The world model used here will be referred to as the ontology, and will be defined as a body of knowledge about the world, hierarchically structured as a directed graph, or, more specifically, a tangled tree. The knowledge is separated into two (albeit interlinked) knowledge bases. The first knowledge base, referred to as the ontology, contains knowledge about concepts. The second knowledge base, called the onomasticon, identifies any specific instantiations of those concepts that enter into the body of general (or domain-specific) knowledge that the speaker may expect from the hearer (or writer may expect from the reader).
In addition to this organization into a taxonomy via IS-A links, the ontology also contains numerous other links between concepts, representing connections between those concepts in the world model; these other links will be seen as crucial for semantic dependency structures in the TMRs below. Figure 3A. illustrates a section of a potential ontology with mostly taxonomic links (IS-A), but also with one other type of link. [Figure 3A not available.]In a realistic ontology, each node might have a number of links to other nodes in the ontology. These other links reflect the variety of other relations that exist between concepts, such as IS-PART-OF, IS-AN-OCCUPANT-OF, MANUFACTURED-BY, as well as relations that have been traditionally referred to as case roles in the Case Grammer and related literatures. These relations are represented as slots on concepts, or, graphically, as labelled links between nodes. For example, the EAT concept may have slots such as AGENT and THEME (reflecting the eater and what is being eaten), as well as slots (probably inherited from an ancestor of EAT in the ontology) that are more general, such as LOCATION. These slots actually reflect relations between two concepts, e.g., between the EAT concept and the concept of what is being eaten. These relations themselves represent concepts from the ontology, and appear in the hierarchy and inherit slots and properties from their ancestors.
In addition to these relational slots, concepts may also have literal slots, such as ENGINE_TYPE or COLOR. In the ontology, the relational slots are specified as applying to particular concepts, and thus, by inheritance, to all concepts below them in the hierarchy. In the definition of a relation, there are constraints placed on what the eligible concepts are for its domain and its range; these constraints are also concepts from the ontology. When the relation appears as a particular slot (on a concept in the allowable domain of the relation), there may be more semantic constraints that are locally defined, necesarily more specific than the constraints already specified in the definition of the relation. For example, there might be a LOCATION relation defined in the relation section of the ontology. The domain of this relation might be specified as any EVENT or any PHYSICAL_OBJECT (in other words, events and physical objects may have locations). The range of the relation might be PLACE (that is, only places can be the locations of events or physical objects). The concept of an AIRPLANE_LANDING_EVENT would have a LOCATION slot, and it is indeed within the domain of the relation. However, it may be useful to further constrain the range of the relation (i.e., the allowed value of the slot) in this particular concept to be LANDING_STRIP or such similar concept, which is already within the range of the relation, but is more constrained. This further constraint may be overriden in some text occurrences (as in texts about forced or crash landings), and the algorithm discussed in the section on processing also incorporates constraint relaxation to allow for such situations. The constraints on the allowed fillers of various slots are maintained in the SEM facet of the slot, whereas the fillers themselves are in the VALUE facet.
Often the instantiations of this knowledge base will have names, and may be referred to as named instances. Typical named instances include instances of COUNTRY, CITY, PERSON, for example, for Japan, Paris, and Abraham Lincoln, respectively. These names would necessarily be in a particular language; however, these names are for the convenience of the knowledge base aquisition and maintenance process, just as the symbols attached to concepts in the ontology are just mnemonics for convenience of use. These named instances could be referenced in whatever language as appropriate. Thus such a knowledge base would have an entry for France, and this entry might be called FRANCE if it were convenient for the knowledge base maintainer to use English names; however, the lexicon for any language would reference that instantiation and give it that language's name.
In addition to instantiations of entities, it may also be useful to encode, in a knowledge base, instantiations of events or other concepts from the ontology. The Battle of Gettysburg may be such an event that could be useful for some domains, and hence may be included in the static knowledge base for a particular application or domain.
Knowledge bases of instantiations of concepts may be either static or dynamic. Instantiations of countries and cities, for example, would fall into a static knowledge base, because this type of information would be obtained from gazetteers or from similar references. Instantiations may also be of a more dynamic nature, where concepts which were instantiated by a system using the approach described here would be retained and archived if they appeared particularly useful. This dynamic acquisition of instances would typically be performed while transferring to a new domain.
The type of knowledge base described above has been called an instance memory or an episodic memory, names which typically refer to knowledge bases of entity instances or instances of events, respectively <<<find references>>. We are introducing the term onomasticon to refer to a knowledge base of instantiated concepts, with or without names, static or dynamic. (1)
Any of the onomasticon entries could be referenced from within the lexicon of any language. Just like a specific concept from the ontology would be referenced, a lexicon entry would reference an instantiation instead. Thus for a given language there would be lexicon entries for Japan, Paris, and John F. Kennedy, pointing to the approriate instantiated concept from within the onomasticon, and with the appropriate name for that language forming the lexeme.