Transfer based Machine Translation System normally needs a large quentity of transfer patterns, not only in pattern-based , but also in example and rule- based transfer module. Some pattern-based analysis modules need patterns too.
There are many different styles in pattern. Translation examples maybe the most concrete style in patterns, and transfer rules maybe the most abstract one. The more abstract, the more flexible, and the more ambiguous too. In the other hand, the more concrete, the more limited in its application.
This thesis porposes a multi-stage transfer pattern to try to over come the problem of pattern style. The multi-stage transfer pattern contains several stages in it, and patterns in different concrete belong to different stages. In this thesis, we presents an approach to construct more flexible and more concrete transfer patterns by pattern clustering. Concrete word or phrase can be and can only be remained when the target pattern will be changed if the unit is abstracted or exchanged to other symbol or unit.
A linguistic knowledge based word and phrase level alignment system is developed for the transfer pattern constraction.