Extracting Business Process Entities and Relations from Text Using Pre-trained Language Models and In-Context Learning

The extraction of business processes elements from textual documents is a research area which still lacks the ability to scale to the variety of real-world texts. In this paper we investigate the usage of pre-trained language models and in-context learning to address the problem of information extraction from process description documents as a way to exploit the power of deep learning approaches while relying on few annotated data. In particular, we investigate the usage of the native GPT-3 model and few in-context learning customizations that rely on the usage of conceptual definitions and a very limited number of examples for the extraction of typical business process entities and relationships. The experiments we have conducted provide two types of insights. First, the results demonstrate the feasibility of the proposed approach, especially for what concerns the extraction of activity, participant, and the performs relation between a participant and an activity it performs. They also highlight the challenge posed by control flow relations. Second, it provides a first set of lessons learned on how to interact with these kinds of models that can facilitate future investigations on this subject.