A controlled natural language for tax fraud detection

Abstract

Addressing tax fraud has been taken increasingly seriously, but most attempts to uncover it involve the use of human fraud experts to identify and audit suspicious cases. To identify such cases, they come up with patterns which an IT team then implements to extract matching instances. The process, starting from the communication of the patterns to the developers, the debugging of the implemented code, and the refining of the rules, results in a lengthy and error-prone iterative methodology. In this paper, we present a framework where the fraud expert is empowered to independently design tax fraud patterns through a controlled natural language implemented in GF, enabling immediate feedback reported back to the fraud expert. This allows multiple refinements of the rules until optimised, all within a timely manner. The approach has been evaluated by a number of fraud experts working with the Maltese Inland Revenue Department.