The Semantics of Search - Social Semantics: The Search for Meaning on the Web - page 180

Semantic Web data. Relevance feedback from hypertext Web data can improve
Semantic Web search, and even vice versa, as we have rigorously and empirically
shown. While relevance feedback is known to in general improve results, our
use of wildly disparate sources of data such as the structured Semantic Web and
the unstructured hypertext Web to serve as relevance feedback for each other is
novel. Furthermore as regards relevance feedback, we show using vector-space
models over hypertext data is optimal while language models are optimal when
operating over Semantic Web. These techniques (as evidenced by the failure of
relevance feedback to beat baseline results with incorrect parametrizations) must
be parametrized correctly and use the correct weighting and ranking algorithm to
be successful. It is shown by our results to be simply false to state that relevance
feedback always improves performance over hypertext and Semantic Web search,
but only under certain (although easily obtainable) parameters. We do this by
treating both data sources as 'bags of words' and links in order to make them
compatible and find from the Semantic Web high quality terms for use in language
models. Also, untraditionally, we turn the URIs themselves into words. Our results
demonstrate that our approach of using feedback from hypertext Web search helps
users discover relevant Semantic Web data. The gain is significant over both baseline
systems without feedback and the state of the art page-rank based mechanism
used by FALCON-S and Yahoo! Web search. Furthermore, the finding of relevant
structured Semantic Web data can even be improved by pseudo-feedback from
hypertext search.
More exciting to the majority of users of the Web is the fact that apparently
relevance feedback from the Semantic Web can improve hypertext Web search.
However, pseudo-feedback also improves the quality of results of hypertext Web
search engines, albeit to a lesser degree. Interestingly enough, using inference only
hurt performance, due to the rather obscure terms from higher-level ontologies
serving functionally as 'noise' in the feedback. Lastly, pseudo-feedback from the
hypertext Web can help Semantic Web search today and can be easily implemented.
Indeed, the key to high performance for search engines is the use of high quality data
of any kind for query expansion, whether it is stored in a structured Semantic Web
format or the hypertext Web. However, the Semantic Web, by its nature as a source
of curated and formalized data, seems to be a better source of high quality data
than the hypertext Web itself, albeit with less coverage. While it is trivial to observe
that as the Semantic Web grows, semantic search will have more importance, it is
surprising to demonstrate that as the Semantic Web grows, the Semantic Web can
actually improve hypertext search.

6.9.2 The Equivalence of Sense and Content

The operative philosophical question is: Why does relevance feedback work
between such diverse encodings? Although there appears to be a huge gulf between
the Semantic Web and the hypertext Web, it is precisely because the same content is