在全球老龄化背景下,老年人的健康问题逐渐凸显,认知刺激对话是保持老年人认知健康的重要手段。前人构建了一个结合情感支持的中文认知刺激对话数据集(Chinese Cognitive Stimulation Dialogue Dataset,CSConv),开启了中文认知刺激对话的研究工作。但是没有充分建模认知刺激对话中的逻辑推理关系,生成对话时没能有效利用策略标签的指导作用。本文将认知刺激对话生成视为一个多任务融合的逻辑思维推理过程,将情感分类任务、决策任务和对话回复生成任务间的逻辑关系,建模为一个推理过程,来引导大语言模型生成。针对决策任务,本文提出分层编码器结构的决策模型。决策实验结果表明,决策模型分别将认知刺激治疗原则及情感支持策略决策任务的准确率提高了3.96%和2.1%。针对多任务逻辑思维推理过程,本文提出多任务融合方法,将分类、决策、生成三个任务对应的模型结合在一起。实验结果表明,相比前人方法,多任务融合方法将双语评估替分-4(Bilingual Evaluation Understudy Score based on 4-gram,BLEU-4)提升了7.95%,表明对话回复能力得到提升,证明了该方法的有效性和先进性。
In the context of global aging, the health problems of the elderly have gradually become prominent, and cognitive stimulation dialogue is an important means to maintain the cognitive health of the elderly. Previous researchers constructed a Chinese cognitive stimulation dialogue dataset (CSConv) that includes emotional support, thereby initiating research in the field of Chinese cognitive stimulation dialogue. However, the authors did not fully model the logical reasoning relationships within cognitive stimulation dialogues and did not effectively utilize the guiding role of strategy labels during dialogue generation. This study regards cognitive stimulation dialogue generation as a multi-task integration thinking and reasoning process, and models the logical relationship among emotion classification tasks, decision-making tasks and dialogue response generation tasks as a reasoning process to guide the generation of large language models. For decision-making tasks, this paper proposes a decision-making model with a hierarchical encoder structure. The results of the decision-making experiment show that the decision-making model improves the accuracy of the cognitive stimulation therapy principles and emotional support strategies decision-making tasks by 3.96% and 2.1%, respectively. For multi-task logical thinking and reasoning process, this paper proposes a multi-task integration method to combine the models corresponding to the three tasks. The experimental results show that the multi-task integration method has improved BLEU-4 by 7.95% compared with the previous baseline, indicating that the dialogue response ability has been improved, proving the effectiveness and advancement of this method.