Utterance Domain Classification (UDC) is essential for Spoken Language Understanding (SLU), a task analogous to short text classification. Short texts are often challenging to understand due to their lack of context, necessitating the enrichment of their semantic representation with supplementary information such as concepts from external knowledge bases. However, the inclusion of concepts introduces noise, making the selection of valuable concepts challenging. This paper proposes a UDC method employing keyword-guided signals to enhance the purity of external knowledge. We use two keyword extraction strategies to construct two types of keywords. A keyword-assisted concept denoising module addresses the concept noise problem, and a knowledge injection module is designed to better integrate concepts into the model. Experimental results on two Chinese SLU datasets demonstrate that our model achieves state-of-the-art performance.