehttps://cgg-projects.github.io/BAWF_V1/, with many lexical, morphological, and psycholinguistic variables, including but not limited to frequency, family size, and morpheme-counts.

Discussion: The contribution suggests that instruction that prioritizes highyield morphemes could give children access to large word families. These morphemes are present and common early in print. The BAWF (V1) helps see which may be useful to teach. Users can generate custom wordlists, targeting particular features and families, and access linked dictionary/etymologicalinformation. The platform includes AI-assisted searches, explanations, parsing review/ corrections, and instructional support. The BAWF project will develop
over time, incorporating additional variables (e.g., morphological transparency), improving accuracy and the user interface.