Abstract: In order to better solve the problems of inaccurate semantic information and out of vocabulary problems in existing text summarization models, a text summarization model based on keyword semantic feature enhancement is proposed. Keybert keyword extractor was used, and multi-head self-attention mechanism was added to enhance the extraction of keyword semantic information, so that the model had better context information fusion ability and key information expression ability. In order to solve the common problems of out of vocabulary words and exposure bias in text summarization, based on pointer generation network, we used a hybrid training strategy based on reinforcement learning, efficiently extracting the generated abstracts from the created vocabulary and the original text. Compared with existing comparison algorithms, experimental results and examples show that the proposed model can effectively improve the accuracy and readability of summary generation on NLPCC2017 dataset.