In data-oriented English-Chinese machine translation, knowledge source is the very important basis for translation processing. This paper presents a kind of construction strategy for knowledge source which contains affluent grammatical and syntactical information. Firstly, taking lexical function grammar as the theoretical basis, treebank including parse trees converted from every sentence in the source language corpus is acquired. Secondly, based on the decomposition algorithm, the corresponding fragment-bank composed of all the legal fragments extracted from the treebank is constructed. Finally, based on the combination algorithm, the fragment-combination-bank including all the possible fragment-combination forms of every parse tree in the treebank is built. Based on the successful construction of the knowledge source, the whole machine translation process can be implemented efficiently and accurately.