TopicRank: Graph-Based Topic Ranking for Keyphrase Extraction

International Joint Conference on Natural Language Processing, pages 543–551, Nagoya, Japan, 14-18 October 2013. TopicRank: Graph-Based Topic Ranking for Keyphrase Extraction Adrien Bougouin and Florian Boudin and B´eatrice Daille Universit´e de Nantes, LINA, France {adrien.bougouin,florian.boudin,beatrice.daille}@univ-nantes.fr Abstract Keyphrase extraction is the task of iden- tifying single or multi-word expressions that represent the main topics of a doc- ument. In this paper we present Topi- cRank, a graph-based keyphrase extrac- tion method that relies on a topical rep- resentation of the document. Candidate keyphrases are clustered into topics and used as vertices in a complete graph. A graph-based ranking model is applied to assign a significance score to each topic. Keyphrases are then generated by select- ing a candidate from each of the top- ranked topics. We conducted experiments on four evaluation datasets of different languages and domains. Results show that TopicRank significantly outperforms state-of-the-art methods on three datasets. 1 Introduction Keyphrases are single or multi-word expres- sions that represent the main topics of a doc- ument. Keyphrases are useful in many tasks such as information retrieval (Medelyan and Wit- ten, 2008), document summarization (Litvak and Last, 2008) or document clustering (Han et al., 2007). Although scientific articles usually provide them, most of the documents have no associated keyphrases. Therefore, the problem of automati- cally assigning keyphrases to documents is an ac- tive field of research. Automatic keyphrase extraction methods are di- vided into two categories: supervised and un- supervised methods. Supervised methods re- cast keyphrase extraction as a binary classifica- tion task (Witten et al., 1999), whereas unsuper- vised methods apply different kinds of techniques such as language modeling (Tomokiyo and Hurst, 2003), clustering (Liu et al., 2009) or graph-based ranking (Mihalcea and Tarau, 2004). In this paper, we present a new unsupervised method called TopicRank. This new method is an improvement of the TextRank method ap- plied to keyphrase extraction (Mihalcea and Tarau, 2004). In the TextRank method, a document is represented by a graph where words are vertices and edges represent co-occurrence relations. graph-based ranking model derived from PageR-A ank (Brin and Page, 1998) is then used to assign a significance score to each word. Here, we pro- pose to represent a document as a complete graph where vertices are not words but topics. We de- fine a topic as a cluster of similar single and multi- word expressions. Our approach has several advantages over Tex- tRank. Intuitively, ranking topics instead of words is a more straightforward way to identify the set of keyphrases that covers the main topics of a docu- ment. To do so, we simply select a keyphrase can- didate from each of the top-ranked clusters. Clus- tering keyphrase candidates into topics also elimi- nates redundancy while reinforcing edges. This is very important because the ranking performance strongly depends on the conciseness of the graph, as well as its ability to precisely represent seman- tic relations within a document. Hence, another advantage of our approach is the use of a com- plete graph that better captures …

Content truncated to 500 words