There must be as many concrete indicators as possible in education, which will become signposts. People will not be confident about their learning and will become confused with tenuous instruction. It is necessary to clarify what they can do and what kinds of abilities they can improve. This paper describes a case of evidence-based education that acquires educational data from students' study activities and not only uses the data to enable instructors to check the students' levels of understanding but also improve their levels of performance. Our previous research called discussion mining was specifically used to collect various data on meetings (statements and their relationships, presentation materials such as slides, audio and video, and participants' evaluations of statements). This paper focuses on student presentations and discussions in laboratory seminars that are closely related to their research activities in writing their theses. We propose a system that supports tasks to be achieved in research activities and a machine-learning method to make the system sustainable for long-term operation by automatically extracting essential tasks. We conducted participant-based experiments that involved students and computer-simulation-based experiments to evaluate how efficiently our proposed machine-learning method updated the task extraction model. We confirmed from the participant-based experiments that informing responsible students of tasks that were automatically extracted on the system we developed improved their awareness of the tasks. Here, we also explain improvements in extraction accuracy and reductions in labeling costs with our method and how we confirmed its effectiveness through computer simulations.