Using Machine Learning to Predict Children’s Reading Comprehension from Lexical and Syntactic Features Extracted from Spoken and Written Language

Advances in natural language processing (NLP) and machine learning (ML) have introduced exciting prospects to educational research and practice. These technologies are poised to contribute to a deeper understanding of the linguistic and cognitive processes associated with successful reading comprehension, which is a critical aspect of children’s educational success. In this thesis, I used ML to investigate and compare associations between children’s reading comprehension and 260 linguistic features extracted through NLP from their speech and writing. Spoken and written language samples were gathered from 172 linguistically diverse children in Grades 4-6 using Talk2Me, Jr., an online language and literacy assessment platform. Lexical and syntactic linguistic features were extracted via a consolidated NLP pipeline. For the first research question, I compared eight supervised ML models predicting reading comprehension from the linguistic features, and then, using the best model, analyzed the 20 top predicting features. For the second question, I checked for differential functioning by examining interactions between top predictors and language-related demographics in predicting reading comprehension. For the third question, I used unsupervised ML to examine the latent factors constituting the linguistic features and explored how these factors predict reading comprehension differently from the ML models in the first research question. All three parts of the study were performed across four datasets: speech- and writing-elicited linguistic features, for both older/more skilled and younger/less skilled readers. The study contributes to the literature by concluding that suggest a substantial amount of variance in children’s reading comprehension can be predicted by productive grammar and vocabulary. A broad implication is that features of both spoken and written language features correlate with successful reading comprehension, but relationships differ whether individual features or multi-feature factors are used, and whether the data pertain to older/more skilled or younger/less skilled readers. There is evidence that some linguistic features may interact with language-related demographics in predicting reading comprehension, suggesting the need for further research. The study highlights how NLP and ML can enable nuanced examination of the language processes associated with reading comprehension and support innovations in language and literacy research and practice, but also that limitations exist and must be considered.Ph.D