Modelling the acquisition of spelling

Learning to spell German words can be hard. Beginning spellers often produce forms thatdeviate from orthographically correct ones in various ways. In this paper, we demonstratethat the proclivity to misspell a given word can be estimated using the Discriminative Lex-icon Model (DLM, Baayen et al. 2019). DLM maps form representations onto meaningrepresentations using simple linear networks, and, in the case of written language, alsophonemic representations onto graphemic representations. We train three different DLMmodels: one relates single phonemes to single graphemes; one relates phoneme bigrams tographeme bigrams, and one relates phoneme trigrams to grapheme trigrams. As a test case,we use a longitudinal corpus of picture story descriptions of primary school children, LitKey(Laarmann-Quante et al., 2019). We let the three models predict words that children actuallyspelled (or misspelled), and we use Generalized Additive mixed-effect models to determinethe alignment, and the influence that factors like age and proficiency have on the alignment.Our results suggest that context becomes more important as children become more profi-cient. Beginning spellers can best be approximated with the unigram model; intermediatespellers are best approximated by the bigram model; and children at the end of primary schoolare best approximated by the trigram model. Even though we only take phonological formsinto account, the trigram model successfully captures spellings that are traditionally describedas morphological. We interpret our results as an argument in favor of statistical learning andthe increasing influence of grain size over time for developing spellers.