ProcessChat: A Dataset for Business Process Grounded Dialogs

Business processes are designed to streamline and optimize work within an organization and are often defined and documented by domain experts or process analysts using formal business process specifications. However, these specifications may be complex for the users executing the tasks of the process. For example, a recruitment process designed by a domain expert is used by many actors in the organization, who may not be skilled in understanding the formal notations that specify the process. With recent advancements in large language models, there has been increasing interest in enabling users to ask questions in natural language and receive relevant responses that are specific to the user’s context and process knowledge. We propose a dialog dataset grounded in domain-specific process knowledge, which is supposed to be followed during the conversation. The dataset consists of 316 dialogs grounded on 73 different process model specifications. We also present a baseline model, which is trained on the proposed dataset. Our experiments find that the fine-tuned model can do zero-shot transfer to unseen processes, and sets a strong baseline for future research.