An alternative approach to exploiting term dependency is term clustering   grouping sets of related 

terms with a view to selecting query expansion terms from these sets. This can be achieved without 

relevance information (using only statistical information on term similarity to choose expansion terms) 

or with relevance information (using a combination of collection dependent information and 

information from the relevant set to choose expansion terms). Both these methods will typically rely on 

term co occurrence methods to generate clusters but the term co occurrence methods used in the 

literature have generally not provided convincing results [PW91].  

The methods for incorporating term dependence outlined in this section have not produced the increase 

in retrieval performance that may be expected. Partly this may be due to the computational limitations 

of calculating and storing dependence information. Although the term independence methods, such as 

the F4 term weighting scheme, do not explicitly capture term dependence, they do implicitly capture 

some degree of term co occurrence. That is, although the term independence methods do not calculate 

explicit values for co occurrence, one would expect that the terms in the term expansion list would have 

a greater than average degree of term co occurrence. This is because good discriminators of relevance 

are those terms that appear more frequently in the relevant than non relevant documents. How to use 

this co occurrence information successfully, and in a computationally efficient manner, remains an open 

research question. 

3.2 The dynamic nature of information seeking

Implicit to much of the early work on RF is the assumption that users have a fixed information need: 

that the information for which they are searching does not change over the course of a search. Whilst 

this may be true in certain cases, evidence from a range of studies on information seeking, e.g. [Kuh93, 

Ell89, SW99], show that information needs should be regarded as transient, developing entities rather 

than a fixed request. 

The techniques discussed previously modify queries based on the difference between relevant and non 

relevant documents but they do not consider when a document was marked relevant: a document 

marked relevant at the start of a search contributes as much to RF as a document marked relevant at the 

current iteration. If we assume that user's information needs are static then this is correct. However if 

the user s need is developing or changing throughout the search, then documents that were assessed as 

being relevant early in the search may not be good examples of what the user currently regards as 

relevant. Campbell, in a series of papers on developing information needs, has addressed this issue 

through the notion of Ostensive Relevance, [Cam95, Cam99, CVR96].  

The basic premise behind Ostensive Relevance, [Cam95], is that documents selected at the current 

iteration of RF are the best indicators of what the user finds relevant; documents assessed as relevant in 

previous iterations are decreasingly useful at describing a user s information need. Relevant documents, 

then, are not seen as a set of equally important documents but sets of documents of varying importance. 

In [CVR96] Campbell and Van Rijsbergen produce an extension to the probabilistic model of retrieval 

that incorporates an  ageing  component to term weighting. When calculating the weight of a term this 

ageing component incorporates when the documents containing the term were assessed relevant: if the 

documents were marked relevant at an early stage in the search then the term receives a lower weight 

than if the document was assessed relevant in recent iterations. The ageing component can be tuned to 

differentiate more or less strongly between older and more recent documents. In [Cam99] a preliminary 

test of this approach indicated that ostensive weighting can improve searches in fewer search iterations 

than non ostensive approaches. Ruthven et al. also showed ostensive weighting as being beneficial for 

query expansion [RLVR02b]. 

Standard RF techniques, such as Rocchio or F4, will also adapt to changing information needs but they 

will require more evidence to do so as they will require an accumulation of new evidence to outweigh 

the old evidence. Campbell s ageing component reduces this mass of evidence required to shift a query 

towards the new information need. 

Berger and Van Bommel, [BVB97], present a model with similar aims. Their work is specifically 

aimed at characterising the content of documents through hyper indices: hypertext representations of 

document indexes, such as the one shown in Figure 14. 

 24