Automatic mashup aims to discover desired mashlets according to user goals automatically and combine them into an entirely new application. However, the user goals are usually high-level and coarse-grained while the mashlets are low-level and fine-grained. How to fill in the gap becomes a challenge when addressing automatic mashup development. This paper proposes a novel goal decomposition and refinement approach to handle this problem. We defined a goal model based on which we proposed a history heuristic based algorithm to build a Mashup Goal Ontology repository to enable the auto-decomposition of user goals. Then mashlets which are matching with the refined user goals can be found out and mashed up. We evaluate our approach through experimental results which demonstrate acceptable performance of the decomposition.