Text Difficulty Classification by Combining Machine Learning and Language Features

Abstract

AbstractIn recent years the application of deep learning algorithms has dominated this research field on document readability prediction. Traditional methods rely excessively on manual feature extraction, and modern deep learning algorithms are severely time consuming in terms of efficiency in deep feature extraction. On the other hand, there is a considerable lack of common datasets for readability studies in the relevant English language education field. In view of these problems, we proposed an improved approach for corpus construction and reconstructed a more reasonable dataset (English Teaching Texts in China, CETT) with five-level difficulty by purposefully integrating and adding some missing datasets. Based on CETT, we extracted three dimensions of text word frequency features, linguistic difficulty features and in-depth features. We compared the effectness of combining features of different features for difficulty prediction under multiple classifiers. The final experimental results show that the accuracy of fused linguistic difficulty features and in-depth features on the text difficulty assessment task reaches 0.9011, and the fused features have an overall significant improvement over using individual features.KeywordsMachine learningReadability assessmentTransformerDifficulty classification