Abstract: The shortcoming of the existing automatic extraction algorithm was analyzed, and a new model was proposed, trying to transform unstructured language knowledge into structural language knowledge. The language knowledge was introduced to a extraction model based on typical patterns, and collocations were classed by noun, verb and adjective as center, and by substantive as backbone. Then, concurrence frequency and MI etc were used to screen in large-scale corpus. Finally, this knowledge was solidified to build collocation database.

Key words: Collocation, I}ypical patterns,MI,Collocation database