Readability Diagnosis System: A Comparative Study of Four Models on Elementary School Textbook

• Main Authors: Yu-Hsiang Tseng, 曾昱翔
• Other Authors: Chih-Wei Hue
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/g9d6em

博士 === 國立臺灣大學 === 心理學研究所 === 107 === Reading is an essential process through which people learn and communicate. In order to predict how comprehensible texts were for readers, readability studies identified text properties and build predicting models. Three genres of studies were differentiated in literatures basing on their text properties considered and modeling methods. In the genre of traditional methods, researchers used easily available text properties (e.g. percentage of difficult words, sentence length, etc.), and used models such as linear regression to predict readability. In the genre of cognitive science-inspired methods, readability studies started to incorporate theories from cognitive science and include more text properties relating to reading comprehension. Some of these properties could be extracted by computerized automatic text analysis tools. As models involved more and more properties, genre of statistical language modeling methods emerged. Researchers employed more elaborate models to predict readability. However, other researchers argued these elaborate models were not easily understandable by average users, and therefore affected how users would adopt the predictions. Purposes of current study were as follows: (1) four modeling approaches were differentiated by input transparency, which based on the relationship between model’s input data and readability literatures, and parameter transparency. Among them, a new modeling approach (i.e. the one with topic modeling) of low input and high parameter transparency was attempted to predict readability of text. This study implemented four models and compared their performances on predicting readability. (2) Literatures showed that text properties related to syntactic complexities affected reading comprehension, but these properties are not easily computable and thus few readability models directly incorporated them. This study built a frequency norm of phrasal patterns, based on which text properties were extracted to reflect syntactic complexities of phrases. These properties were then tested if they improved readability models. (3) In a survey study, we interviewed teachers in elementary schools, and collected their opinions on how willingly they were to adopt predicitons made by different readability modeling approaches. (4) A readability diagnosis system was developed. The system not only predicted readability but provided extra information on the properties affecting readability. Survey studies on teachers showed the diagnosis system could help them understand text properties and edit text.