.Q
rel(D
Q
i
+
.D
j
if
j
)
i+1
=
.Q
i
-
.D
j
if rel(D
j
)
Equation 19: Iterative RF
where Qi = query for iteration i, Qi +1 = query for iteration i + 1,
and
are weights to bias retrieval in favour of the query or relevance information
This technique does not require the user to explicitly request RF, thus side stepping the difficulty of
getting users to interact. However it may not allow users to make relative relevance assessments, which
has been shown to affect users assessments and method of making relevance assessments, e.g. [FM95,
EB88]. The particular implementation also forced users to make a relevance decision. Users, however,
may not always be able to decide on the relevance of a document at the time they view it.
The model was tested in [Aal92] against Rocchio's formula, the Ide dec hi and Ide regular. The model
was also tested against Ide's variable RF, section 2.2.2. This model forms a new query from the first
relevant document and all preceding non relevant documents. This is, then, analogous to the Ide dec hi
that uses all relevant and the first, retrieved, non relevant document, section 2.2.2. The test collection
evaluation showed iterative RF can perform better than the Rocchio, and Ide variants but performs
roughly the same as variable RF.
In a separate experimental investigation Iwayama, [Iwa00], suggests that incremental relevance
feedback of the form proposed by Aalsberg works better for well specified topics. These are topics for
which the set of relevant documents has a high similarity. This is because iterative feedback retrieves
documents that are very similar to the ones used for feedback. It does not, however, perform as well in
retrieving relevant documents that cover a number of topics.
6.2 Ostensive browsing
Campbell's ostensive weighting technique, described in section 3.2, was combined in [Cam99] with a
novel browsing interface, an example of which is shown in Figure 16.
Figure 16: Ostensive browser interface, taken from [Cam99]
39
<
New Page 1
UK Web Hosting