A System for Generating Student Progress Reports in Cram School

Abstract

In many cram schools, instructors write reports on students’ progress after each class. The generation of these reports is a heavy burden for instructors, and there is a need to reduce this burden. Therefore, in this paper, we propose a system that automatically generates a student learning status reports. Students’ learning status is often evaluated from several specific items, and rule-based sentence generation can be considered for those items. However, since viewpoints other than the specific items are often incorporated into the report document, a keyword-based sentence generation function is required to incorporate expressions that are difficult to be generated by the rule-base methods. Here we consider two keyword-based methods: the Sequence-to-Sequence-based method, which learns the correspondence between keywords and sentences, and the Information Retrieval-based method, which directly retrieves and reuses past reports. In this paper, we compare and evaluate the two methods and implement the model with better performance into our report generation system. We evaluated the two methods based on actual data of about 200,000 reports written by instructors, and confirmed that the Seq2Seq-based model with Attention had the best performance, and was able to generate more accurate sentences by learning positive and negative expressions separately.