WordTopic-MultiRank: A New Method for Automatic Keyphrase Extraction

International Joint Conference on Natural Language Processing, pages 10–18, Nagoya, Japan, 14-18 October 2013. WordTopic-MultiRank : A New Method for Automatic Keyphrase Extraction Fan Zhang† Lian’en Huang† Bo Peng‡ †The Shenzhen Key Lab for Cloud Computing Technology and Applications Peking University Shenzhen Graduate School, Shenzhen 518055, P.R.China fan.zhgf@gmail.com, hle@net.pku.edu.cn ‡Institute of Network Computing and Information Systems Peking University, Beijing 100871, P.R.China pb@net.pku.edu.cn Abstract Automatic keyphrase extraction aims to pick out a set of terms as a representa- tion of a document without manual assign- ment efforts. Supervised and unsupervised graph-based ranking methods have been s- tudied for this task. However, previous methods usually computed importance s- cores of words under the assumption of single relation between words. In this work, we propose WordTopic-MultiRank as a new method for keyphrase extraction, based on the idea that words relate with each other via multiple relations. First we treat various latent topics in documents as heterogeneous relations between words and construct a multi-relational word net- work. Then, a novel ranking algorithm, named Biased-MultiRank, is applied to s- core the importance of words and topics si- multaneously, as words and topics are con- sidered to have mutual influence on each other. Experimental results on two differ- ent data sets show the outstanding perfor- mance and robustness of our proposed ap- proach in automatic keyphrase extraction task. 1 Introduction Keyphrases refer to the meaningful words and phrases that can precisely and compactly represen- t documents. Appropriate keyphrases help users a lot in better grasping and remembering key ideas of articles, as well as fast browsing and reading. Moreover, qualities of some information retrieval and natural language processing tasks have been improved with the help of document keyphrases, such as document indexing, categorizing, cluster- ‡Corresponding author. ing and summarizing (Gutwin et al., 1999; Krul- wich and Burkey, 1996; Hammouda et al., 2005). Usually, keyphrases are manually assigned by authors, which is time consuming. With the fast development of Internet, it becomes impractical to label them by human effort as articles on the Web increase exponentially. Therefore, automat- ic keyphrase extraction plays an important role in keyphrases assignment task. In most existing work, words are assumed under a single relation and then scored or judged with- in it. Considering the famous TextRank (Mihal- cea and Tarau, 2004), a term graph under a sin- gle relatedness was built first, then a graph-based ranking algorithm, such as PageRank (Page et al., 1999), was used to determine the importance s- core for each term. Another compelling example is (Liu et al., 2010), where words were scored un- der each topic separately. In this study, inspired by some multi-relational data mining techniques, such as (Ng et al., 2011), we assume each topic as a single relation type and construct an intra-topic word network for each re- lation type. In other words, it is to map word relat- edness within multiple topics to heterogeneous re- lations, meaning that words have interactions with others based on different topics. A multi-relational words example of …

Content truncated to 500 words