Training an ML model to "learn" a rules engine strikes me as an incredibly bad practice. It'd make more sense to just have an actual corpus of labeled data.