Dialogue is a useful communication way between human and robot. Dialogue management is important to gather the user's intent, select a task to satisfy the user's request, and make a QA session to perform a specific task, all of which should be performed by natural conversation. We proposed a multi-scenario task-oriented dialogue system based on finite state machine(FSM). Our state machine has common state transitions and they enable users to write scenarios easily and flexibly. FSM based dialogue scenarios are suitable for common state transitions. Multimodal dialogue also requires to care the elderly for dementia diagnosis and daily conversation. We applied the proposed system to diagnoses dementia.
Dialogue systems are a flexible and easy way for humans
and robots to communicate. Task oriented dialogue can
be used to solve specific domain tasks (e.g. hotel
reservation, guides, etc.)[
Recently, the number of patients is increasing due to an aging society. Early detection of dementia is one of important factors in terms of elderly care. Although specialists diagnose dementia, this is a problem that is increasing
Domain-specific tasks rely on pre-defined rules based on ifnite state machines. A task is defined by a XML file defined uniquely. In the state, it is defined that slots requested from users and system’s actions, and cooperation with KB and external services. Figure 1 and 2 show an example of scenario and dialogue in HDS-R.
No Contens Mean Q1 Did you achieve the objectives of the dialogue? 4.06 Q2 Were you able to interact with them in a natural, human way? 3.59 Q3 Did the dialogue go smoothly? 4.06 Q4 Would you like to interact with the robot again? 3.82 Q5 Did the robot perform as expected? 3.59 Q6 Comments STD 0.97 1.28 1.14 1.01 1.23
The system normally chooses the next state when the required slot is satisfied in that state, although we propose special state transition to have flexibility in state transitions.For example, "repeat" transition repeats the same state, "skip" transitions to the next state regardless of whether slots are filled, "cancel" terminates task’s state machine.
The role of input events is to perform slot filling and the role of output events is to generate actions. Figure 3 shows the graph of input modules. Multimodal process is possible by linking each module in a graph.
Two experiments were carried out to evaluate success rate and fluency of dialogue and validity of HDS-R diagnose. First, a survey were conducted to evaluate success rate and fluency of dialogue on 18 people. The survey was rated on 5 scale. Table 1 shows contents and results of survey. The results showed high success rate was achieved while the fluency needs to be improved.
Second, our system diagnosed HDS-R to 15 subjects pretending to be elderly and compared the results of manual and system scoring. Figure 4 shows a radar chart of the results for a subject. The RMSE of the total score was 2 points, confirming that it can be scored as well as a human.
We proposed a multimodal dialogue system based on state machine. Our system enables to write scenarios easily and flexibly because of common state transitions. However, our system is inferior to the end-to-end approach in terms of dialogue fluency. In the future, we also plan to conduct a demonstration test to conrfim the validity of the HDS-R scoring.
This work was supported by JSPS KAKENHI Grant Number JP20H04289 ”Functional Independence Measurement System based on ADL Ontology for Aged Person” URL:https://www.sciencedirect.com/science/ article/pii/S209657961930004X. doi:https://doi. org/10.3724/SP.J.2096-5796.2018.0010. [4] S. Kato, H. Shimogaki, A. Onodera, [development of the revised hasegawa simplified intelligence assessment scale (hds-r)] kaitei hasegawashiki kannichino hyoka suke-ru (hds-r) no sakusei(in japanese), Ronen seishin igaku zasshi(Japanese Psychogeriatric Society) (1991). [5] M. F. Folstein, “mini-mental state”. a practical method for grading the cognitive state of patients for the clinician, J Psychiatr Res (1975).