Abstract: Data records of simi-structure Web page are similar in structure.This virtue represents as repeat tag strings in the tag sequence of first order traversing DOM tree，which generally can be mined through constructing a suffix-tree.Since the tag sequence can be generated both in block tag level and text tag level，and the different granularity patterns’ performances play an uncertain way，a semi-structure information extraction algorithm with automatic granularity selection is introduced in this paper.Firstly，it generates two different granularity candidate pattern collections by search maximal repeats and tandem repeats in respective suffix-trees，and then evaluates these patterns by statistic metrics.A new metric of extraction result regularity and a weighted approach are introduced for selecting the target pattern.