This paper proposes a graph-based projection approach to improve the robustness of cross-lingual spoken language understanding (SLU) when applied to a new language. The approach constructs graphs using trigrams in the dataset connected by word alignments and semantic similarities. It then performs label propagation to induce labels for unlabeled nodes. An evaluation on English-Korean SLU tasks shows the graph-based projection approach improves over direct projection and training only on the source language data, achieving higher precision, recall and accuracy for named entity recognition and dialog act identification.