Extending the class of queries answerable in ADL generally entails extending the metadatabase. While the traditional library community has established well-defined sets of terms and procedures relating to cataloging issues [1], there is no generally accepted conceptualization of metadata for DL's. Hence the current design of metadatabases and catalogs for DL's is essentially an ad hoc procedure. This is a critical issue because digital technology is providing tools that are relaxing the limits on the extraction, storage, maintenance, and search of metadata. It is likely that metadatabases in the catalogs of DL's will grow to be extremely large and that there will be a blurring of the distinction between items in the ``main'' collection of a library and representations of these items in the metadata.
One strategy for avoiding an
ad hoc growth in the metadata and catalogs of DL's,
is to develop a general model of metadata
that accommodates likely extensions to catalog information.
We therefore provide a sketch of a simple model
of metadata for the ADL catalog component
.
We believe that this model provides sufficient
generality to cover many of the conceivable extensions
to the metadatabase and the catalog that are supportable
with digital technology,
while still providing a useful conceptual tool for designing
and implementing these components in an evolving DL.
We view the assignment of ``meaning'' to a DOBJ as being provided by a set of interpretive mappings or interpretations that assign to the handles of the DOBJ information characterizing any aspect of the DOBJ that may be of value to a user. The information may or may not be inferable from the data of the DOBJ. Examples of such information for the case of a document include an HTML representations of the document, textual representations of the lineage of the document (which may not be inferable from the document itself), symbolic representations of entities existing in the real world that are referred to in the document, and a ``reduced'' graphical representation of the illustrations represented in the document.
We distinguish two main classes of such interpretive mappings: 1) formal interpretations that take as input the handle of a DOBJ and return some ``displayable'' representation (e.g. ascii, html, sgml, postscript) of the ``data'' associated with the DOBJ; 2) informal interpretations that either take as input the output of some formal interpretation and return representations of characteristics of the item (e.g. the creator of an image or the edges computable from the image) or that map the output of some other informal interpretation into the representations of other entities (e.g. a set of edges into faces). Formal interpretations are invertible mappings from the data of the DOBJ, while informal interpretations are generally neither invertible nor computable solely from the data of the DOBJ. One may construct an unlimited number of informal interpretations, and it may be possible to form compositions of such mappings.
As noted by many other researchers (see, for example, [3]), it is useful to distinguish content-independent and content-dependent informal interpretations. Examples of the former include the author catalog and more general ``lineage'' information about items. While it is also useful to define many other subclasses of metadata, for purposes of exposition we only discuss a classification of content-dependent interpretations into domain-independent and domain-dependent interpretations. Examples of domain-independent interpretations include the LoC subject headings and general classes of image processing procedures that extract domain independent features such as the sizes, colors, shapes, and textures of regions from gray scale images. Examples of the domain-dependent interpretations include footprints of library items, gazetteers, and representations of such features as trees or valleys.
We define the metadata of a catalog to be the total set of instantiated interpretations of DOBJ's. It is sometimes possible to view such metadata in terms of a set of metadata records that each contains the handle of a DOBJ and a set of ``fields''. The fields contain representations of the output of all valid compositions of interpretative mappings applied to that DOBJ.
We argue, however that it is more useful to view metadata as a set of mappings rather than a set of ``records''. At one extreme, the application of all compositions of interpretations to all DOBJ's would generally require impossible amounts of precomputation. It would also lead to the need for continual record updates as new interpretive mappings are defined (i.e. as the catalog is extended). At the other extreme, the application of all interpretations at search time is also computationally infeasible. Some combination of precomputed interpretations and interpretations computed on the fly will generally be optimal for a given set of interpretations, as well as permitting the easy addition of new interpretations. The concept of interpretive mappings therefore provides a uniform conceptualization of metadata.
It is straightforward to conceptualize retrieval in terms of this model. For simple cases of retrieval based entirely on a precomputed set of interpretations 1) a user query is translated into a query record, which takes the form of a metadata record but without a DOBJ handle, with the query record employing only those representations used to characterize the output of interpretive mappings; 2) the query record is matched against precomputed metadata records and additional interpretations are computed as required; 3) an appropriate set of DOBJ handles is returned; 4) additional transformations are applied to the retrieved set of DOBJ's in order to extract the requested information.
In general, this procedure is likely to be complex. First, additional informal interpretations may be computed as required. These may take the form, for example, of image processing procedures that are applicable by the user or the form of a browse process in which users apply the informal interpretations that arise in their visual and cognitive processing of items. Second, it is generally necessary to match query records and metadata records based on similarity values, such that a set of DOBJ's will be returned. Third, the search for appropriate metadata records must be aided by the application of indexing strategies.
The metadata and retrieval process of the Alexandria RP is clearly a special case of the preceding model. The metadata corresponding to a single DOBJ may be viewed as taking the form of a single ``record'' in some ``relation'', whose ``field'' values are drawn from USMARC (incorporating FGDC) standards. In terms of the preceding model, we may view this as a set of precomputed informal interpretations of various classes, including originator, publication (content-independent), theme keywords (domain-independent, content-dependent), and bounding coordinates (domain-dependent, content-dependent.)
During search in the RP, the user's query is essentially translated into a query record involving a relatively small number of key pieces of information such as: the ``theme'' of the item requested, the spatial footprint of the item, and the ``time'' at which the item was created. This query record is then matched against the metadata records. In particular, the class and time values are matched exactly, while a truth value is determined for an intersection predicate defined on the footprint values in the query and metadata records. A formal interpretation of the DOBJ's is displayed to the user, who may then request delivery of the data in the DOBJ or the application of transformations to this data.