Acquisition and evaluation of a human-robot elderly spoken dialog corpus for developing computerized cognitive assessment systems

Abstract

Personal communication robots are expected to assist daily living of elderly people. Aiming at developing computerized cognitive assessment systems, we collected human-robot spoken dialog of a cognitive impairment test scenario based on TICS (Telephone Interview for Cognitive Status) and COGNISTAT (Cognitive Status Examination) in Japanese. For the efficient acquisition of the spoken dialog corpus of this scenario, we implemented a WOz (Wizard of Oz) style spoken dialog system on a commercial personal robot, PaPeRo by NEC. By using this system, we collected 147 dialogs spoken by 48 elderlies whose ages varied about 75-85 and MMSE (Mini Mental State Examination) scores varied around 26.5. Each dialog took about 30 min, and contains around 100 human utterances. In order to evaluate feasibility of automatic assessment, we conducted speech recognition experiments with the speech corpus. In the recognition experiments of the elderly speech of answering to the temporal orientation test of asking today's month and date, the accuracies of discriminating correctness of the answers exceeds 80% by using speech dictation engines. This preliminary results show potential feasibilities for computerized cognitive assessment systems.