Enhance Robustness of Language Models Against Variation Attack through Graph Integration