This interface contains two features: paths and nodes. A node consists of a retrieved object. In Figure 

16 these objects are images. Clicking on a node will cause the system to perform a RF iteration using 

all the objects in the path that contains the node. A small number of the top retrieved objects are then 

displayed to the user, who may choose to continue the path by clicking a new object or return to a 

previously followed path. If a user selects more than one retrieved object, this corresponds to a 

diverging path: two paths with the same initial components. 

Each selection of a node by a user is taken to be an implicit relevance assessment or expression of 

interest in the object by the user. No explicit request for RF is necessary by the user. The paths 

themselves correspond to multiple iterations of feedback; each object is the result of RF performed on 

the objects preceding it in the path. Objects may appear in different paths as the result of being 

retrieved in response to different RF modified queries. 

This is similar to an extent to the iterative method of RF described in the previous section in that only 

one additional document is added to the relevant set at each iteration. The major interface difference is 

that the user is not asked to make an explicit assessment of relevance or decision on the relevance of a 

document. The major implementational difference is that Campbell uses the ostensive weighting 

extension to the probabilistic model, described in section 3.2. The use of paths also means that RF 

decisions are reversible: the user can backtrack to a previously selected document at any point in the 

search.  

One of the main aims of Campbell's work on ostension is to remove the need for a user to manipulate a 

query. However this also removes the control from the user in modifying the content of the query. A 

user cannot manually manipulate the query as is generally possible with the traditional RF systems. 

Whether or not this hiding of the IR system's functionality benefits the user or not requires further 

investigation.  

Both the interfaces described in this section force users to employ RF. However, in most interfaces the 

user has RF as an option. As shown previously users can be reluctant to initiate relevance feedback 

iterations. Partly this is because the decisions made by RF are not clear to the users and the possible 

effects of RF are not obvious before initiating feedback. Ruthven et al. [Rut02, RLVR03] developed an 

interface to an RF system that used explanations to help users understand what decisions RF had made 

and why these decisions had been made. An example of an explanation is shown in Figure 17. The 

results from these experiments indicate that presenting a more meaningful description of RF can lead to 

more use of feedback techniques by the searcher. 

`As you have not found many useful documents, I have added the following 

words to try to broaden your search couldst inescapeably hillle banquo 

macduff laurenson. You can remove any word you do not think is useful for 

your search'.   

Figure 17: Explanation of RF 

Much more experimentation is required into good interfaces for RF; ones that encourage users to 

initiate feedback and make good relevance decisions. In particular this need for further experimentation 

is necessary because the range of factors that lead to the success or failure of interaction with an IR 

system are very diverse. Many researchers have argued that the process of retrieving relevant 

information is richer and more complex than the relatively simple model described so far, e.g. [Bat90, 

Kuh93, Ing92]. In the next section we shall outline some of the features of user searching 

characteristics that affect how RF is used and its success in improving searching.  

7 User issues 

We can separate out some factors that will affect success of failure of RF algorithms: user experience of 

on line searching, section 7.1, user characteristics, 7.2, and the process of making relevance 

assessments and term selection, section 7.3. 

7.1 User experience 

In [CPB+96, KQC+95] Cool, Koenemann et al looked at the effects on new types of IR systems 

(ranked output, best match model) on the searching behaviour of users who were expert in Boolean, 

 40