Enhancing Automatic Readability Assessment with Verb Frame Features

Abstract

Patterns of verb usage, as encoded in verb frame elements, have been shown to correlate with text readability. While automatic readability assessment models have incorporated a wide range of features on lexical, syntactic, semantic and discourse properties of a text, none of these features directly exploit the semantic information provided in verb frames. This paper investigates whether features based on semantic frames can improve the performance of Chinese readability assessment models. We conduct experiments with manually annotated verb frame elements in Mandarin VerbNet, including the use of non-core frame elements, subject omission, metaphoric usage and clause as verb argument. Experimental results on a corpus of Chinese texts, spanning twelve school grades, show that these frame features significantly raise readability assessment accuracy over a baseline model that relies on surface features only.