View article

We introduce ml-BERT, an effective machine learning method for automatic short answer grading when training data, ie, graded answers, is limited. Our method combines BERT (Bidirectional Representation of the Transformer), the state-of-the-art model for learning textual data representations, with meta-learning, a training framework that leverages additional data and learning tasks to improve model performance when labeled data is limited. Our intuition is to use meta-learning to help us learn an initialization of the BERT parameters in a specific target subject domain using unlabeled data, thus fully leveraging the limited labeled training data for the grading task. Experiments on a real-world student answer dataset demonstrate the promise of ml-BERT method for automatic short answer grading.