This is an excellent example where phrase-based statistical machine translation is troublesome; the language model score of "his wife" is going to be dramatically larger "her wife" in any world where there are a lot more hereto married people. Neural MT should have a much better shot at getting it.