Users: Can't Work With Them, Can't Work Without Them?
Alistair Moffat · Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval · 2022
If we could design the ideal IR "effectiveness" experiment (as distinct from an IR "efficiency" experiment), what would it look like? It would probably be a lab-based observational study [3] involving multiple search systems masked behind a uniform interface, and with hundreds (or thousands) of users each progressing some "real" search activity they were interested in. And we'd plan to (non-intrusively, somehow) capture per-snippet, per-document, per-SERP, and per-session annotations and satisfaction responses. The collected data could then be compared against a range of measured "task completion quality" indicators, and also against search effectiveness metric scores computed from the elements contained in the SERPs that were served by the systems. That's a tremendously big ask! So we often use offline evaluation techniques instead, employing test collections, static qrels sets, and effectiveness metrics [6]. We abstract the user into a deterministic evaluation script, supposing for pragmatic reasons that we know what query they would issue, and at the same time assuming that we can apply an effectiveness metric to calculate how much usefulness (or satisfaction) they will derive from any given SERP. The great advantage of this approach is that aside from the process of collecting the qrels, it is free of the need for users, meaning that it is repeatable. Indeed, we often do repeat, iterating to set parameters (and to rectify programming errors). Then, once metric scores have been computed, we carry out one or more paired statistical tests and draw conclusions as to relative system effectiveness.