A study of Automatic scoring model for essay-type questions using a small dataset expansion method with The Back Translation

Abstract

With the introduction of essay-type questions in The Common Test for University Admissions in Japan being considered, there is a growing demand for essay-type questions to test students' thinking ability. Although there have been many previous studies on the use of natural language processing to construct an automatic scoring model to reduce the cost of scoring, the feasibility of such a model in actual scoring situations is low because it requires a large data set to construct a highly accurate model. In this study, we expand the dataset by using The Back Translation method on the English dataset obtained by translating the dataset of answers written in Japanese into English. Then, we fine-tune the English pre-training model of BERT using the dataset obtained in this way, aiming to output a distributed representation that reflects the meaning of the answers. Finally, by building a classifier that outputs scores based on the distributed representations, we consider building a highly accurate automatic scoring model for essay-type questions from a small dataset.