Social Network Mining from Natural Language Text and Event Logs for Compliance Deviation Detection

Abstract

Social network mining aims at discovering and visualizing information exchange of resources and relations of resources among each other. For this, most existing approaches consider event logs as input data and therefore only depict how work was performed (as-is) and neglect information on how work should be performed (to-be), i.e., whether or not the actual execution is in compliance with the execution specified by the company or law. To bridge this gap, the presented approach considers event logs and natural language texts as input outlining rules on how resources are supposed to work together and which information may be exchanged between them. For pre-processing the natural language texts the large language model GPT-4 is utilized and its output is fed into a customized organizational mining component which delivers the to-be organizational perspective. In addition, we integrate well-known process discovery techniques from event logs to gather the as-is perspective. A comparison in the form of a graphical representation of both, the to-be and as-is perspectives, enables users to detect deviating behavior. The approach is evaluated based on a set of well-established process descriptions as well as synthetic and real-world event logs.