Abstract: With the development of the Internet,webpage data,new media text and other data are increasing,the efficiency of information retrieval based on full text is not enough to support the retrieval of massive data,so the keyword extraction technology is widely used in search engines (such as Baidu search) and new media services (such as news retrieval).The fusion model is a model that uses the BiLSTM-CRF structure and fuses multiple manual features,which can more effectively complete the task of keyword extraction.Based on the features of words embedding,the fusion model incorporates the features of part of speech,word frequency,word length and word position.Themultidimensional feature information can help the model to extract deep keyword feature information more comprehensively.The fusion model combines the features of deep learning,such as wide coverage and high learning ability,with the ability of accurate expression of manual features to further improve the feature mining ability and shorten the training time.In addition,a labeling method called LMRSN is adopted in this modelto extract key phrases moreeffec-tively.Experimental results show that the fusion model achieves F1 score of 62.08 in comparison with the traditional model,and its performance is much better than that of the traditional model.

Key words: Deep learning, Feature fusion, Information retrieval, Keyword extraction, Long and short term memory network