On the Application of Sentence Transformers to Automatic Short Answer Grading in Blended Assessment

In Natural Language Processing, automatic short answer grading remains a necessary launch-pad for the analysis of human responses in a blended learning setting. This study presents pre-trained neural language models that use context dependent Sentence-Transformers to automatically grade student responses with two different input settings. It is found that the use of these models achieves promising results when compared to conventional Bidirectional Encoder Representation Transformer, (BERT), approaches when applying various text similarity-based tasks. This work presents experiments using the benchmark Mohler dataset to test these new models. In summary, an excellent Pearson Correlation score of 0.82 and a Root Mean Square Error of 0.69 is exhibited across a representative experiment sample size.