Flowlogue: A Novel Framework for Synthetic Dialogue Generation With Structured Flow From Text Passages

Dialogue systems play a pivotal role in domains ranging from customer service to virtual assistance and education, using natural language to deliver information and resolve inquiries. Integrating Large Language Models (LLMs) has significantly boosted their capabilities and applications, underscoring their potential to facilitate more nuanced human-computer interactions. Despite these advances, a significant challenge persists in curated dialogue data scarcity, especially in Conversational Question Answering (ConvQA) systems that require domain-specific information. Traditional Passage to Dialogue (P2D) methods attempt to mitigate this by converting textual passages into dialogue form but often need help with issues such as unnatural responses and information redundancy due to the direct use of passage sentences as dialogue answers. To overcome these limitations, we introduce Flowlogue, a novel ConvQA framework that enhances dialogue generation by merging related sentences within passages to maintain natural flow and coherence. This approach leverages LLMs to generate questions and contextually relevant answers based on newly formed dialogue flows, significantly improving the quality and relevance of dialogues compared to existing P2D methods. Our experimental results, validated through reference-free metrics and GPT-4 evaluations, confirm that Flowlogue produces superior dialogues, establishing a robust framework for generating natural, high-quality ConvQA dialogues that effectively harness the depth and nuance of human conversations.