Information Technology Reference
In-Depth Information
Fig. 6.13 Comparing the relevance feedback on the Semantic Web ( RDF ) and hypertext web
( HTML ) both without and with Semantic Web inference
drops even more, from 0.6240 to 0.4557 ( p
05). The use of inference actually
makes the results worse than the baseline performance of language models of 0.5043
( p
<
0
.
05) without either relevance feedback or inference.
The results of using inference to boost hypertext Web results using Semantic Web
equally fail to materialize any performance gains. In this case, inference is used to
expand Semantic Web documents, which are then used via relevance feedback to
improve the ranking of hypertext search. Using the same parameters as above, the
feedback from the expanded Semantic Web data to the hypertext Web results in an
average precision of 0.4273, which is insignificantly different from the baseline of
not using relevance feedback at all of 0.4284 ( p
<
0
.
05) and considerably worse than
not using inference at all, which has a MAP of 0.6549 ( p
<
0
.
05). When pseudo-
feedback is used, the results fall to the rather low score of 0.3861, which is clearly
below the baseline of 0.4284 ( p
<
0
.
05). So, at least one obvious way of use of
simple type and sub-class based Semantic Web inference seems to only lead to a
decline in performance.
Why does inference hurt rather than help performance? One would naively
assume that adding more knowledge in the form of the Semantic Web would help
the results. However, this assumes the knowledge gained through inference would
somehow lead to the discovery of new relevant terms. However, in the case of
much inference with the Semantic Web, this is not the case. For example, simply
consider the Semantic Web data about the query for the singer 'Britney Spears.'
While the first Semantic Web document about Britney Spears gives a number
of useful facts about her, such as the fact that she is a singer, determining that
Britney Spears is a person via inference is of vastly less utility. For example, the
Cycontology (Lenat 1990) declares that Britney Spears is a person, namely that
“Something is an instance of Person if it is an individual Intelligent Agent with
<
0
.
Search WWH ::




Custom Search