Work-in-progress abstract:
Search engines have seen many of the submitted queries before, which might provide valuable relevance feedback for recurring queries that did not change their intent. However, this relevance feedback can often not be applied directly, as documents that were previously relevant to a query might not exist anymore or might have substantially changed content. By counterfactually assuming that the previously relevant document still exists, we can formulate so-called keyqueries so that the previously relevant documents, if they would still exist, would be retrieved at the top positions of the ranking. Our evaluations in the LongEval scenario with varying time gaps between 1
month to 1year and naturally and simulated removed/changed documents show that XYZ.
TBD.