background image
DigiCULT 23
a demanding and time-consuming task that places an
increased cognitive burden on those involved. Doc-
uments may be lengthy or complex, users may have
time restrictions or the initial query may have gen-
erated a poor result set. Therefore, searchers may be
unwilling to provide relevance feedback.
RF systems typically adopt a binary notion of rel-
evance: either a document is relevant, or it is not. If
a document is only partially relevant the approach
requires the searcher to choose between these two
extremes, which means that there is no middle
ground. In such circumstances it can be difficult for
searchers to make decisions on what documents to
assess as relevant.
Implicit RF, in which an IR system unobtrusively
monitors search behaviour, removes the need for the
searcher to explicitly indicate which documents are
relevant. The technique uses implicit relevance indi-
cations, gathered unobtrusively from searcher inter-
action, to modify the initial query. Although not as
accurate as explicit feedback, it has been demonstrat-
ed that implicit feedback can be an effective substitute
for explicit feedback in interactive information seek-
ing environments.
I
MPLICIT
F
EEDBACK
S
YSTEMS
I
mplicit feedback systems remove the responsibil-
ity of providing explicit relevance feedback from
the searcher. These systems infer which pages are rel-
evant by analysing user actions such as the time taken
to read pages. A variety of `surrogate' measures can be
used (hyperlinks clicked, mouseovers, scrollbar activi-
ty, etc.) to unobtrusively monitor user behaviour and
estimate searcher interests. Through such means, it is
possible to estimate document relevance implicitly. If
a user `examines' a document for a long time, or if a
document suffers a lot of `read wear', it is assumed to
be relevant. The motivation for this idea is similar to
that used to promote adaptive search systems, which
develop and enhance their knowledge of search-
er needs incrementally from inferences made about
their interaction. Such systems aim to help struggling
searchers, who may have problems finding what they
are looking for. In a similar way, implicit feedback
based systems in concert with the searcher use implic-
it monitoring of interaction to generate an expand-
ed query that estimates information needs. We have
explored these ideas in two different applications: Web
Search Interfaces and Image retrieval.
Personalised Interfaces for Web Search Systems
We present an interface (Figure 1), which we use
to engender interaction and hence generate more
evidence for the techniques we employ. Our inter-
face works as an adjunct to search systems. The system
takes user queries and despatches them to the Google
Web search engine.
The returned results are processed and present-
ed to the user. We use the top 30 documents from
the Google results and summarise each of them with
respect to the query. By this, we generate a maximum
of four sentences from each document, which are also
known as top ranked sentences. These sentences pro-
vide a much more granular representation of the doc-
ument, with respect to the query.
In general, top ranked sentences, document titles,
the document summary, and each sentence in the
summary within its context of occurrence in the doc-
ument are the various representations a user can view
on the interface.
In our approach searchers can interact with differ-
ent representations of each document. These represen-
tations are of varying length, are focused on the query
and are logically connected at the interface to form
an interactive search path. The representations present
a higher level of granularity than the full text of doc-
uments, allowing the implicit approach to concentrate
on the most relevant parts of individual documents.
Our objective is to help the users in information
seeking activities. During the search process there are
a number of possibilities for personalising. One is the
crystallisation of the user query as the user is exposed
to more and more relevant information. In anoth-
er situation, the underlying information need itself
changes. Information needs are dynamic and may