A cognitive stimulation dialogue system for elders with cognitive impairment

Abstract

In managing communications with elders suffering from cognitive impairment, cognitive stimulation (CS)-based dialogues is able to help to maintain or improve cognitive functions of elders. A significant challenge in developing a Chinese CS dialogue system is the lack of sufficient CS dialogue data in Chinese. To overcome this, we initially build a dataset comprising around two thousand six hundred groups of dialogues according to cognitive stimulation therapy handbook and videos. However, in order to train a higher-quality Chinese CS dialogue model, using more data to train model is essential. Therefore, we propose a knowledge-driven progressive thought prompting method, which uses a knowledge generator to intervene in large language model (LLM) to generate responses with therapy principles and progressive thought generator to ensure that the generated multi-turn dialogue does not generate excessive semantic deviation. Finally, by integrating knowledge and progressive thought through a multi-turn dialogue generator to form prompts, guiding the LLM to generate dialogue, aiming to augment data. After the completion of the dialogue dataset construction, the subsequent step is to build the dialogue model. Chit chat with emotional support not only increases interest in conversation in older adults, it also improves cognitive function. Most existing cognitive dialogue systems neglect to chitchat with emotional support. We also propose a multi-source knowledge fusion method for model to better generate open-ended responses incorporating therapy principles and emotional support strategies. Based on the encoder-decoder structure, we first use a progressive mask method to adapt the encoder to the CS scenario, thereby better classifying utterances into three categories (i.e., therapy principle, emotion and support strategy) and extracting the utterance characteristics. The decoder outputs three categories of utterances classification results and utterances representation through encoder, perceives therapy principle and emotional support strategy in combination with extra paying more attention to emotion and therapy information, to generates reasonable responses. Extensive experiments and analyses demonstrate the exceptional effectiveness of our proposed methods, while there is still a large space for improvement compared to human performance. Subsequent research will further expand the high-quality dataset and increase the parameter volume of the model, introducing more cognitive stimulation principles, with the aim of achieving effects suitable for clinical trials.