Sarcasm is generally characterized as ironic or satirical that is intended to blame, mock, or amuse in an implied way. Recently, pre-trained language models, such as BERT, have achieved remarkable success in sarcasm detection. However, there are many problems that cannot be solved by using such state-of-the-art models. One problem is attribute infor- mation of entities in sentences. This work investigates the potential of external knowledge about entities in knowledge bases to improve BERT for sarcasm detection. We apply em- bedded knowledge graph from Wikipedia to the task. We generate vector representations from entities of knowledge graph. Then we incorporate them with BERT by a mechanism based on self-attention. Experimental results indicate that our approach improves the accuracy as compared with the BERT model without external knowledge.