MDSD: Multi-turn Diverse Synthetic Dialog Generation for Domain Specific Incomplete Requests Understanding

Handling user natural language instructions and implementing system requests for specific tasks is crucial for large language models (LLMs). However, users often struggle to provide complete and accurate instructions, leading to potential challenges in task execution. To address this issue, we propose a Multi-turn Diverse Synthetic Dialog Generation (MDSD) method aimed at generating large-scale, diverse dialogue datasets spanning various task types, request modes, and guiding logics. By leveraging these datasets, we train an LLM to effectively translate natural language requests into precise system inputs, even in scenarios where users provide incomplete or ambiguous information. To further enhance user interaction, we design an agent-based workflow that actively guides users towards formulating clear and precise requests through a dynamic, chat-style interface. This approach not only improves the model's ability to comprehend and fulfill user intentions but also empowers users to iteratively refine their instructions. Evaluation across different test scenarios demonstrates the effectiveness of the MDSD method, achieving a system request success rate exceeding 98%. These results underscore the potential of our approach in improving the accuracy and reliability of LLMs for real-world applications.