Simulated test collections may find application in situations where real datasets cannot easily be accessed due to confidentiality concerns or practical inconvenience. They can potentially support Information Retrieval (IR) experimentation, tuning, validation, performance prediction, and hardware sizing. Naturally, the accuracy and usefulness of results obtained from a simulation depend upon the fidelity and generality of the models which underpin it. The fidelity of emulation of a real corpus is likely to be limited by the requirement that confidential information in the real corpus should not be able to be extracted from the emulated version. We present a range of methods exploring trade-offs between emulation fidelity and degree of preservation of privacy.
We present three different simple types of text generator which work at a micro level: Markov models, neural net models, and substitution ciphers. We also describe macro level methods where we can engineer macro properties of a corpus, giving a range of models for each of the salient properties: document length distribution, word frequency distribution (for independent and non-independent cases), word length and textual representation, and corpus growth.
We present results of emulating existing corpora and for scaling up corpora by two orders of magnitude. We show that simulated collections generated with relatively simple methods are suitable for some purposes and can be generated very quickly. Indeed it may sometimes be feasible to embed a simple lightweight corpus generator into an indexer for the purpose of efficiency studies.
Naturally, a corpus of artificial text cannot support IR experimentation in the absence of a set of compatible queries. We discuss and experiment with published methods for query generation and query log emulation.
We present a proof-of-the-pudding study in which we observe the predictive accuracy of efficiency and effectiveness results obtained on emulated versions of TREC corpora. The study includes three open-source retrieval systems and several TREC datasets. There is a trade-off between confidentiality and prediction accuracy and there are interesting interactions between retrieval systems and datasets. Our tentative conclusion is that there are emulation methods which achieve useful prediction accuracy while providing a level of confidentiality adequate for many applications.
Many of the methods described here have been implemented in the open source project SynthaCorpus, accessible at: https: //bitbucket.org/davidhawking/synthacorpus/
David Hawking is an Honorary Professor in the College of Engineering and Computer Science at the Australian National University in Canberra. His research in Information Retrieval began in the early nineteen-nineties at the Australian National University where he worked on parallel and distributed IR, and proximity models of retrieval. With ANU colleagues, he coordinated the Very Large Collection and Web tracks at TREC and co-created a number of test collections, notably WT10g and VLC2. Between 1998 and 2008 he worked for the government research organisation CSIRO and, with colleagues, developed an internet and enterprise retrieval system which was spun off as Funnelback Pty Ltd. David worked as Chief Scientist for Funnelback until 2013 when he joined Microsoft and worked on the Bing search engine. He retired from paid work in 2018.Bodo Billerbeck is an applied data scientist at Microsoft Bing, working mostly on core search problems. His interests lie in finding often data-driven solutions to multiple problems in the search stack, including indexing, query reformulation, matching, answer selection and insertion, as well as evaluation. Since completing his Ph.D. at RMIT University in 2005 he briefly worked at Sensis.com.au, but soon moved on to Microsoft. After spending some time embedded at Microsoft Research in Cambridge, UK, he returned to Australia, and recently has come full circle and is enjoying an honorary fellow position at RMIT.Paul Thomas is an applied scientist at Microsoft. His research is in information retrieval: particularly in how people use web search systems and how we should evaluate these systems, as well as interfaces for search including search with different types of results, search on mobile devices, and search as conversation. He has previously worked at Australia's CSIRO and the Australian National University.Nick Craswell is a research manager at Microsoft Bing. Since obtaining his Ph.D. from the Australian National University in 2000 on the topic of distributed information retrieval, Nick has worked on enterprise search, expert search, anchor text, click graphs, image search, offline and online evaluation of Web search, query-independent evidence in ranking, evaluation metrics and web ranking and neural ranking models. He has been a driving force in the TREC web tracks and enterprise tracks and was instrumental in creating various widely used test collections.
