Current studies on a natural language interlace rely on train data and syntactic parsing for an input sentence. These dependencies cause the discordance between expressions covered by a system and meaning representations of the given domain. This paper proposes a method to generate at least one expression corresponding to a meaningful meaning representation. The proposed method chooses the most appropriate meaning representation by using the generated expressions. The experiment results on Question answering for U.S. geography ontology showed that the proposed method outperforms methods of previous studies.