Cognitive Complexity Detection in Student Writing Through Transformer-Based Language Models
Abstract
The cognitive complexity in student writing reflects the depth of learning and critical thinking that is essential for educational assessment. However, traditional rubric-based ratings are often inconsistent among evaluators, subjective, and time-consuming. This study introduces a transformer-based formal Cognit-Write using a RoBERTa language model to detect cognitive complexity in student writing texts. This approach begins with the collection and categorization of student articles based on Bloom's taxonomy, classifying them into distinct cognitive layers. Pre-processed speeches are fed into a sophisticated RoBERTa model, which is trained to classify writing under complex conditions, such as tokenization, recall, understanding, analysis, and set. The sample attention mechanisms and environmental embeddies help identify cognitive markers, such as argument depth, compression, rationality, and conceptual layer. Rated educational database, Cognit-Write reaches 89.2% F1-scoring, performs better than basic classification and traditional NLP techniques. The study concludes that the detection of cognitive complexity in the RoBERTa-based cognitive model is a measuring, accurate, and interpreting method for improving automatic educational feedback systems and learning analysis.