Unsupervised Keyword Extraction Methods Based on a Word Graph Network

Abstract

Supervised keyword extraction methods usually require a large human-annotated corpus to train the model. Expensive manual labeling has made unsupervised technology using word graph networks attractive. Traditional word graph networks simply consider the co-occurrence relationship of words or the topological structure of the network, ignoring the influence of semantic relations between words on keyword extraction. To solve these problems, an unsupervised keyword extraction method based on word graph networks for both Chinese and English is proposed. This method uses word embedding to applying a “word attraction score” to semantic relevance between words in a document. Combination of the bias weight of the node and a weighted PageRank algorithm is used to compute the final scores of words. The experimental results demonstrate that the method is more effective than the traditional methods.