Translation memory (TM) is one of the cruxes for CAT systems. With the help of TM,translators could save time and reduce cost in translation. Right now,there are kinds of TMs and the best-known TMs are:Deja vu,SDL Trados,and WordFast. When you really use one TM,you will find that similarity is the linchpin to influence TM's outputs. Per Bowker's introduction,there are several elements to influence similarity. The paper will analyze the main elements to influence similarity and try to find a way to improve the similarity level.