A Comparison of Centrality Measures for Graph-Based Keyphrase Extraction

International Joint Conference on Natural Language Processing, pages 834–838, Nagoya, Japan, 14-18 October 2013. A Comparison of Centrality Measures for Graph-Based Keyphrase Extraction Florian Boudin LINA - UMR CNRS 6241, Universit´e de Nantes, France florian.boudin@univ-nantes.fr Abstract In this paper, we present and compare various centrality measures for graph- based keyphrase extraction. Through ex- periments carried out on three standard datasets of different languages and do- mains, we show that simple degree cen- trality achieve results comparable to the widely used TextRank algorithm, and that closeness centrality obtains the best results on short documents. 1 Introduction Keyphrases are the words and phrases that pre- cisely and compactly represent the content of a document. Keyphrases are useful for a variety of tasks such as summarization (Zha, 2002), informa- tion retrieval (Jones and Staveley, 1999) and docu- ment clustering (Han et al., 2007). However, many documents do not come with manually assigned keyphrases. This is because assigning keyphrases to documents is very costly and time consuming. As a consequence, automatic keyphrase extraction has attracted considerable attention over the last few years. Previous works fall into two categories: super- vised and unsupervised methods. The idea behind supervised methods is to recast keyphrase extrac- tion as a binary classification task (Witten et al., 1999). Unsupervised approaches proposed so far have involved a number of techniques, including language modeling (Tomokiyo and Hurst, 2003), clustering (Liu et al., 2009) and graph-based rank- ing (Mihalcea and Tarau, 2004). While supervised approaches have generally proven more success- ful, the need for training data and the bias towards the domain on which they are trained remain two critical issues. In this work, we focus on graph-based methods for keyphrase extraction. Given a document, these methods construct a word graph from which the most important nodes are selected as keyphrases. TextRank (Mihalcea and Tarau, 2004), a ranking algorithm based on the concept of eigenvector cen- trality, is usually applied to compute the impor- tance of the nodes in the graph. Here, centrality is used to estimate the importance of a word in a document. The concept of centrality in a graph has been extensively studied in the field of social network analysis and many different measures were pro- posed, see (Opsahl et al., 2010) for a review. Sur- prisingly, very few attempts have been made to ap- ply such measures to keyphrase extraction. (Lit- vak et al., 2011) is one of them, where degree cen- trality is used to select keyphrases. However, they evaluate their method indirectly through a summa- rization task, and to our knowledge there are no published experiments using other centrality mea- sures for keyphrase extraction. In this study, we conduct a systematic evaluation of the most well- known centrality measures applied to the task of keyphrase extraction on three standard evaluation datasets of different languages and domains1. The rest of this paper is organized as follows. We first briefly review the previous work, followed by a description of the centrality measures. Next, we present our experiments and …

Content truncated to 500 words