Chinese short text matching is a fundamental task in natural language\nprocessing. Existing approaches usually take Chinese characters or words as\ninput tokens. They have two limitations: 1) Some Chinese words are polysemous,\nand semantic information is not fully utilized. 2) Some models suffer potential\nissues caused by word segmentation. Here we introduce HowNet as an external\nknowledge base and propose a Linguistic knowledge Enhanced graph Transformer\n(LET) to deal with word ambiguity. Additionally, we adopt the word lattice\ngraph as input to maintain multi-granularity information. Our model is also\ncomplementary to pre-trained language models. Experimental results on two\nChinese datasets show that our models outperform various typical text matching\napproaches. Ablation study also indicates that both semantic information and\nmulti-granularity information are important for text matching modeling.\n