Relevance feedback is an effective method for iteratively improving a query without increasing the computational requirements to perform the query. Relevance feedback uses the terms contained in relevant documents to supplement and enrich the user's initial keyword query, allowing greater retrieval performance (in terms of precision and recall; see Section 2.1). Since LSI explicitly represents both terms and documents in the same space, a relevance feedback query is constructed in essentially the same way as a regular query.
Given two vectors:
the pseudo-document representing the relevance feedback query is given by
Then, by matching the pseudo-document against each term or document vector (as described in the previous section) and sorting the results, the highest-ranking terms or documents (or all those that exceed some threshold value) can be returned to the user.