CGKAT: a Knowledge Acquisition and Retrieval Tool
Using Structured Documents and Ontologies

Philippe MARTIN
University of Adelaide - Computer Sciences department, Australia
e-mail: pm .@. phmartin dot info
This work was completed at the INRIA (ACACIA Project), France

1  Introduction

In Knowledge Acquisition (KA), the knowledge engineer must model and represent expertise into a knowledge base (KB). To do so s/he often searches for information in documents (e.g. interview retranscriptions and technical reports) and structures these documents in order to ease search and modelling. S/he also has to do searches on the knowledge representations to compare, organize and validate them.

In Information Retrieval (IR), the indexation of (parts of) documents by direct hypertext links, keywords or SGML-like tags do not allow the IR system to adequately answer queries expressed at different levels of generality or generate an organised view of the document contents. To allow this, an adequate knowledge representation (KR) language must be used. The more detailed the indexation is, the more precise the answers of the IR system will be. Like KA, precision-oriented IR implies the construction of an organised KB from documents and searches in documents via searches in this KB. It is also eased by a KA/IR system exploiting the structure of documents (i.e. the fact that the document elements are typed and may be linked by composition links or hypertext links, and that various presentation models may be associated and applied to them).

CGKAT [2,3,4] helps KA and IR in two ways.

1. CGKAT integrates a knowledge processor, the Conceptual Graph workbench CoGITo [1], with the structured document editor Thot [6], so that the user may use and combine a) an advanced technique for representing, organizing, accessing and handling knowledge, and b) an advanced technique for displaying, organizing, accessing and handling document elements (DEs). More precisely, these two kinds of techniques may be applied to conceptual graphs (CGs) and to any DE since we allowed a) CGs to be edited and structured with and as any other DEs, and b) any DE (even a whole document) to be indexed by one or several CGs via hypertext links of types Representation and Annotation [3].
   Since documents may store knowledge representations mixed with other DEs, there is no need to maintain a separate KB. Discrepancies between the KB and the KB documentation are thereby avoided. Moreover, Thot allows users to "include" a DE (e.g. a paragraph, a CG, a concept) into several other DEs and then enables hypertext navigation between the inclusions and their sources (in both directions), and automatically modifies the inclusions if the source is modified. This facility allows hypertext navigation between CGs via the concepts they share and eases the handling of modules or views.
   A user may navigate from a DE to its indexations, then navigate between CGs according to the relations between their concepts or their context (described by the DEs which embed them), and then navigate to a DE indexed by a CG. Knowledge-based navigation is thus possible.
   CGKAT can also merge words and their representations into an alphabetically sorted index table, and uses inclusions of this information for building the index table, thus providing a complementary way to compare and access knowledge representations, their authors, the viewpoints they use and the DEs they index.
   We have designed command language for a) combination of queries in the KB or documents, b) the generation of virtual documents (i.e. views on parts of other documents) as answers to these queries, c) the storage of queries into scripts that may be associated to some DEs, thus allowing the use of virtual (dynamic) hypertext links as in some advanced knowledge-based hypertext systems, such as MacWeb. Using queries, the CGKAT user may generate documents that "include" the CGs satisfying conceptual constraints (e.g. the CGs specialising a given CG) and/or the DEs that are represented or annoted by these CGs. The use of inclusions allows users to combine searches by queries and searches by navigations. The scripts may combine commands on the KB with commands accessible from the Unix shell (and then with any tool behaving as a Unix filter). Such scripts may for example be used for testing the KB and generating explanations, (parts of) technical documents or new knowledge representations (e.g. with the provided maximal join command).
   
   
2. CGKAT proposes libraries for easing and guiding the structuration of documents, their representation and the reuse or extension of the representations [2]:
   - structure models and default presentation models for various types of DEs, e.g. Article, Section, Paragraph, Image, Graphics and CG;
   - an initial concept type ontology which merges various KA/KR top-level concept type ontologies and the models associated to these types (e.g. the KADS tasks models) and also the natural language ontology WordNet [5];
   - an initial relation type ontology which merges various relation types ontologies, e.g. thematic, spatial, temporal and argumentative relation type ontologies.
   Searches in such ontologies may be done by navigation, lexical queries (i.e. type name substrings) or conceptual queries (e.g. constraints on authors, domains and supertypes). The natural language ontology WordNet (90,000 types) is not proposed to be wholly included in the user ontology but the types retrieved in WordNet by queries or navigation may be included (with their supertypes) in the user ontology.

2  Architecture

CGKAT has a client/server architecture and also includes the above cited libraries.

The server is made up of the CG workbench CoGITo plus an additional functional interface to allow a) building and retrieval of CGs via Thot menus or textual commands callable inside Thot documents or from an Unix shell or script, and b) browsing ontologies (e.g. WordNet) and modifications on the user ontologies.

The client is the structured document editor Thot plus additional code to allow a) CGs to be edited, handled and stored inside structured documents using the Thot interface, b) the indexation of DEs by CGs, and c) the generation of virtual documents. When a Thot document including CGs is opened, CGKAT also automatically creates them in the base of CoGITo, and removes them when the document is closed. Thus, Thot documents may be used to load, display, browse, structure, document, edit and store selected parts of the KB (an editing operation on a CG via Thot is allowed only if it is accepted by CoGITo, i.e. if it does not violate conceptual constraints previously defined). Conceptual queries may also be done on these selected parts in order to retrieve some CGs or type definitions, or the DEs they index.

Fig. 1.  The CGKAT architecture

3  Applications

CGKAT is a domain-independent KA tool and precision-oriented IR tool. Arbitrary precise representations are enabled by the CG formalism. However, the representations are done manually, therefore their precisions depends only on the users goals. CGKAT has already been used for modelling road accident expertises.

4  Limitations

The main limitations of CGKAT for IR, and to a lesser extent for KA, are the facts that: a) it does not help knowledge extraction (DE representation) by natural language processing techniques, b) no index on knowledge representations is exploited for accelerating their retrieval (except via their membership to documents), e.g. the search for the specialisations of a CG is done by projection of this CG on each CG loaded in main memory), and c) it does not allow the retrieval of paths of concepts and relations inside CGs (inside each CG or inside the CGs seen as a global semantic network). For these reasons, CGKAT is mainly interesting from a KA viewpoint.

5  Conclusion

CGKAT helps KA and IR by combining a CG workbench with a structured document editor, and by providing default general ontologies and functions to search and handle them. Thus, compared to other current KA or IR systems, it provides more ways or more precise ways to represent or index DEs and structure these DEs or the knowledge representations, and more guidance or freedom for representing or structuring information (KA or IR systems generally do not provide a default ontology or provide a non-extensible ontology). Articles related to CGKAT are accessible at http://www.phmartin.info/CGKAT/.

CGKAT could be quickly extended by using extensions of CoGITo (e.g. with rules and CGs index), of Thot (e.g. Alliance for document cooperative edition and Amaya for Web-browsing) and of WordNet (e.g. EuroWordNet and International WordNet).

Many of the CGKAT functionnalities could also be obtained and complemented using other knowledge processors, other structured document editors or browsers and other information management systems (see The WebKB set of tools in these CGTOOLS'97 proceedings for more details on our work in that direction).

6  References

1. O. Haemmerlé, CoGITo: une plate-forme de développement de logiciels sur les graphes conceptuels. Ph.D thesis, Montpellier II University, France, Jan. 1995.

2. P. Martin, Using the WordNet Concept Catalog and a Relation Hierarchy for KA, in Proceedings of Peirce'95, Santa Cruz, California, Aug. 18, 1995.

3. P. Martin, and L. Alpay, Conceptual Structures and Structured Documents, in Proceedings of ICCS'96, Sydney, Australia, Aug. 19-22, 1996.

4. P. Martin, Exploitation de graphes conceptuels et de documents structurés et hypertextes pour l'acquisition de connaissances et la recherche d'informations. Ph.D thesis, University of Nice - Sophia Antipolis, France, Oct. 14, 1996.

5. G.A. Miller, WordNet: A Lexical database for English, in Communications of the ACM. Nov. 1995.

6. V. Quint, and I. Vatton, Combining Hypertext and Structured Documents in Grif, in Proceedings of ECHT'92, D. Lucarella, ed., ACM Press, Milan, Dec. 1992.