This paper combines Wmatrix with BNCweb to look at the grammar of linguistic metaphors formed by concrete nouns and their derivatives across word classes in order to shed light on learning of polysemous words. The methodology is to search different word classes of such a noun in BNCweb, use the concordances to construct corpus data and upload the data to Wmatrix for semantic domain tagging, choose the self-contained corpus in Wmatrix — British English 2006 — for a reference corpus to do the frequency comparison, then study the resultant key semantic tags as clues to identify the source domains of metaphor. The underlying idea of Cognitive Metaphor Theory, the Great Chain of Being and the semantic makeup of a concept will help with the identification of a metaphor. The case study on duck shows the detailed process and feasibility of the research method. The research results reveal how different meanings of a concrete noun are related by metaphor or metonymy, which facilitates the learning of polysemous words.