Learning Word Relatedness over Time

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1168–1178 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Learning Word Relatedness over Time Guy D. Rosin1, Eytan Adar2, Kira Radinsky1,3 1Technion – Israel Institute of Technology, Haifa, Israel 2University of Michigan, Ann Arbor, USA 3eBay Research, Israel {guyrosin,kirar}@cs.technion.ac.il, eadar@umich.edu Abstract Search systems are often focused on pro- viding relevant results for the “now”, as- suming both corpora and user needs that focus on the present. However, many corpora today reflect significant longitudi- nal collections ranging from 20 years of the Web to hundreds of years of digitized newspapers and books. Understanding the temporal intent of the user and retrieving the most relevant historical content has be- come a significant challenge. Common search features, such as query expansion, leverage the relationship between terms but cannot function well across all times when relationships vary temporally. In this work, we introduce a temporal relation- ship model that is extracted from longitu- dinal data collections. The model supports the task of identifying, given two words, when they relate to each other. We present an algorithmic framework for this task and show its application for the task of query expansion, achieving high gain. 1 Introduction The focus of large-scale Web search engines is largely on providing the best access to present snapshots of text – what we call the “Now Web”. The system constraints and motivating use cases of traditional information retrieval (IR) systems, coupled with the relatively short history of the Web, has meant that little attention has been paid to how search engines will function when search must scale not only to the number of documents but also temporally. Most IR systems assume fixed language models and lexicons. They focus only on the leading edge of query behavior (i.e., what does the user likely mean today when they type “Jaguar”). In this context, features as basic as dis- ambiguation and spelling corrections are fixed to what is most likely today or within the past few years (Radinsky et al., 2013), query expansions and synonyms are weighted towards current infor- mation (Shokouhi and Radinsky, 2012), and re- sults tend to include the most recent and popular content. While this problem would seem to be speculative in that it will be years until we need to address it, the reality is the rate of change (Adar et al., 2009) of the Web, language, and culture have simply compressed the time in which critical changes happen. The “Now Web” assumptions are entirely rea- sonable for temporally coherent text collections and allow users (and search engines) to ignore the complexity of changing language and concentrate on a narrower (though by no means simpler) set of issues. The reality is that this serves a significant user population effectively. There are nonethe- less a growing number of both corpora and users who require access not just to what is relevant at a particular instant (e.g., Hathitrust (Willis and Efron, 2013), historical news corpora, …

Content truncated to 500 words