Combining NLP with evidence-based methods to find text metrics related to perceived and actual text difficulty

Abstract

Measuring text difficulty is prevalent in health informatics since it is useful for information personalization and optimization. Unfortunately, it is uncertain how best to compute difficulty so that it relates to reader understanding. We aim to create computational, evidence-based metrics of perceived and actual text difficulty. We start with a corpus analysis to identify candidate metrics which are further tested in user studies. Our corpus contains blogs and journal articles (N=1,073) representing easy and difficult text. Using natural language processing, we calculated base grammatical and semantic metrics, constructed new composite metrics (noun phrase complexity and semantic familiarity), and measured the commonly used Flesch-Kincaid grade level. The metrics differed significantly between document types. Nouns were more prevalent but less familiar in difficult text; verbs and function words were more prevalent in easy text. Noun phrase complexity was lower, semantic familiarity was higher and grade levels were lower in easy text. Then, all metrics were tested for their relation to perceived and actual difficulty using follow-up analyses of two user studies conducted earlier. Base metrics and noun phrase complexity correlated significantly with perceived difficulty and could help explain actual difficulty.