Abstract: The goal of legal text matching is to quickly extract and compare element information and discover related cases to ensure the uniformity of legal application. The existing methods can not make full use of prior case knowledge of specific types, extracting a lower accuracy for core elements. Most of the methods can only perform the weight calculation of word vectors, but ignore the vector information of word meaning, sentence meaning and syntax, affecting the matching effect. In this paper, a multi-granularity legal text matching method was proposed to incorporate domain element knowledge. Firstly, a domain knowledge base was established to accurately extract legal elements of specific case types. And then, three granularity attention mechanisms of word, sentence and syntax character were introduced to calculate the weight of different text vectors, so as to improve the effect of matching model. Experimental results show that this method can achieve the best results on public datasets.