Abstract: Aiming at the problems of missing key information and difficult control of content in the original text during the generation process of existing generative summary models， we proposed a generative text summarization method guided by extraction methods. This method first obtained key sentences from the original text through an extraction model, and then adopted dual encoding strategy to encode key sentences and news text respectively, so that key information was guided to generate a summary during the decoding process. Finally, expert network was introduced to screen information during decoding to further guide the  generation of summary. The experimental results on CNN/Daily Mail and XSum datasets show that the proposed model can effectively improve the performance of abstractive text summarization. This method improves the content of key information in the original text for generating summary to a certain extent, while alleviating the problem of  difficult  control of generated content.

Key words: abstractive text summarization, double encoder, key information, expert network, guided perception