iii. 

a separate working space was included to view the developing query. The source of query 

terms was also colour coded (initial query, IQE added query, user added query, etc.)  

iv. 

each time the user made a relevant document selection the interface was dynamically updated 

to show the effect of choosing this document.  

The premise behind this interface was that the user would gain more information on the effects of 

actions such as making relevance assessments. The uptake rate for this system was 19.5% and it led to 

the retrieval of further relevant items in 46% of the searches. This system had higher take up and 

effectiveness rates than the first IQE interface but the figures are still lower than the AQE interface. The 

indication is that, although an improved interface can increase the level of use of IQE and the 

effectiveness of term selection, it remains an open problem how to get users to employ IQE in 

operational environments. 

Beaulieu and Jones [BJ98] extended this study by looking in more detail at three factors that affect 

interaction: functional visibility, cognitive load and balance of control between the user and system, 

specifically relating them to this set of experiments. The functional visibility   allowing the user more 

information on how the system works   is important at two levels. Not only must the user be aware of 

what options are available at any stage but they must also be aware of the effect of these options. For 

example, the initial IQE interface was more difficult for user as it separated the act of modifying the 

query and that of assessing relevance.  

The cognitive load, or effort that a user must put into an action, may deter the user from trying an action 

that would be beneficial such as choosing more query terms. Cognitive load is also related to the notion 

of control: generally the more control the user has the higher the overall cognitive load is placed upon 

the user. Thus, as Bates [Bat90] reported, the balance of control, between the system and a user, is a 

question not necessarily of how much control the user has but over what to give the user control. In this 

context it may be preferable to use AQE as a default expansion technique, and to use IQE as an option 

for certain types of search or search stage, rather than use a single method of query expansion.  

Fowkes and Beaulieu, [FB00], in a separate investigation, hypothesised that the complexity of the 

search may be an indicator of when to use AQE or IQE. Searches for which the desired information is 

clearly defined and for which the user can retrieve relevant information easily benefit more from AQE. 

Searches for vague information needs or in cases where little relevant information is being retrieved 

benefit more from IQE. In addition, users are more likely to employ IQE in a complex or difficult 

search. A related point is that users may employ RF, either AQE or IQE, less often when the retrieval 

system is performing well   when it is easy to retrieve relevant information. 

Belkin and Koenneman [KB96] also investigated the use of IQE versus AQE. In this study they looked 

at the performance and behaviour of 64 novice users in the use of three different types of RF 

mechanism: completely automatic query expansion, automatic which showed the expanded query after 

retrieval, and interactive which allowed users to modify query before re evaluation. They also had a no 

feedback control and each user was trained on this baseline system. On the whole the findings were 

positive: the subjects who could control the expansion terms (the third, interactive, case) had better 

performance, and feedback itself gave better performance than no feedback. Users tended to choose 

semantically related feedback terms, and entered fewer terms manually than were suggested 

automatically.  

This set of experiments demonstrated that interactive expansion could give positive results over 

automatic expansion. One particular feature of the experimental design may hold the key to the 

experiments  success. The task that users were given was to develop a good query for an information 

filtering system

32

,  good  in this sense meaning one which was good at retrieving relevant documents. 

The task the users were given, then, was one that concentrated the users  attention on the development 

of good queries, a situation that would lend itself to the use of techniques such as IQE. How to 

encourage users to develop good queries and develop more sophisticated queries does remain a difficult 

area as shown by Beaulieu et al.'s experiments. 

Dennis et al, [DMB98], in a study looking at different types of query expansion techniques found that 

although users could successfully use novel expansion techniques and could be convinced of the 

32

 An information filtering system matches a query (or search profile) against a changing set of documents. Most 

IR systems operate on a fixed set of documents. 

 37