Это – рабочий авторский перевод. Профессиональный
перевод опубликован в журнале Pattern Recognition and Image Analysis, Vol. 15.
No. 2, 2005, 362 - 364
В.Н. Белоозеров, И.Б.
Гуревич, Ю.О. Трусова
Научный совет по
комплексной проблеме «Кибернетика» РАН
1199991 Москва ГСП-1 ул.
Вавилова, 40
igourevi@ccas.ru, lcmi@ccas.ru
The problem of information search in the domain
of operations on image data is considered. The concept of lexical thesaurus is
proposed as a model of the subject domain ontology that aims an amelioration of
the search machines output.
Information
supply has a paramount significance for investigations in the urgent and
quickly developing sphere of image data analysis and understanding. The search
of information in the modern global nets and big universal data banks cannot be
successful if mere coincidence of the words in the request and a document works
as a relevance criterion. Search machines use various complex criteria which
take into account statistical properties of the words and their grammatical
variability. However the practice of the information retrieval systems
development in the domain of scientific and technological information showed
that the high retrieval quality could be achieved only on the bases of the
semantical analysis of the request and the information resources texts. An
ontological model of the subject domain of the search should be incorporated in
the system. Such model would permit the sense analysis by indication of the
proper place for each text among the real objects of the notions treated in the
text. The proximity of these ontological representations of the texts would be
the criterion for the relevance decision. The human cognitive sphere comprises
such an ontological model in the form of the net of psychic gestalten (ideas)
and logic notions. The systems dealing with natural language texts could
comprise the analogous net in the form of a dictionary which lists semantic
characteristics and cross-references for each term denoting ontological realia.
The
necessity of developing universal and global dictionaries impedes the
full-scale use of the ontological dictionary conception for searching in the
universal multi-branch documentation resources and global nets. Nevertheless a
restricted representation of the isolated subject domain ontology is sufficient
for a search machine to successfully detect the relevant information in that
domain. So, for the information retrieval in the domain of image analysis,
processing and understanding only specific notions have to be provided (with
the restricted amount of general terms). Developing of such restricted
dictionary becomes a quite executable problem.
Search
machine may comprise ontology representation with different degrees of
completeness. As a simplest way, a classification scheme of notions may represent
it. The notions themselves are the kind of generalized classes of mental
gestalten which mirror the objects of reality. These classes are to be realized
in automated system by software connections between the terms naming the
notions.
The
features that indicate the affiliation of the notions to one or another class
arise as the signs of interactions with an investigator. In the domain of image
data handling, the four main notion classes are obviously distinguished: images
(objects of handling), processes of handling, instruments of handling, and
handling tasks. Qualitative words make up two more classes: properties (of
objects, processes, instruments and tasks), and other words of general
character. Finally, the system must reflect its own nature where all the
objects are represented by some texts; so, the names of the texts will form a
category of description terms.
According
to classical theory of classification, the upper level categories should be
divided consecutively to lesser classes by a general feature on each step of
dividing. The resulting tree-structure of classes exists almost in each
information system as its catalogue. However, the search in the catalogue
restricts the search field by the resources that were previously classified and
included in any catalogue division. Classification system of notions permits to
organize an effective retrieval not only in the previously developed
catalogues, but also in a free navigation mode across the information resources
space. In accordance to a request for any object, a search instruction is to be
formed automatically which includes not only the name of the search object, but
also all terms of the subclasses of the search object class. In the result, the
relevance criterion calculation will take into account all kinds of the search
object.
Tree-structures
do not reflect the relations of ontologically and pragmatically significant
notions in full. The one and the same class may be divided by different bases,
and different intersecting rows of subclasses will appear. (Such subclasses of
the description class are image description and structural description
intersecting in the notion structural image description). Many notions may be
treated in the different aspects as subclasses of two and more broader notions.
These phenomena show that the structure of notions relations is not a tree, but
a general case oriented graph without cycles. Such a notion net can be
represented in a tree-structure of terms if the identical terms were placed
sometimes in different nods. This entails to a “tree with glued together nods”.
For instance, the notion high-pass linear filter takes places in two nods: a
subclass of high-pass filter as well as linear filter. A list of the class names
with indication of the immediately broader superclasses and narrower subclasses
is the adequate representation of the notion structure. This leads us to the
representation of the classification system in the form of a dictionary which
indicates by lexical links the interconnections between notions caused by the
inclusion relations for notion extensions. An immediate generalization of this
idea is the inclusion into consideration of the terms with coinciding notion
extensions (synonyms) and partly coinciding as well. Special links are to be
established between the last terms in the dictionary.
Synonymy
link indicates that these terms have the same ontological denotation. These
interconnections permit retrieval of the search objects under alternative ways
of nomination, and evidently raise the search recall. In particular synonymy
links must be registered between different orthographic variants, which has the
great importance for English where British and American norms disagree; for
instance, color (Am) = colour (Brit). If notion extensions intersect partly, a
lexical link is expedient when the notions include mainly the objects from the
intersection. These associative links permit to recall the documents that may
contain the needed information with a great probability. That may be very
valuable when the direct search leads to poor recall.
Additional
information can be obtained also by search on terms with no notion intersection
if they are connected by ontological substances or processes. The most obvious
substantial interconnection is the link between a whole object and its parts.
So, this link should enter the dictionary structure. The operational
interconnections which have a great influence on image analysis processes are the following:
Kind of images - Processing methods
Processing methods - Result
Instrument - Processing method
Property - Bearer of the property
Inclusion
of all these interconnection in a dictionary brings us to the conception of
information retrieval thesaurus (IRT). It is a conventional retrieval tool in
the systems of scientific and technological information what was confirmed by
the series of national and international standards. IRT provides possibility
for classifying the information resources during the search process, on the bases
of requirements of individual request. If the search instruction contain the
ontological links, not only new documents shall appear in the output, but those
documents shall appear in the first rows of relevancy ordered output that
mention the search object not casually, but treat it in detail, taking into
consideration its properties, parts, functions in supersystems, processing
operations. This shall be achieved by accounting for the linked terms in
relevancy calculations. Statistical investigation of Internet search machines
showed that the ratio of pertinent resources was doubled on the first pages of
output. Even more effective retrieval performance will be achieved when the
specially established software analyses thesaurus links in information resources
texts.
More
precise simulation of the domain ontology might include a procedural part that
transforms the thesaurus links into activity corresponding to ontological
processes. This role is performed by the information search software that
spreads the search operations from general terms to the specific ones, from the
whole to the parts, goes from a cause to its effects, from input material to
output data, brings to consideration instrument terms for executing given
processes etc. A system with other destination might be a different model of
the same ontology, with other procedural part, for instance – with the
functions of automatic image processes planning.
The
spontaneity of Internet resources creation excludes the hope on preliminary
classification of the materials or indexing with IRT terms, as it is in the
sphere of scientific and technological information. As a whole the procedure of
thesaurus application for intelligent search in global nets and big documentation
banks may be described as follows. A software interface will be elaborated to
generate a search instruction from the text of the end-user information
request. The search instruction will be designed as to optimize the retrieval
performance of the search machine used. The instruction will include the
request terms and all the synonyms, specific terms, and terms of immediately
connected notions, indicated in the thesaurus. The output will consist of the
fist pages of the search machine recall ordered in accordance with a relevance
criterion. The items of this output will be reordered by new relevance
criterion which calculation accounts for different weights of the request terms
indicated by end-user, and evoked for the search from the thesaurus as well.
The items in the first rows of this list will be the pertinent documents for
the user with the greatest probability.
RAN
Scientific Council on Cybernetics now develops an Internet resource on image
processing, analysis and recognition that will concentrate the access to
information sources [1]. Information access will be supported by semantic
interface on the bases of the Thesaurus for Image Analysis (TIA). The resource
will contain: 1) bibliographical data base of papers, monographs, and
electronic publications; 2) catalogue of Internet resources with the relevant
data; 3) terminological reference base on image processing, analysis and
recognition; 4) means of semantic access to the above mentioned components and external
Internet resources. All these components lean on the use of semantic TIA links
for discovering the data needed on the bases of the request and document
meanings conformity.
Now we have
a TIA version that indicates the above defined links among about 1340 terms (in
English and Russian forms), including 230 of image category, 535 – image
processing, 165 – image analysis and 110 – pattern recognition. In view of the
fact that the conceptual system of our subject domain is insufficiently
structured and rather unstable, TIA will be continuously replenished with new
terms, alternative ways of object nomination and it will incorporate new
semantic relations as the experience of practical TIA exploitation be
accumulated.
Acknowlegement
This paper is partially supported by the Russian Foundation for Basic Research, grant No. 04-07-90187.
References
1.
В. Н. Белоозеров, И.
Б. Гуревич, Д. М. Мурашов, Ю. О. Трусова. Концепция создания интернет-ресурса
по обработке, анализу и распознаванию изображений // Тезисы докладов 7-ой
Всероссийской с участием стран СНГ конференции «Методы и средства обработки
сложной графической информации», Нижний Новгород, 15-18 сентября, 2003. - C. 11-12.