THE STRUCTURE AND ORGANIZATION OF A SEMANTIC NET FOR COMPREHENSION AND INFERENCE

We have developed a network representation for propositional knowledge that we believe to be capable of encoding any proposition expressible in natural language. The representation can be regarded as a computer-oriented logic with associative access paths from concepts to propositions. Its syntax is closely modeled on predicate calculus but includes constructs for expressing some kinds of vague and uncertain knowledge. The representation allows the encoding and efficient use of caselike semantic constraints on predicate arguments for the purpose of language comprehension: these constraints are simply implications of the predicates concerned. Our approach to language comprehension is based on nonprimitive representations. We argue that primitive representations of simple propositions are often extremely complex, and offer no real advantages. We have demonstrated these ideas with a mini-implementation capable of mapping certain kinds of declarative sentences into the network representation. The implementation emphasizes the proper handling of iterated adjectival modifiers, especially comparative modifiers. More recently, we have worked on the problem of rapid access to the facts that are relevant to a query. Our solution involves the use of back-link structures from concepts to propositions, called “topic access skeletons,” which conform with general topic hierarchies in memory. For example, the proposition “Clyde is grey” is classified under the “coloring” topic for Clyde, which is subsumed under the “appearance” topic, and in turn under the “external quality” topic, and finally under the “physical quality” topic for Clyde. The form of a query (or of an assertion) can be used to determine what concepts in memory should be accessed as starting points, and what paths in the associated access skeletons should be followed in order to access the relevant information. We have demonstrated the feasibility of building such hierarchies, inserting information into them automatically, and accessing the inserted information with a second experimental implementation. The hierarchic organization appears capable of providing order-of-magnitude improvements in question-answering efficiency, with only a doubling in storage costs.