This lead to the notion of relevance feedback (RF):  users marking documents as relevant to their needs 

and presenting this information to the IR system. The system can then use this information 

quantitatively   retrieving more documents like the relevant documents   and qualitatively   retrieving 

documents similar to the relevant ones before other documents. The process of RF is usually presented 

as a cycle of activity: an IR system presents a user with a set of retrieved documents, the user indicates 

those that are relevant and the system uses this information to produce a modified version of the query. 

The modified query is then used to retrieve a new set of documents for presentation to the user. This 

process is known as an iteration of RF.  

The mechanism by which an IR system uses the relevance information given by the user is the main 

focus of this paper. The paper covers several aspects of RF: the representations used in RF, how these 

representations lead to deciding how to modify a query and the role of interaction in RF. Section 2 

presents a brief discussion of the retrieval process as a whole and outlines how RF has been 

incorporated into the major retrieval models. In section 3 we discuss extensions and modifications to 

the traditional models of RF. 

Historically, most RF approaches have been based on automatic techniques for modifying queries. In 

section 4 we summarise these approaches. More recently, a number of researchers have examined the 

role of the user in RF and have presented techniques designed to increase the interaction between the 

user and system in RF. These interactive techniques are the main topic of section 5. In section 6 we 

describe interfaces specifically designed to facilitate RF, in section 7 we outline some of the important 

aspects the user that are important to RF, and we conclude this overview in section 8.  

2 The information retrieval process

The IR process is composed of four main technical stages. The first stage, indexing the document 

collection, during which the documents are prepared for use by an IR system, is discussed in section 

2.1. Document retrieval, the process of selecting which documents to display to the user, is described in 

section 2.2. The presentation of retrieved documents and the evaluation of the retrieval results are 

discussed briefly in sections 2.3 and 2.4 respectively. In the section on retrieval we shall outline the 

basic approaches to RF in the major retrieval models. In section 2.5 we shall summarise the difference 

between these main approaches to RF. 

2.1 Indexing 

For small collections of documents it may be possible for an IR system to assess each document in turn, 

deciding whether or not it is likely to be relevant to a user's query. However, for larger collections, 

especially in interactive systems, this becomes impractical. Hence it is usually necessary to prepare the 

raw document collection into an easily accessible representation; one that can target those documents 

that are most likely to be relevant, for example those documents that contain at least one word that 

appears in the user's query.  

This transformation from a document text to a representation of a text is known as indexing  the 

documents. There are a variety of indexing techniques but the majority rely on selecting good document 

descriptors, such as keywords, or terms, to represent the information content of documents. A  good  

descriptor for IR is a term that helps describe the information content of the document but is also one 

that helps differentiate the document from other documents in the collection. A  good  descriptor, then, 

has a certain discriminatory power

1

. This power of a term in discriminating documents can be used to 

differentiate between relevant and non relevant documents, as will be discussed in the section on 

retrieval. 

Figure 1 outlines the basic steps in transforming a document into an indexed form. The first stage is to 

convert the document text (Document text, Figure 1a) into a stream of terms, typically converting all 

the terms into lower case and removing punctuation characters (Tokenisation, Figure 1b).  

1

See [VR79], Chapter 2, for a more detailed explanation of the trade off between the descriptive and 

discriminatory power of terms. 

 2