10th International Congress on Information and Communication Technology

Authors - Koichi Akashi, Hibiki Ito, Atsuko Yamashita, Katsuhiko Murakami, Sayaka Matsumoto, Kunihiko Takamatsu, Tetsuhiro Gozu
Abstract - In recent times, a ground-breaking approach termed abridgement has been advocated for assessing reading comprehension of students in a distinct manner. The operation requires them to shorten the size of the given passage by only deleting the words or characters in it, while not permitting adding, paraphrasing, or swapping any of them. Thanks to its simple operational property, an efficient computational marking scheme has been invented employing the technique of dynamic programming (DP) by the following studies. These endeavors enabled further analyses to proceed, allowing for an ensuing research of reflections about teachings of abridgement collected from students using co-occurrence networks, to give an instance. Furthermore, evolution of this educational approach has led the researchers to suggest its unprecedented application to detect the presence of students who have potential difficulty with learning or specific tasks at an early stage, which is expected to contribute to the reduction of dropout rate in the long run. To make this happen, this research adopted the idea of utilizing large language models (LLMs), specifically Bidirectional Encoder Representations from Transformers (BERT) developed by researchers at Google in 2018, to automatically quantify the confidence levels of their understanding based on the technique of sentiment analysis of collected reflections. Combining grades calculated from DP-based algorithm with LLM newly invited to the research of abridgement, certain criteria have been established to issue warnings to identify those who are experiencing challenges early on.