Information Technology Reference
In-Depth Information
Fig. 6.15
Summary of best average precision scores: relevance feedback from Semantic Web to
hypertext
information retrieval components used in our system is not entirely arbitrary, as
argued above in Sects.
6.5.1.2
and
6.5.2.2
. The gain of our relevance feedback
system, a respectable 16% in average precision over the engine FALCON-S,
intuitively makes the system's ability to place a relevant structured Semantic Web
data in the top rank acceptable for most users.
More surprisingly, by incorporating human relevance judgments of Semantic
Web documents, we make substantial gains over state of the art systems for
hypertext Web search, a 25% gain in average precision over Yahoo! search. One
important factor is the constant assault of hypertext search engines by spammers
and others. Given the prevalence of a search engine optimization and spamming
industry, it is not surprising that the average precision of even a commercial
hypertext engine is not the best, and that it performs less well than Semantic Web
search engines. Semantic Web search engines have a much smaller and cleaner
world of data to deal with than the unruly hypertext Web, and hypertext Web search
must be very fast and efficient. Even without feedback from the Semantic Web, an
average precision of 40% is impressive, although far from the 65% precision using
relevance feedback from the Semantic Web.
Interestingly enough, it seems that pseudo-feedback only helps marginally
in improving hypertext Web search using Semantic Web data. Therefore, it is
somewhat unrealistic to expect the Semantic Web to instantly improve hypertext
Web search. Even with the help of the Semantic Web, hypertext search is unlikely to
achieve near perfect results anytime soon. This should not be a surprise, as pseudo-
feedback in general performs worse than relevance feedback. However, the loss of
performance given by pseudo-feedback in comparison with traditional relevance
feedback shows that for the Semantic Web using pseudo-feedback for concepts
and entities is difficult, as many results that are about highly different things and
subject matters may be returned. However, both pseudo-feedback and traditional
Search WWH ::
Custom Search