Heterogeneous Networks and Their Applications: Scientometrics, Name Disambiguation, and Topic Modeling

Heterogeneous Networks and Their Applications: Scientometrics, Name Disambiguation, and Topic Modeling Ben King, Rahul Jha Department of EECS University of Michigan Ann Arbor, MI {benking,rahuljha}@umich.edu Dragomir R. Radev Department of EECS School of Information University of Michigan Ann Arbor, MI radev@umich.edu Abstract We present heterogeneous networks as a way to unify lexical networks with relational data. We build a unified ACL Anthology network, tying together the citation, author collaboration, and term-cooccurence networks with affiliation and venue relations. This representation proves to be convenient and allows problems such as name disambiguation, topic modeling, and the mea- surement of scientific impact to be easily solved using only this network and off-the-shelf graph algorithms. 1 Introduction Graph-based methods have been used to great ef- fect in NLP, on problems such as word sense disam- biguation (Mihalcea, 2005), summarization (Erkan and Radev, 2004), and dependency parsing (McDon- ald et al., 2005). Most previous studies of networks consider networks with only a single type of node, and in some cases using a network with a single type of node can be an oversimplified view if it ignores other types of relationships. In this paper we will demonstrate heterogeneous networks, networks with multiple different types of nodes and edges, along with several applications of them. The applications in this paper are not pre- sented so much as robust attempts to out-perform the current state-of-the-art, but rather attempts at being competitive against top methods with little effort be- yond the construction of the heterogeneous network. Throughout this paper, we will use the data from the ACL Anthology Network (AAN) (Bird et al., 2008; Radev et al., 2013), which contains additional metadata relationships not found in the ACL Anthol- ogy, as a typical heterogeneous network. The results in this paper should be generally applicable to other heterogeneous networks. 1.1 Heterogeneous AAN schema We build a heterogeneous graph G(V, E) from AAN, where V is the set of vertices and E is the set of edges connecting vertices. A vertex can be one of five semantic types: {paper, author, venue, institution, term}. An edge can also be one of five types, each connecting different types of vertices: • author — [writes] — paper • paper — [cites] — paper • paper — [published in] — venue1 • author — [affiliated with] — institution2 • paper — [contains] — term All of this data, except for the terms, is available for all papers in the 2009 release of AAN. Terms are extracted from titles by running TextRank (Mihal- cea and Tarau, 2004) on NP-chunks from titles and manually filtering out bad terms. We show the usefulness of this representation in several applications: the measurement of scien- tific impact (Section 2), name disambiguation (Sec- tion 3), and topic modeling (Section 4). The hetero- geneous network representation provides a simple framework for combining lexical networks (like the term co-occurence network) with metadata relations from a source like AAN and allows us to begin to develop NLP-aware methods for problems like sci- entometrics …

Content truncated to 500 words