Enhancing Dialogue Generation Through Iterative Prompt with Difference Analogy Learning

In-context learning (ICL) is increasingly employed in open-domain dialogue tasks, showcasing remarkable capabilities in areas such as persona consistency, dialogue hallucination, and safety issues. By harnessing ICL's exceptional analogy learning ability, recent methods have achieved significant success in dialogue prediction when interacting with large language models (LLMs). However, due to the diversity of dialogues, it is often challenging for analogy learning to learn appropriate responses from various examples. Moreover, their interactions with large LLMs are typically one-off, making it difficult to fully exploit the potential of LLMs. To tackle these challenges, we introduce the concept of difference analogy learning and propose an iterative prompt method centered around this concept. This method obtains initial prediction results through iterations over examples and then guides the LLMs to learn the gap between predictions and gold response, prompting them to make more appropriate responses in dialogue generation. We validate our method on the widely used public dataset DailyDialog, the results indicate that our approach significantly improves response consistency and diversity.