A data science approach to readability

In this paper we study and test a novel hypothesis: namely, sentence complexity is a measure of readability. Our goal is not to replace nor necessarily improve readability tests but to investigate an experimental, data science application of word complexity. Word complexity is the number of words needed to define a word according to a reference dictionary. Sentence complexity is the aggregation or reduction of word complexity for words in a sentence. With the aid of natural language processing, we test our hypothesis using a corpus of 24 classic fairy tales. We chose these stories for their accessibility, popularity, and brevity. The data shows their reading difficulty to be centered around the 6
^th
-7
^th
grade level. The data also show the Flesch-Kincaid test and sentence complexity are negatively correlated in reading ease (r=-0.31, P<0.01) and positively correlated in grade level (r=0.46, P<0.01), respectively. That the correlations are moderate suggests sentence complexity is related to Flesch-Kincaid but not a proxy for it. Finally, the data further show that sentence complexity is weakly correlated with maximum number of syllables in a sentence (r=0.12, P<0.01), weakly correlated with maximum word length in a sentence (r=0.20, P<0.01), and moderately correlated with the number of words in a sentence (r=0.52, P<0.01). These findings support our hypothesis.