Estimating Personal Model Parameters from Utterances in Model-based Reminiscence

TL;DR

The paper tackles the challenge of personalizing reminiscence therapy by estimating a user’s internal state through a model-based reminiscence framework built on ACT-R. It introduces an interactive estimation scheme where the model’s activation and utility parameters are updated based on user utterances during lifelog recalls, enabling inference of both the user’s mood and the model’s internal state. Empirical results show partial success in classifying the model’s internal states from responses and stronger performance for post-hoc mood classifications, with real-time mood classification being more challenging and variable across individuals. The work demonstrates the feasibility of utterance-driven human-model interaction to tailor reminiscence support and highlights avenues for integrating physiological data and real-time feedback to enhance personalization and well-being outcomes.

Abstract

Reminiscence therapy is mental health care based on the recollection of memories. However, the effectiveness of this method varies amongst individuals. To solve this problem, it is necessary to provide more personalized support; therefore, this study utilized a computational model of personal memory recollection based on a cognitive architecture adaptive control of thought-rational (ACT-R). An ACT-R memory model reflecting the state of users is expected to facilitate personal recollection. In this study, we proposed a method for estimating the internal states of users through repeated interactions with the memory model. The model, which contains the lifelog of the user, presents a memory item (stimulus) to the user, and receives the response of the user to the stimulus, based on which it adjusts the internal parameters of the model. Through the repetition of these processes, the parameters of the model will reflect the internal states of the user. To confirm the feasibility of the proposed method, we analyzed utterances of users when using a system that incorporates this model. The results confirmed the ability of the method to estimate the memory retrieval parameters of the model from the utterances of the user. In addition, the ability of the method to estimate changes in the mood of the user caused by using the system was confirmed. These results support the feasibility of the interactive method for estimating human internal states, which will eventually contribute to the ability to induce memory recall and emotions for our well-being.

Figures (7)

• Figure 1: Scheme of the Human-Model interaction. The model provides a stimulus to a user. The user receives a stimulus and responds to it. The model updates its internal parameters based on the received response and determines a new stimulus.
• Figure 2: Example of photos network. The number A given to each photo means the activation. The photos are connected to each other via a network of attributes. Each attribute has a utility $U_i$ that is assigned to the corresponding production rule.
• Figure 3: Interface for mood state rating. Users rate their mood state using the interface in the red square.
• Figure 4:
• Figure 6: Classification results of the experimental conditions. (A: Activation, R: Reward)