Yoko Nishihara, Seiya Tsuji, Wataru Sunayama, Ryosuke Yamanishi, Shino Imashiro: A generation Method for Discussion Process Model during Research Progress Using Transitions of Dialogue Acts, Journal On Advances in Systems and Measurements, 14(1&2), pp.17-26, 2022年1月

People in business and academic fields work in cooperation rather than alone. They may discuss their progress with others, like co-workers and supervisors, to help them obtain the best results, and sometimes, they may feel that such discussions are not conducted well. However, people do not evaluate the quality of each discussion on every occasion because it is tough work for them, and they usually do not have enough time for that. In the process of evaluating discussions, people might look back on their discussions and make a plan to have an improved discussion next time. This paper proposes a generation method for a discussion process model during research progress. First, discussions are recorded to generate transcripts in which each line has a speaker name and his/her utterance. Then, the transcripts are classified manually into high-quality and low-quality discussion groups. Next, dialog acts are assigned to utterances as labels. The labels of dialog acts are originally designed for discussion analysis for research progress. After the labeling, transitions of dialog acts with a high appearance rate are extracted. The transitions are connected if the same dialog act is in both transitions to make a network of dialog acts. The network represents a model of the discussion process. A model for a high-quality discussion group is compared with a model for a low-quality discussion group. By investigating dialog acts and transitions found only in one group, suggestions for low-quality discussions to high-quality discussions would be found. We used discussions between a supervisor and a student who was studying for a degree at a university. We generated models for high-quality and low-quality discussion groups by the proposed method and revealed suggestions for low-quality discussions to high-quality discussions. Our contributions are summarized in two points: (1) we proposed a new method to generate models that represent the discussion process, and (2) we found suggestions for low-quality discussions to high-quality discussions using the models obtained.