18

16

14

12

Freezing

Average  10

Test/control

precision 8

Residual (removal)

6

Residual (no removal)

4

2

0

0

1

2

3

4

Feedback iterations

Figure 13: Average precision over 4 iterations of feedback 

2.5 Summary of RF 

In this section we shall summarise the main points from the previous sections and outline some of the 

major issues in the core RF models. In section 2.5.1 we shall summarise the comparison between 

Boolean and best match models, in section 2.5.2 we shall compare the types of best match model, and 

in section 2.5.3 we shall compare the two main components of RF   query term reweighting and query 

expansion. 

2.5.1 Boolean vs Best match 

Although Boolean models are still popular and have strong advocates, e.g. [FST+99], in general there 

are many advantages to best match models over exact match models. The first advantage is that the 

user does not need to generate a query expression in the same way as with the Boolean model. Instead 

they can enter a natural language expression. This means that users can initiate retrieval sessions 

without knowledge of the collection, previous searching experience or experience in creating Boolean 

queries. 

A second difference is that ranking documents allows the users to interact in a more meaningful fashion 

with the system, [Beau97]; documents are presented in order of match and documents are not excluded 

if they miss out elements of the query.  

Thirdly the system can automatically alter a query through RF. The main strength of best match models 

is that they allow for iterative improvement, often using similar techniques to retrieve documents as to 

modify queries. The strength of ranking models for RF is that, after initial querying, the user can 

interact without further describing the information for which they are searching. The RF algorithms 

discussed in the main body of this paper deal almost exclusively with best match algorithms. In the next 

section we shall look at the relative performance of the best match models discussed previously. 

2.5.2 Relative performance of best match models 

In [SB90] Salton and Buckley investigated the relative performance of 12 feedback algorithms on six 

standard test collections

18

. Several of the feedback algorithms (Ide dec hi, F4, Rocchio, and three 

versions of Rocchio with scaling factors for query, relevant and non relevant set) have already been 

discussed.  

A further version of the Ide scheme was used, the Ide regular scheme, [IdS71], which uses all 

retrieved, non relevant documents. The Ide regular is based on the Rocchio formula but omits the 

18

 CACM, CISI, Cranfield, Inspec, MEDLARS and NPL collections. These are relatively short document 

collections ranging from 1, 033 documents (MEDLARS) to 12, 684 documents (INSPEC). 

 20