AI-built database of 378,000 English word families reveals how morphology shapes children’s reading

A team of reading researchers has unveiled what they describe as the largest morphological database of English ever constructed, and used it to map, with unprecedented scope, the word-building patterns that children encounter as they learn to read. The British and American Word Families project, or BAWF, covers approximately 378,000 words and was created by fine-tuning the large language model GPT-4.1 on thousands of expertly coded words before unleashing it on the vocabulary of contemporary British and American English. The findings, published in Frontiers in Psychology, suggest that the morphological complexity of children’s books rises steadily across the school years, that classical Greek and Latin roots appear far earlier in children’s print than many educators assume, and that a strikingly small set of prefixes and suffixes accounts for nearly all affixed words in the language.

Morphology is the system by which meaningful units, called morphemes, combine to form words. English inherits three distinct morphological systems: Germanic compounding, which fuses short concrete elements into forms like blackboard and afternoon; Romance affixation, which attaches prefixes and suffixes to Latin-derived bases as in disorder or preview; and Greek-Latin root morphology, the bound combining forms behind words such as biology and synthesis. Because robust links between morphological awareness and reading outcomes have been documented across diverse learner populations, researchers have long wanted a comprehensive picture of which morphemes children actually meet in print. Until now, existing databases were simply too small to answer that question at scale.

To build the new resource, the researchers, Clarence Green of the University of Hong Kong, Elfrieda Hiebert of TextProject, and Averil Coxhead of Victoria University of Wellington, started with the full open-source Wiktionary dictionary of roughly 8.7 million entries. They filtered it against two state-of-the-art corpora, the 100-million-word British National Corpus 2014 and the billion-word Corpus of Contemporary American English, retaining only words attested in contemporary use. That left about 378,000 words. The team then fine-tuned GPT-4.1 using 6,000 morphologically coded words drawn from MorphoLex, a rigorous human-generated database of about 70,000 words, training across 1,650 learning cycles to teach the model to segment words into roots, base words, prefixes, and suffixes.

The resulting parses were evaluated against human-constructed resources. For the 65,900 words appearing in both the BAWF and MorphoLex, overall agreement reached 73 percent, with component-level agreement of 79 percent for roots and 81 percent for base words. Those figures fall comfortably within the range of disagreement that exists between the human-built databases themselves, which varies from roughly 49 percent to 81 percent depending on which pair is compared. Morpheme counts estimated by the model also correlated strongly with MorphoLex counts at r = 0.80. The authors are candid that the database contains error and noise, and they emphasize that a fully accurate static coding of hundreds of thousands of words is limited by genuine theoretical ambiguity in morphology itself, where the boundary between a transparent derived form and an unanalyzable item is often gradient rather than categorical.

Applied to recent databases of children’s texts, the BAWF revealed a clear developmental trajectory. In picture books, about 29.91 percent of word types were morphologically complex, meaning they contained derivations or compounds. That figure climbed through juvenile fiction at 37.57 percent, and through UK children’s books, where complexity rose from 42.27 percent for ages seven to nine to 45.80 percent for readers aged thirteen and up. In United States core reading programs, the pattern was starker still: morphologically complex word types roughly doubled from 17.49 percent in Grade 1 to 37.34 percent in Grade 6, while compounds grew from 5.46 percent to 8.80 percent. Morpheme count also correlated with both grade-level placement and word-reading difficulty, showing that morphological demands accumulate systematically as texts become more challenging.

One of the most striking findings concerns classical roots. Approximately 20 percent of unique word types in picture books were estimated to contain a Greek-Latin root, rising to about 30 percent in books for older children and to 30 percent in Grade 6 reading materials, up from 19 percent in Grade 1. In terms of running words, classical-root vocabulary accounted for roughly 6 percent of tokens in picture books, climbing to about 8.5 percent in books for older readers and to around 13 to 15 percent of tokens in graded school reading programs. Words children encounter early include excited, from a root meaning to rouse or set in motion, captain, from a root meaning head, and incredible, from a root meaning to believe. The researchers argue that classical morphology should therefore not be treated as an exclusively upper-grade concern.

The database also allowed the team to quantify the productivity of individual roots. Of about 1,437 unique Greek-Latin roots identified, each supports an average of roughly 54 different word forms. The 100 most productive roots, including familiar teaching staples such as spec meaning look, ject meaning throw, tract meaning pull, and struct meaning build, together generate nearly 39,000 word forms. This multiplicative leverage is the theoretical foundation of morphological instruction: knowing one root potentially unlocks dozens of related words. Earlier estimates, notably work by Wei and Nation identifying 25 high-utility roots that access about 300 mid-frequency word families, had hinted at this concentration, but no corpus-based account of when such roots appear at scale in children’s print existed before.

The affix findings may prove the most immediately useful for educators. The BAWF contains 1,016 distinct prefix entries and 1,125 distinct suffix entries, yet the top 100 prefixes account for about 94 percent of all prefixed word forms, and the top 100 suffixes account for about 95 percent of suffixed forms. In every children’s corpus examined, suffixation dominated prefixation: for example, 22.62 percent of picture-book word types contained a top-100 suffix compared with 8.73 percent containing a top-100 prefix. Words carrying affixes outside the top 100 were extremely rare. In graded reading programs, top-100 prefix coverage rose from 5.85 percent in Grade 1 to 13.82 percent in Grade 6, while suffix coverage rose from 12.34 percent to 27.71 percent, and words carrying both a prefix and a suffix nearly quadrupled across the elementary grades.

These estimates extend decades-old benchmarks. Nagy and Anderson’s influential 1984 analysis concluded that roughly 60 percent of unfamiliar words in grades three through nine are derivational, and that learning one word might allow children to infer three or four more. White and colleagues showed in 1989 that just 20 prefixes covered nearly 97 percent of prefixed school words. The new database, with coverage far beyond any single frequency dictionary, confirms and refines this picture of a highly concentrated morphological system. The authors suggest that learning approximately 2,100 to 3,100 familiar word families is sufficient to cover 90 percent of the running words in the school-aged texts they examined, with 95 percent coverage requiring between 4,100 and 7,900 families.

The implications for literacy instruction are direct. If a small inventory of high-yield morphemes, including un-, re-, de-, -er, -ion, and -ize, plus a compact set of productive classical roots, captures most of the complex vocabulary children will ever meet, then morphology teaching can be prioritized, sequenced, and made efficient from the earliest grades. The BAWF is released openly under a Creative Commons license with an interactive web platform offering frequency data, family sizes, age-of-acquisition estimates, grade-level assignments, AI-assisted searches, and user-submitted parse corrections. The authors stress that version 1.0 is a revisable starting point, and that future work will add semantic transparency ratings, the key missing ingredient for updating the classic estimate that every word a child learns unlocks several more.

Subject of Research: Morphological patterns in children's print vocabulary estimated using an AI-built database of British and American word families

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

Article References: Green, C., Hiebert, E. H., & Coxhead, A. (2026). Patterns of morphology in children’s print: estimates of polymorphemic vocabulary using an innovative, AI-informed database of British and American word families. Frontiers in Psychology, 17, Article 1985734. https://doi.org/10.3389/fpsyg.2026.1985734

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1985734

Keywords: morphology, vocabulary, children's literacy, word families, large language models, GPT-4.1, reading development, Greek-Latin roots, prefixes and suffixes, corpus linguistics, morphological instruction, lexical database