Bilgi Erişimde İlgi Sıralamalarının Artırımlı Olarak Geliştirilmesi

Abstract

Relevance ranking algorithms rank retrieved documents based on the degrees of topical similarity (relevance) between search queries and documents. However, in some cases, sources that address various aspects of a queried topic are needed in addition to the articles that demonstrate a high level of similarity with the search query. Therefore, topical diversity of retrieved articles is also essential, especially in literature search results. Moreover, relevance rankings should be personalized based on users’ information needs. The aim of this study is to develop a new relevance ranking method. To that end, firstly, the relevance rankings for 65 search queries were obtained by applying the LDA (Latent Dirichlet Allocation) probabilistic topic modeling algorithm to the abstracts of some 435,000 physics articles in the iSearch corpus taken from arXiv. Then, these rankings were supported by the pennant retrieval method based on relevance theory, information retrieval, and bibliometrics, and incrementally refined new relevance rankings were created. Findings show that when the relevance rankings obtained by the topic modeling algorithm are fused with the citation data: (1) more enriched relevance rankings containing higher relevance levels with more diverse articles can be created; (2) the rankings can be personalized based on users’ information needs; and (3) the literature can be followed more easily by visualizing the retrieval outputs. Our research is the first to show that LDA-based relevance rankings can be incrementally refined with the pennant retrieval techniques based on citation data. The data used to create relevance rankings such as titles, abstracts, and the total number of citations and co-citations are readily available in the citation indexes. Hence, the method we developed can be used in, for instance, Web of Science, Scopus, and TR-Dizin in the near future once the computation, robustness, reproducibility, and scalability issues are resolved.