A Context-Aware, Feature-Enriched Bidirectional Long Short-Term Memory Framework for Automatic Document Readability Assessment
Abstract
Automatic readability assessment is important in the domain of educational content selection, language learning, digital accessibility, and personalized reading assistance. The traditional readability formulas quickly show a high dependence on bare-bones factors, and as such, they can not adequately assess the context of meaning, syntactic config, semantic entropy, discourse cohesion, and arranged structure types for degrees of ease. This work presents a framework for automatic document readability assessment which takes into consideration the context and other related features. They mixed several elements in their model, including RoBERTa-based contextual embeddings, a Bidirectional Long Short-Term Memory (Bi-LSTM) encoder, an attention mechanism, multidimensional linguistic features, gated feature fusion, and an ordinal-aware learning objective. These lexical, syntactic, semantic, and discourse cues are fused using an attention-weighted contextual representation to obtain a document-level feature vector. In the two-stage transfer-learning strategy, first, we pre-train our encoder on the CEFR-Based Sentence Profile corpus and then fine-tune it with the OneStopEnglish document-level corpus. The final objective involves the use of weighted categorical cross-entropy in conjunction with an ordinal-distance penalty to take into account class imbalance, as well as the seriousness of the error between more distant readability classes. Evaluation metrics: The evaluation metrics used include accuracy, macro-F1, Mean Absolute Error (MAE), Quadratic Weighted Kappa (QWK), adjacent level accuracy, and Spearman rank correlation. The approximation evaluation shows that after using the CEFR-SP pretrained model, our accuracy is 93.19%, macro-F1 is 92.96%, MAE is 0.078, and QWK is 0.956. Ablation studies also show that the model performance increases progressively through incorporating bidirectional encoding, attention, linguistic-feature fusion, auxiliary pretraining, and ordinal supervision. This framework forms a basis for effective context-sensitive and order-aware document readability prediction.