Why not just build a big platform contains all the data, then write kernels for each type of 'violation', learn a pattern, find suspects, then use results to improve the kernel. Human labor + Machine intelligence, Yay! Eventually you'll have all kinds of tax law text covered, and patterns of violations.