基于图嵌入的关键词抽取方法OA

Keyword extraction is one of hot research topics in the field of text mining,and its accuracy has significant impact on downstream tasks.To enhance the accuracy of keyword extraction for test,this paper proposes a graph embedding-based keyword extraction(GeKWE)approach.A candidate word graph is built for a preprocessed document,where nodes denote candidate keywords while edges denote the strength of co-occurrence relationship between candidate words.On this basis,drawing on the embedding method of the Word2Vec model,this paper uses a random walk strategy to obtain sampling nodes sequences from the graph and then trains a skip-gram model to learn representations of candidate words graph,obtaining the semantic vectors of candidate words.Further,candidate words are divided into several clusters using the embedding representation.For each cluster,candidate words with a high score are chosen as the final keywords.Experiments on two benchmark datasets,Inspec and SemEval2010,are conducted and comparisons with six state-of-the-art approaches are made.Experimental results show that the proposed approach achieves the highest F1 score on both datasets,which are 61.26%and 42.73%respectively,better than the selected comparison approaches,and is able to effectively alleviate information overload.