To address the problem of the performance of embedding-based keyword extraction methods on long documents deterio-rates,a summarization-based keyword extraction approach was proposed,denoted by summarization-based document embedding rank(SDERank).The document embedding was taken as the weighted sum of sentence vectors,and weights were assigned to each sentence according to its semantic relevance to the document topic.Existing embedding-based methods fail to take into account the relation between candidate words.For this problem,in SDERank+,an improved version of SDERank,the co-occurrence weight of candidate words was calculated to amend the original similarity score by PageRank.Experimental results demonstrate that SDERank and SDERank+achieve 2.2％and 3.29％higher F1 scores respectively than that of the current best,MDERank,on four widely used datasets.
维普
万方数据