Topical PageRank: A Model of Scientific Expertise for Bibliographic Search

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 501–510, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Topical PageRank: A Model of Scientific Expertise for Bibliographic Search James Jardine Simone Teufel Natural Language and Information Processing Group Computer Laboratory Cambridge University, CB3 0FD, UK {jgj29,sht25}@cam.ac.uk Abstract We model scientific expertise as a mixture of topics and authority. Authority is calcu- lated based on the network properties of each topic network. ThemedPageRank, our combi- nation of LDA-derived topics with PageRank differs from previous models in that topics in- fluence both the bias and transition probabili- ties of PageRank. It also incorporates the age of documents. Our model is general in that it can be applied to all tasks which require an estimate of document–document, document– query, document–topic and topic–query sim- ilarities. We present two evaluations, one on the task of restoring the reference lists of 10,000 articles, the other on the task of au- tomatically creating reading lists that mimic reading lists created by experts. In both eval- uations, our system beats state-of-the-art, as well as Google Scholar and Google Search in- dexed againt the corpus. Our experiments also allow us to quantify the beneficial effect of our two proposed modifications to PageRank. 1 Introduction For search, the presence of links in a document collection adds valuable information over that con- tained in the text of the documents alone. Each act of linking can be interpreted as a latent judgement of authority or trust which is bestowed onto the linked documents (Kleinberg, 1998). This makes author- ity an objective measure of how important that pa- per is to a community who confer that authority. The citation count is the simplest of these, which has been used successfully for decades for biblio- metrics (Garfield, 1972) and for mapping out scien- tific fields via bibliometric coupling (Kessler, 1963) and co-citations (Small, 1978). More recently, cita- tion counts have been shown to improve effective- ness of ad-hoc retrieval (Meij and De Rijke, 2007; Fujii, 2007). In science, the peer review process ensures that the right to cite is hard-earned, but on the web, hy- perlinking is infinitely cheap. This means that that the authority of webpages cannot simply be approx- imated as the number of incoming links. Algorith- mically more complex authority such as the random- surfer model PageRank (Brin and Page, 1998) or the authorities/hub based algorithm HITS (Kleinberg, 1998)) have spectacularly improved search results in comparison to standard IR models relying on simi- larity calculations based on the words in the text and other text-internal informatioh. Much recent work in bibliographic search has been driven by the intuition that what works for the web should also work for science, even though ci- tations are more comparable to each other in weight than hyperlinks. Case studies comparing PageRank- based authority measures against citation counts alone report some cases where PageRank is supe- rior (Chen et al., 2007; Ma et al., 2008), but exper- …

Content truncated to 500 words