In this paper, we proposed an example-based approach aiming at recovering ill-formed inputs to improve robustness of spoken dialogue systems. In this approach, a treebank, which contains example sentences and their correct parse trees, is used to provide clues for fixing the errors of ill-formed inputs. Particularly, the proposed error recovery method is suitable for spoken dialogue application because of computationally efficiency. In addition, when evaluated in a Mandarine spoken dialogue system, the proposed method has shown to improve the system's understaning rate very significantly.