计算机科学
推论
言语行为
自然语言处理
语音识别
语言模型
插件
间接引语
人工智能
语言学
程序设计语言
哲学
作者
Jeremy Chang,Chung‐Hsien Wu
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-01-01
卷期号:33: 2661-2673
被引量:5
标识
DOI:10.1109/taslpro.2025.3578773
摘要
Developing an empathetic dialogue system has been a challenge for years. Although most system responses are grammatically correct and convey emotions to some extent, their quality still falls short of human standards. In this study, an integrated system called the Plug-and-Play Rational Speech Act (PPRSA) was developed by integrating a Transformer-based language model with two frameworks, namely a Plug-and-Play (PP) framework and a rational speech act (RSA) framework, to generate empathetic and engaging responses. The PP framework was used to control the generation through two attribute models, which ensure the empathetic intent and engagement of responses, thereby increasing empathy and relevance in the generated responses. The RSA framework was used to accelerate and improve response generation by reasoning the perturbation variation from the attribute models. It reduced the response generation time by 73.65% and exhibited enhanced performance compared to the Plug-and-Play language model (PPLM). EmpatheticDialogues was used as the evaluation dataset. In terms of objective evaluation, the PPRSA-based system accurately provided empathetic intent and engaging responses based on users' emotions, thereby providing more comfort and support. The system achieved a BERTScore of 0.869, a Distinct-1 score of 5.19, a Distinct-2 score of 31.88, an empathetic intent accuracy of 41.84%, and an engagement accuracy of 72.24%. In terms of human evaluation, the system outperformed other baseline systems in terms of empathy, relevance, and fluency.
科研通智能强力驱动
Strongly Powered by AbleSci AI