Estranged Predictions: Measuring Semantic Category Disruption with Masked Language Modelling