We don't need to run HERD on the test data. We could just use a dependecy parser to find the object and subject in a sentence.