benefits of these techniques in a laboratory or training environment, they often stopped using these 

techniques in operational environments. The question may be, then, can we design systems that will 

lead users into spending time developing queries through IQE. 

5.5 Summary of interactive query expansion 

In this section we summarise the case for IQE over AQE. The general intuition that some increased 

control for the user in selecting query expansion terms would be beneficial seems to be valid. Although 

systems have access to internal statistical information that allows them to select good discriminatory 

terms, users can make more informed relevance decision. The question is how this process of query 

modification should be constructed to translate the potential benefits of IQE into actual increases in 

retrieval performance.  

There are several issues involved in this problem. The first is to decide what is the actual role of the 

user: should we ask the user to interactively create queries or perform an editing role on system 

generated queries? How much of the query generating process should be interactive and at what stages 

should we expect and desire user involvement?  

Several of the reasons given by users for not using AQE are also applicable to IQE, [BCK+96, RTJ01], 

e.g. these are time consuming actions, the relation between cause and effect is not clear and on what 

principles the selection of terms should be made is not obvious. The latter point   how terms should be 

chosen   is significant. It may be the case that users are better at eliminating potentially poor terms than 

they are at selecting good terms for query expansion. IR systems need to be able to help users make 

difficult decisions regarding term quality. 

In the next section we shall describe interfaces that were specifically designed for RF. These interfaces 

are an attempt to overcome the user's reluctance to initiate RF. The success of interactive approaches to 

RF may, of course, not simply be a result of the interface or algorithms used by the system. For 

example the characteristics of the user, such as experience with on line searching, and the search itself 

may affect the use and the success of more user oriented methods of interaction. We shall examine 

some of these characteristics in section 7. 

6 Interfaces and RF 

The reluctance of users to engage in RF often comes from a poor understanding of why RF may be 

useful and how RF should be used in a search. This may be because RF is presented as a separate task 

to querying and to assessing retrieved documents. In the next two sub sections we discuss two systems 

that attempt to incorporate RF as a seamless task   the process of RF is integrated into querying and 

assessment of documents.  

The two approaches have a common underlying principle: each relevance assessment given by the user 

initiates a cycle of RF. The major difference between the two approaches   incremental feedback, 

section 6.1 and ostensive browsing, section 6.2   is the interface design and principles. 

6.1 Incremental feedback

Most RF systems treat the process of relevance assessment as a batch process: users are shown a set of 

documents and provide relevance assessments on a number of documents before requesting RF.  

Aalsberg, [Aal92], proposed the alternative technique of incremental RF. Rather than asking a user to 

batch process relevance assessments by assessing a number of documents in a ranking, he suggests 

presenting only one document at a time. The user is asked to make an assessment on the displayed 

document before being shown the next document. With each relevance assessment made by the user, 

the query can be iteratively modified through feedback. The formula used by Aalsberg simplifies the 

Rocchio, Ide dec hi and Ide regular formulae

33

 to the one shown in Equation 19.  

33

 Section 2.2.2. 

 38