An exploration of automated narrative analysis via machine learning

The accuracy of four machine learning methods in predicting narrative macrostructure scores was compared to scores obtained by human raters utilizing a criterion-referenced progress monitoring rubric. The machine learning methods that were explored covered methods that utilized hand-engineered features, as well as those that learn directly from the raw text. The predictive models were trained on a corpus of 414 narratives from a normative sample of school-aged children (5;0-9;11) who were given a standardized measure of narrative proficiency. Performance was measured using Quadratic Weighted Kappa, a metric of inter-rater reliability. The results indicated that one model, BERT, not only achieved significantly higher scoring accuracy than the other methods, but was consistent with scores obtained by human raters using a valid and reliable rubric. The findings from this study suggest that a machine learning method, specifically, BERT, shows promise as a way to automate the scoring of narrative macrostructure for potential use in clinical practice.