Metaphor comprehension has become an important issue of linguistics, cognitive science and computer science. It is also an unavoidable task of natural language processing. This paper presents a novel metaphor comprehension method to make full use of global information based on relevance constraints. The method uses implied perspective to calculate the relevance degree between the target and source domains. First, multi-level semantic representation is obtained based on the semantic representation of word, topic features of word and topic features of discourse. Next, the degree of relevance relations is calculated and the relevance model is generated. Additionally, relevance relations is used to connect cross-level nodes from different perspectives. Then, using random walk algorithm, the relevance relations are acquired from latent perspectives through iterative computations. Finally, the target attribute that has the maximum relevance degree with the target domain is selected as the comprehension result. Experimental results show that the presented method is effective in metaphor comprehension.