Keywords Extraction Based on Word2Vec and TextRank

In order to improve the performance of keyword extraction by enhancing the semantic representations of documents, we propose a method of keyword extraction which exploits the document's internal semantic information and the semantic representations of words pre-trained by massive external documents. Firstly, we utilize the deep learning tool Word2Vec to characterize the external document information, and evaluate the similarity between the words by the cosine distance, thus we obtain the semantic information between words in the external documents. Then, the word-to-word similarity is used to replace the probability transfer matrix in the TextRank of word graph of the target document. At the same time, the information of the title and the abstract of the internal document are exploited to construct the words' semantic graph for keyword extraction. The experiments select the related academic paper data from AMiner as experimental data set. The experimental results show that our method outperforms the TextRank algorithm and the precision, recall and F-score of the five keywords are increased by 28.60%, 10.70% and 12.90% respectively compared to the single TextRank algorithm.