In addition to assuming documents are well-formed, IR researchers often assume that queries are well-formed, meaning that they contain a large number of keywords, and a human can judge whether a document satisfies a query with a reasonably high rate of accuracy.
Various researchers have found that users actually enter between one and three keywords [31, ]. The direct result of this is that the query is under specified, thereby inundating the user with irrelevant documents. Thus, there would seem to be a clear need for Relevance Feedback. However, it is also likely since the initial results are poor, there may be little relevance information initially available -- perhaps only one document. There has been some earlier research with little relevance information [23], however this needs to be revisited and applied to current practices.