Beyond Lexical and Syntactic: Better Text Simplicity Analysis Informed by Psycholinguistic Research
Abstract
AbstractAutomatic text simplification measures correlate poorly with human judgments and are disconnected from psycholinguistic research on comprehension. We propose a multidimensional analysis framework grounded in Cognitive Load Theory and the Construction-Integration model, organizing over 400 linguistic features into six simplicity dimensions targeting distinct comprehension stages. Using this framework, we analyze four human-annotated and six aligned simplification datasets. We find that (1) our dimensions predict human simplicity judgments better than established measures on most datasets; (2) current datasets overwhelmingly target lexical and syntactic dimensions while under-targeting personalization the dimension psychological research identifies as most impactful for comprehension; and (3) existing measures overall do not capture any specific dimension. These findings provide actionable guidance for building measures and datasets better aligned with how readers actually process text.