Improving Unsupervised Keyphrase Extraction using Background Knowledge

WikiRank:Improving Keyphrase Extraction Based on Background Knowledge Yang Yu, Vincent Ng Human Language Technology Research Institute University of Texas at Dallas {yangyu,vince}@hlt.utdallas.edu Abstract Keyphrase is an efficient representation of the main idea of documents. While background knowledge can provide valuable information about documents, they are rarely incorporated in keyphrase extraction methods. In this paper, we propose WikiRank, an unsupervised method for keyphrase extraction based on the background knowledge from Wikipedia. Firstly, we construct a semantic graph for the document. Then we transform the keyphrase extraction problem into an optimization problem on the graph. Finally, we get the optimal keyphrase set to be the output. Our method obtains improvements over other state-of-art models by more than 2% in F1-score. Keywords: Keyphrase Extraction, Knowledge Graph, Semantic Graph 1. Introduction As the amount of published material rapidly increases, the problem of managing information becomes more difficult. Keyphrase, as a concise representation of the main idea of the text, facilitates the management, categorization, and retrieval of information. Automatic keyphrase extraction concerns “the automatic selection of important and topical phrases from the body of a document”. Its goal is to extract a set of phrases that are related to the main topics discussed in a given document(Hasan and Ng, 2014). Existing methods of keyphrase extraction could be divided into two categories: supervised and unsupervised. While supervised approaches require human labeling, at the same time needs various kinds of training data to get better gen- eralization performance, more and more researchers focus on unsupervised methods. Traditional methods of unsupervised keyphrase extraction mostly focus on getting information of document from word frequency and document structure(Hasan and Ng, 2014), however, after years of attempting, the performance seems very hard to be improved any more. Based on this observation, it is reasonable to suspect that the docu- ment itself possibly cannot provide enough information for keyphrase extraction task. To get good coverage of the main topics of the document, Topical PageRank (Liu et al., 2010) started to adopt topical information in automatic keyphrase extraction. The main idea of Topical PageRank is to extract the top topics of the document using LDA, then sum over the scores of a can- didate phrase under each topic to be the final score. The main problems with Topical PageRank are: First, The top- ics are too general. Second, since they are using LDA, they only classify the words to several topics, but don’t know what the topics exactly are. However, the topical informa- tion we need for keyphrase extraction should be precise. As shown in Figure 1, the difference between a correct keyphrase sheep disease and an incorrect keyphrase incur- able disease could be small, which is hard to be captured by rough topical categorization approach. To overcome the limitations of aforementioned ap- proaches, we propose WikiRank, an unsupervised auto- matic keyphrase extraction approach that links semantic meaning to text The key contribution of this paper could be summarized as follows: 1. We leverage the topical information in knowledge bases to improve the …

Content truncated to 500 words