Основные понятия
Developing a suggestion question generator using dynamic contexts to enhance user interaction with conversational systems.
Аннотация
When users interact with Retrieval-Augmented Generation (RAG) based conversational agents, they often struggle to craft queries accurately, leading to ambiguous questions that require clarification. This work focuses on developing a suggestion question generator that utilizes dynamic contexts, including dynamic few-shot examples and retrieved contexts. By experimenting with this approach, the study demonstrates that dynamic contexts can generate better suggestion questions compared to other methods. The system aims to alleviate users from the task of formulating questions, ensuring a smoother conversational flow. Dynamic Contexts approach is designed to improve user understanding of conversational systems' capabilities by providing more informed suggestions and enhancing the overall user experience.
Статистика
ChatGPT: 44 correct samples
Claude-2: 44 correct samples
GPT-4: 46 correct samples
Zero-Shot method by Claude-2: 30 correct questions out of 48 samples
Few-Shot method showed improvement in performance but tended to replicate example structures closely
Dynamic Few-Shot approach aimed at introducing variety in question types and structures but lagged behind dynamic contexts approach