Article Preview
TopIntroduction
The task of keyword extraction (KE) is to automatically identify a set of terms that best describe the document (Mihalcea & Tarau, 2004). Automatic keyword extraction establishes a foundation for various natural language processing applications: information retrieval, the automatic indexing and classification of documents, automatic summarization and high-level semantic description (Brian & Pradeep, 2010; Cheng & Qu, 2009), etc.
Although the keyword extraction applications usually work on single documents (document-oriented task) (Boudin, 2013; Lahiri, Choudhury, & Caragea, 2014; Palshikar, 2007), keyword extraction is also applicable to a more demanding task, i.e. the keyword extraction from a whole collection of documents (Dostal & Jezek, 2011; Grineva, Grinev, & Lizorkin, 2009; Jones & Paynter, 2002) (collection-oriented task) or from an entire web site (Wu & Agogino, 2003). In the era of big-data, obtaining an effective method for automatic keyword extraction from huge amounts of multi-topic textual sources is a nowadays necessity.
State-of-the-art keyword extraction approaches are based on statistical, linguistic or machine learning methods (Siddiqi & Sharan, 2015; Beliga, Meštrović, & Martinčić-Ipšić, 2015). In the last decade the focus of research has shifted towards unsupervised methods, mainly towards network or graph enabled keyword extraction. In a network enabled keyword extraction the document representation may vary from very simple (words are nodes and their co-occurrence is represented with links), or can incorporate very sophisticated linguistic knowledge like syntactic (Lahiri et al., 2014; Liu & Hu, 2008; Mihalcea & Tarau, 2004) or semantic relations (Joorabchi & Mahdi, 2013; Grineva et al., 2009; Wang, Wang, Senzhang, & Zhoujun, 2014). Typically, the source (document, text, data) for keyword extraction is modeled with one network. This way, both the statistical properties (frequencies) as well as the structure of the source text are represented by a unique formal representation, hence a complex network.
A network enabled keyword extraction exploits different measures for the task of identifying and ranking the most representative features of the source – the keywords. The keyword extraction powered by network measures can be on the node, network or subnetwork level (Beliga et al., 2015). Measures on the node level are: degree, strength (Lahiri et al., 2014); on the network level: coreness, clustering coefficient, PageRank motivated ranking score or HITS motivated hub and authority score (Boudin, 2013; Mihalcea & Tarau, 2004); on the subnetwork level: communities (Grineva et al., 2009). Most of the research was motivated by various centrality measures: degree, betweenness, closeness and eigenvector centrality (Abilhoa & de Castro, 2014; Boudin, 2013; Lahiri et al., 2014; Matsuo, Ohsawa, & Ishizuka, 2001; Mihalcea & Tarau, 2004; Palshikar, 2007).