A Hierarchical BERT-Based Transfer Learning Approach for Multi-Dimensional Essay Scoring

The task of automated essay scoring (AES) continues to attract interdisciplinary attention due its commercial and educational importance as well related research challenges. Traditional AES approaches rely on handcrafted features, which are time-consuming labor-intensive. Neural network have recently given fantastic results in without feature engineering, but they usually require extensive annotated data. Moreover, most the existing models only report a single holistic score providing diagnostic information about various dimensions writing quality. Focusing these issues, we develop novel approach using multi-task learning (MTL) with fine-tuning Bidirectional Encoder Representations from Transformers (BERT) for multi-dimensional tasks. As state-of-the-art pre-trained language model, BERT-based can improve tasks limited training Meanwhile, deal long texts by proposing hierarchical method mechanism automatically determine contribution different fractions input final score. For multi-topic ASAP dataset, reveal that our outperforms average quadratic weighted Kappa (QWK) 4.5% compared strong baseline. We propose self-collected dataset Chinese EFL Learners’ Argumentation (CELA) provide valuable quality multiple rating dimensions, including five analytic scales. multi-rating dimensional CELA experimental demonstrate model increases QWK 8.1%