Reducing correspondence ambiguity in loosely labeled training data