Databases Reference
In-Depth Information
request are extracted by the semantic associative search, accurate data items which
must be the retrieval results are specified as suggestions. Then, the learning
mechanism is applied to the semantic associative search system to extract the
appropriate retrieval results in subsequent requests.
1.2
A semantic associative search system for Kansei information retrieval
As an associative search system related to Kansei information, we review a semantic
associative search system for image databases dealing with Kansei information
4,
5)
.
The mathematical model of meaning has been designed for realizing the semantic
associative search with semantic computation machinery for context recognition
4,
5)
. This model can be applied for retrieving multimedia information, such as images
and music, with Kansei information. This model can be applied to extract images
by giving the context words which represent the user's impression and contents of
the images.
Images and contexts are characterized by the specific features (words) as Kansei
information and those features are represented as vectors. Those vectors are named
“metadata items for images” and “metadata items for key words.” The important
feature of this model is that semantic similarity between a given context and images
is measured by using the following mathematical operations on the orthogonal
semantic space.
The mathematical model of meaning consists of:
1)
A set of m words is given, and each word is characterized by n features.
That is, m by n matrix is given as the data matrix M.
2)
The correlation matrix of M with respect to the n features is constructed.
Then, the eigenvalue decomposition of the correlation matrix is computed
and the eigenvectors are normalized. The orthogonal semantic space is
created as the span of the eigenvectors which correspond to nonzero
eigenvalues.
3)
Images are characterized by the specific features (words) which correspond
to the n features in the step 1), and the metadata items for the images are
represented as vectors with the n elements, the metadata items for keywords
are also characterized by the same features and represented as vectors.
4)
The metadata items for images and keywords are mapped into the
orthogonal semantic space by computing the Fourier expansion for the
vectors.
5)
A set of all the projections from the orthogonal semantic space to the
invariant subspaces (eigen spaces) is defined. Each subspace represents a
phase of meaning, and it corresponds to a context or situation.
6)
A subspace of the orthogonal semantic space is selected according to the
user's impression, which is given as a context representing Kansei
information with a sequence of context words.
