| Keyphrase can be efficiently used to retrieve and understand text content,enabling users to catch the main idea of text quickly.The quality of keyphrase extracted by current keyphrase extraction technology is still not good enough,which further influence subsequent tasks.In order to improve the accuracy of keyphrase extraction,this thesis focuses on the study of keyphrase extraction methods from two aspects: graph method and neural network method.Most keyphrase extraction methods only use few features.This thesis proposes an improved keyphrase extraction method for multi-feature fusion of TextRank.Our opinion is the word appears in more sentences,the word is more important.Therefore,we calculate the total number of sentences in which the target word appears and integrate it into TextRank as a feature.Furthermore,we use word span,word position reciprocal sum to modify TextRank transition probability and use the LDA topic information to modify the reset probability.We conduct experiments on three data sets SemEval2010,KDD and WWW.Experimental results show that our algorithm has significantly improved precision,recall,and F1-score over baseline algorithms TF-TDF,TextRank,SingleRank,and TopicRank.The previous keyphrase extraction method is based on graph model only.This thesis further combines graph model with neural network method and proposes a keyphrase extraction method based on BiGRU and PositionRank.We use BiGRU network to learn forward and backward context information of words in a document and use PositionRank algorithm to learn positional feature and then combine these two parts.Eventually,the word with the highest score will be selected as final keyphrase.We perform experiments on three datasets SemEval2010,KDD,and WWW.Experimental results show that the quality of keyphrase extraction based on BiGRU and PositionRank is higher than that of baseline methods.And the combination of graph model and neural network shows better result than graph-only method. |
PDF Full Text Request