Why do we need man-machine interaction in MT?

There have been great gaps between MT research and NLU (Natural Language Understanding) research in AI. MT researchers with few exceptions [1] have generally claimed against exploitation of extra-linguistic knowledge and discourse information in MT, while exploitation of extra-linguistic knowledge and discourse information have been main research objectives in NLU. MT research so far has concentrated on formulating problems in MT as linguistic problems not as 'understanding' problems. Though I admit that a lot of things will have to be done in this framework, it is my contention in this paper that translation requires some extra factors, which we may call 'understanding'. We cannot avoid discussion about possible relationships between 'understanding' and translation, if we take 'high quality MT' seriously.