Comparison of Questioning Methods to Elicit QOL Expression Speech of Elderly Subjects
Takumi Kudo (AY 2023)
With the aging of society, the isolation of the elderly from society has become a serious problem. In particular, the problem of lonely death of the elderly living alone has been pointed out, and the background of this problem is that it is difficult to communicate their health and living conditions to their families and environment. In order to solve these problems, many studies on the quality of life (QOL) of the elderly have been conducted. In previous studies, it was found that in communication between family members and the elderly, the elderly's QOL can be communicated to family members by supporting the generation of utterances in which the listener can infer the elderly's QOL from the elderly's utterances (hereafter: QOL-expressed utterances). However, most of the behaviors and conditions described in the utterances were family-centered, and very few of them were elder-centered. Therefore, in this study, we constructed an application equipped with ChatGPT that enables natural dialogue and investigated whether it would be possible to elicit QOL-expressed utterances from elderly people by increasing elderly-oriented utterances using smart speakers. In addition, for future practical application, we felt it necessary to adapt the set of questions used to individual users. Therefore, we devised a "scoring method (proposed method) that takes into account the fact that a group of questions with high scores should be retained in order to adapt to individual differences" and investigated whether it would be possible to efficiently elicit QOL-expressed utterances.
In this study, smart speakers were installed in the homes of older adults, and participants kept a daily diary by answering three questions selected from five groups of questions (diet, exercise, shopping, hobbies, and conversations with others) created by the experimenter through the device. The smart speaker asks a question as a conversation starter, and the participant responds. Based on the utterances, ChatGPT was constructed to ask another set of questions. The experimenter then scores the collected conversation data and determines three groups of questions to be used the next day according to an algorithm. The algorithms used were "suggested method," "random sampling," and "fixed order," and each method was rotated every five sessions. The experiment was conducted on four elderly people (two males and two females) living in Ibaraki Prefecture for a period of three months.
The results of the experiment showed that QOL-expressed utterances were elicited in all question groups. In addition, a comparison of the questioning methods revealed that the "suggestion method" was more effective in eliciting QOL information than the "fixed order method". However, no effect was observed between the "suggestion method" and "random sampling" in terms of the difference in the method of selecting question groups. From these results, it was found that it is better to vary the order of question groups, but it was not clear from this study whether the proposed method is better or randomization is sufficient in terms of method of variation. This study verified that ChatGPT can be used to derive QOL-expressed utterances by a dialog system, which has been difficult to achieve in the past. In the future, we would like to increase the number of ChatGPT tokens and adjust the timeout period to achieve a dialogue closer to that of humans, and confirm the effect on the degree of QOL expression of the elderly.
(Translated by DeepL)