single leaf node corresponds to the user's information need; specifically it dictates how the query 

elements are to be used to score the documents (what operators are used in the query). 

d

d

1 

d

2 

d

3 

n 

r

r

1 

2 

r

m 

q

1 

q

2 

q

k 

I 

Figure 15: Inference network 

Each node contains a `link matrix' that calculates the belief for a node given the belief on its parent 

nodes. RF can alter the weights used to calculate the beliefs, in a manner analogous to term reweighting 

approaches. Query expansion is accomplished by adding new query terms as parents of the original 

query nodes. 

Combination of evidence is possible in two ways: by using multiple representations in the concept 

recognition layer, e.g. single term and phrase versions of the same terms, and by the addition of query 

operators. An example of the latter is for the user to ask for documents containing the phrase 

`information retrieval' and any documents containing the word `information' and the word `retrieval'. 

Haines and Croft s tested a number of RF variables: how to select terms for expansion, how to reweight 

terms, the relative weighting of query and new terms and number of terms to add. Although 

performance was variable across the collections tested, they found that query expansion and 

reweighting was effective. They also found support for Salton and Buckley's, [SB90], hypothesis that 

original query terms should be weighted higher than added query terms. They also provide limited 

support for the potential of RF in structured queries   ones that contain operators such as phrases and 

proximity information.  

3.4.2 Multiple retrieval algorithms 

Using more than one retrieval algorithm to score documents is a common way to combine evidential 

sources in IR. Simmonot, [Sim96], proposed a technique for selecting good retrieval algorithms 

techniques based on a user s relevance assessments. In her approach a number of indexing techniques 

were used to represent the content of documents, e.g. keyword representation, conceptual graph 

representation. Each indexing technique was associated with a retrieval algorithm. The user's query was 

submitted to each retrieval algorithm to obtain a number of document rankings and these rankings were 

combined to form a single document ranking that was presented to the user.  

The user was asked to provide a set of relevance assessments, in a similar manner to standard RF. The 

degree of match between the rankings provided by the individual retrieval algorithms and the user s 

relevance assessments was used to score the quality of the retrieval algorithm for the search. This 

quality measure was then used to bias the combination in favour of `good' individual retrieval 

algorithms. A low match between the user s assessments and an individual retrieval function s ranking 

resulted in that retrieval function having a low contribution to the combined ranking at the next 

iteration. A high match meant that the retrieval algorithm would give a high contribution to the 

 29