Leveraging Graph Structures and Large Language Models for End-to-End Synthetic Task-Oriented Dialogues
Maya Medjad, Hugo Imbert, Bruno Yun, Raphaël Szymocha, Frédéric Armetta
TL;DR
GraphTOD tackles the data bottleneck in task-oriented dialogue systems by offering a two-agent, end-to-end LLM pipeline guided by a JSON-defined action transition graph $G=(V, Ac, E, t, s, f)$. The framework formalizes the dialogue flow with an action graph, a set of function-call APIs, and a suite of prompt templates to orchestrate system and user utterances. Key contributions include the JSON-based graph specification, automatic user preference generation, and an end-to-end prompting scheme that yields domain-adaptive synthetic dialogues across multiple domains. In experiments, 150 dialogues across four domains were generated and evaluated with UniEval, showing strong naturalness and coherence and approaching human-in-the-loop baselines like LAPS. GraphTOD thus lowers the cost and complexity of TOD dataset creation, broadening accessibility for non-experts to produce scalable synthetic data.
Abstract
Training task-oriented dialogue systems is both costly and time-consuming, due to the need for high-quality datasets encompassing diverse intents. Traditional methods depend on extensive human annotation, while recent advancements leverage large language models (LLMs) to generate synthetic data. However, these approaches often require custom prompts or code, limiting accessibility for non-technical users. We introduce GraphTOD, an end-to-end framework that simplifies the generation of task-oriented dialogues. Users can create dialogues by specifying transition graphs in JSON format. Our evaluation demonstrates that GraphTOD generates high-quality dialogues across various domains, significantly lowering the cost and complexity of dataset creation.