Readability assessment of educational content helps ensure that texts are understandable for learners with different reading abilities. However, traditional models mainly rely on basic language features and often miss deeper patterns in organizing and expressing information. While rapid advancements in different Artificial Intelligence subfields, especially modern text embedding methods, have significantly improved readability assessment tools, existing models still need improvement. This is particularly true for educational texts, where continuous scoring is more effective than fixed-level classification to address diverse learner needs. This study proposes a graph-based method for readability assessment using Graph Convolutional Networks (GCNs) and a novel graph construction technique to represent textual structures. It models textual complexity by incorporating syntactic dependencies and assigning edge weights based on the part-of-speech tags of intermediate words. Additionally, Bayesian Optimization is used to fine-tune hyperparameters and graph construction configurations, improving the robustness and accuracy of the final model. Tested on the CLEAR dataset, the proposed method achieved an Formula: see text score of 0.9729. Also, the proposed method was evaluated on a classification-based dataset, and the predicted scores in the same classes were meaningfully close. The results confirm the effectiveness of the proposed method in modeling text complexity and producing accurate readability scores for educational purposes.
Zhang et al. (Fri,) studied this question.