A data-driven cognitive feature–based model for english text readability assessment to support college english instruction

Text readability assessment is important for college English instruction and reading material selection. Traditional readability formulas mainly rely on surface linguistic features, such as sentence length, word length, syllable counts, and word familiarity, but they often fail to represent reading-related processing demands beyond surface form. To address this limitation, this study proposes a cognitive-feature-based readability assessment framework that integrates three cognitively motivated, text-derived indicators with conventional linguistic features. The indicators comprise a Rarity Index, a Logical Complexity Index based on knowledge-network structure, and an Understanding Difficulty Index reflecting textual sequencing. Experiments are conducted on four benchmark datasets, including CEFR, CLEC, OneStopEnglish, and RACE, and the framework is compared with traditional readability formulas, conventional machine learning models, pretrained language models, and hybrid models. Evaluation metrics include Accuracy, F1-score, Quadratic Weighted Kappa, Mean Absolute Error, and RRNSS. A controlled feature-set ablation compares otherwise identical model configurations before and after adding the cognitive feature block. Across 24 matched comparisons, QWK improves in all 24 comparisons, ACC in 22, F1 in 23, and MAE decreases in 23. The best hybrid model, RoBERTa + Cog-Core, achieves the highest average ACC among the evaluated models (0.648) and improves over RoBERTa alone on all four datasets. These findings support the incremental and interpretable value of cognitively motivated textual proxies for readability assessment. The framework does not claim to directly measure human cognitive states, but it can provide interpretable support for matching reading materials with learners’ proficiency levels in college English instruction.