Scholarlink

<p>In building smart and adaptive dialog systems, natural language processing research is concerned with complications involving multiple-turn conversations in which context maintenance, diversity in the response generation, and knowledge integration matter the most. In this article, we present EMRO-Dial as a new dialog system that incorporates Reinforcement Learning (RL) with the enhanced version of the Rime Optimizer (EMRO) to improve policy search, quicken the convergence process, and encourage diversity in responses. The population-based evolutionary strategy employed within EMRO helps the system bypass some inherent drawbacks of RL methods, such as rare rewards and being sucked into local optima. Experiments were conducted extensively on the MultiWOZ and SQuAD datasets, with evaluation metrics based on the task success rate, BLEU score, novelty, and knowledge accuracy. The results show that EMRO-Dial is superior to rule-based, supervised, and RL-enhanced models. The parameter sensitivity analysis also gives good configurations for stable and efficient learning. The results show that injecting EMRO into RL-based dialog systems largely improved fluency and factual grounding and thus represents a promising avenue for innovative conversational AI applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reinforcement learning-driven multi-turn dialogue systems enhanced by modified rime optimizer

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Haizhi Wei

摘要

In building smart and adaptive dialog systems, natural language processing research is concerned with complications involving multiple-turn conversations in which context maintenance, diversity in the response generation, and knowledge integration matter the most. In this article, we present EMRO-Dial as a new dialog system that incorporates Reinforcement Learning (RL) with the enhanced version of the Rime Optimizer (EMRO) to improve policy search, quicken the convergence process, and encourage diversity in responses. The population-based evolutionary strategy employed within EMRO helps the system bypass some inherent drawbacks of RL methods, such as rare rewards and being sucked into local optima. Experiments were conducted extensively on the MultiWOZ and SQuAD datasets, with evaluation metrics based on the task success rate, BLEU score, novelty, and knowledge accuracy. The results show that EMRO-Dial is superior to rule-based, supervised, and RL-enhanced models. The parameter sensitivity analysis also gives good configurations for stable and efficient learning. The results show that injecting EMRO into RL-based dialog systems largely improved fluency and factual grounding and thus represents a promising avenue for innovative conversational AI applications.