When the client searches a Q&A collection, the system uses another relevance feedback technique similar to the one proposed by Aalbersberg (1992). The system retrieves Q&A's one by one, as Figure 3 shows. At each retrieval, the client can tell the system whether or not the retrieved Q&A answers the question. The client can also request more Q&A's without evaluating the quality of the current pair, in which case the system retrieves the next best match.
Our technique differs from the one proposed by Aalbersberg (1992) in that the number of negative interactions the client and the system can go through is explicitly bounded. If the client is dissatisfied with the retrieved result when the bound is reached, the system suggests the client browse the current Q&A collection or move the search elsewhere.
Define 
to be 1 when the client believes D, retrieved
at the i-th interaction, to be relevant, and 0 otherwise. Fix
to be an integer upper bound on the number of negative
interactions. Let 
be the vector of the question
submitted by the client. For 
, 
is computed
by 
defined as
where 
is the Q&A vector retrieved by 
,
and is different from every previously retrieved vector;
, 
, 
, 
.