Abstract: Sememe is the core component of concept description in HowNet, and the predication of sememe description for new concepts is the key issue involved in automatic or semi-automatic expansion of HowNet. This paper proposes a sememe prediction method based on network embedding and the pre-training models. It realized the dynamic matching between the new concept and the candidate sememe by learning representation of the character-word-concept-sememe and their relationships in HowNet, and combining the pre-training language models to construct the partial "concept-sememe" relationship network. The predicted F1 value of the experimental results was 0.6237, which indicated that this method could solve the problem of semantic prediction of OOV words in HowNet more effectively.