×

Semantic networks: an evidential formalization and its connectionist realization. (English) Zbl 0703.68089

Research Notes in Artificial Intelligence. London: Pitman Publishing; Los Altos, CA: Morgan Kaufmann Publ., Inc. 222 p. £19.95 (1988).
SHASTRI’s book describes how a high-level specification of hierarchically structured knowledge about concepts and their properties may be encoded as a massively parallel network of a simple processing elements. The evidential formalization of semantic networks leads to a principled treatment of exceptions, multiple inheritance and conflicting information during inheritance, and the best match or partial match computation during recognition. This formalization offers semantically justifiable solutions to a larger class of problems than existing formulations (e.g. default logic). The network operates without the intervention of a central controller or interpreter. The knowledge as well as mechanisms for drawing limited inferences on it are encoded within the network. It uses controlled spreading activation to solve inheritance and recognition problems in time proportional to the depth of the conceptual hierarchy independent of the total number of concepts in the conceptual structure. The number of nodes in the connectionist network is at most quadratic in the number of concepts.
The book has six chapters and one appendix. After the introduction in chapter 1 semantic networks their properties and formalizations are discussed in chapter 2. Especially the significance of inheritance and recognition and the evidential approach to it is pointed out here. Chapter 3 specifies a knowledge representation language. The problems of inheritance and recognition are reformulated in this language. In chapter 4 the evidential formalization and its application to inheritance and recognition are demonstrated. Section 4.1 derives an evidence combination rule. In the following two sections this rule is compared to the DEMPSTER-SHAFER evidence combination rule (section 4.2) and to the BAYES’ rule for computing conditional probabilities. The next two sections develop solutions to evidential inheritance (section 4.4) and evidential recognition (section 4.5) together with constraints for a conceptual structure. The connectionist realization of the memory network is developed in chapter 5. First the need for parallelism is discussed (section 5.1), then the connectionist model (section 5.2) and other massively parallel models of semantic memory (section 5.3) are reviewed. The connectionist encoding of the high-level specification is described in section 5.4 together with the connectivity and computational characteristics of node types. This is followed by examples of network encoding (section 5.5) and the elaboration of some implementation details (section 5.6). In section 5.7 and appendix A there is a proof that the proposed network solves the inheritance and recognition problem in accordance with the evidential formulation and in time proportional to the depth of the conceptual hierarchy. Section 5.8 describes the simulation of the proposed system on a conventional computer together with simulation runs of test examples often cited as being problematic. The book ends with a general discussion (chapter 6).
Reviewer: C.Moebus

MSC:

68T30 Knowledge representation
68-02 Research exposition (monographs, survey articles) pertaining to computer science
PDF BibTeX XML Cite