Deep learning-based automatic assessment of aesthetic expression in english writing: A multi-task learning approach with cross-cultural validation

Although AES systems have made significant progress in evaluating such surface-level linguistic features as grammar, vocabulary, and text structure, the systematic assessment of aesthetic expression in writing is still a significant research gap due to its subjectivity, multi-dimensionality, and cultural variability. This study aims to develop a deep learning-based automatic assessment system for aesthetic expression in English writing. Three research questions are proposed: one regarding the effectiveness of deep learning for aesthetic evaluation, another concerning the advantages of multi-task learning over a single-task approach, and the third addressing the cross-cultural generalization capability of the proposed model. A four-dimensional aesthetic assessment framework including rhetorical application, imagery description, emotional conveyance, and stylistic unity was constructed. Besides, the proposed system leveraged a multi-task learning architecture based on RoBERTa to achieve knowledge transfer and joint optimization across dimensions. Then, it systematically validated across four cultural background groups, including East Asian, Southeast Asian, European, and American learners. The proposed method showed substantial agreement with human ratings and outperformed all the baseline models. Multi-task learning demonstrates consistent improvements for all aesthetic dimensions. The mixed-cultural training strategy effectively mitigated cross-cultural performance degradation. This study confirms that aesthetic expression is computable in writing and provides theoretical and technical means of extending AES systems further than surface linguistic features to deep aesthetic quality assessment.