Abstract: Encoder-decoder-based generative approaches have been widely used and achieved good performance for keyphrase extraction tasks. However, the main challenges of the encoder-decoder-based approach are modeling an effective document vector representation and generating a set of keyphrases covering the entire document topic, which can directly affect the keyphrase extraction results. In this paper, a document-level keyphrase extraction model incorporating neighborhood knowledge is proposed to address the challenges mentioned above simultaneously. Specifically, the original document is extending to a document set by adding some nearest-neighbor documents. Then, each document in the set is constructed into a word graph based on the distance between words, and all the word graphs in the set are merged into a large graph, which is then encoded using graph convolutional networks. Besides, to fully cover diverse keyphrases and topics, the context modification mechanism and coverage mechanism are introduced at the decoding step. Finally, by comparing with the existing baseline model on four benchmark datasets, the experimental results show that the method can effectively improve the performance of extracting keyphrases.