Abstarct: The keyphrase indicates the basic concepts for a text. In the case of keyphrase extraction, the goal is to design a system that can automatically extract keyphrases by receiving input text. In conventional methods, modeling for keyphrases was modeled baxsed on the words appearing in the document. If keyphrases do not appear, they form a significant part of a document's keywords and are not adequately identified. Another problem is the production of keyphrases in terms of the arrangement and composition of the final words, which are not readable and of high quality. In general, the three proposed methods are presented in this thesis with an innovative approach baxsed on particular attention to modeling the phrases that do not appear, and the final product is more appropriate. In the first method, a model is presented to extract the basic concepts of the text by estimating similar texts and adding unseen keyphrases to the hidden laxyers of the deep network. According to the research in this study, ubiquitous documents have an excellent commonality in keywords that do not appear directly. In the second method, an attempt is made to improve the network structure in terms of different laxyers baxsed on the attention mechanism. In addition to improving the memory of the sequences and the quality of the final results, these changes make the method more susceptible to other applications, including text summarization. Finally, the third proposed method creates an exceptional quality, especially in producing the sequence of final exxpressions, using the combined approach of reciprocal generating network with reinforcement learning optimization. All methods have been tested on four common data in this field. In addition to solving the challenges, these experiments show that the third method (GANUnseen) shows an average improvement of 4.2% compared to the CorrRNN method in extracting keyphrases.