Modelling sentence deficits in Arabic agrammatism: A machine learning approach to the Tree Pruning Hypothesis

This study models sentence-level deficits in Palestinian Arabic agrammatism using supervised machine learning, with a focus on evaluating the Tree Pruning Hypothesis (TPH). TPH attributes sentence production impairments to disrupted access to higher syntactic projections - particularly CP-level structures - while predicting relative preservation of lower projections such as TP and AgrP. TPH is widely cited but has rarely been tested in morphologically rich, typologically diverse languages. We compared two machine learning (ML) models: a syntax-only model using structural features (question type, tense) and a full-feature Random Forest model incorporating morphosyntactic and clinical variables (Production Score, Error Type, Severity, demographics). The syntax-only model achieved 69% accuracy but failed to classify incorrect responses. In contrast, the full-feature model reached 96% accuracy with strong generalisation. Shapley Additive Explanations (SHAP) and permutation analyses revealed that clinical and morphosyntactic features were the strongest predictors, whereas syntactic structure contributed less than 3% of model variance. Our results support TPH's claim of CP-related vulnerability but challenge its hierarchical prediction: no gradient was observed between Wh-arguments and adjuncts, and deficits extended to lower projections. These findings suggest a flatter impairment profile, with similar difficulty across CP structures, shaped by processing demands and individual variability. This study demonstrates the value of theory-driven, multifactorial modelling for capturing sentence deficits in agrammatism and contributes new evidence from an underrepresented language variety.