Monday, December 06, 2004
It Takes an Agent to be Semantic
Conceptualizations are used to make discriminations. Such discriminations happen on representations of the world. They happen on things with intentionality, on things that are about something else. This requires a boundary creating an interior space which separates a part of the world from the rest of the world. Within this interior space there must be a capacity to form a representation of that which is exterior. It is upon these interior representations of exterior space that discriminations can be made. The world outside is not divided by conceptualizations, it is one, whole, indivisible. But representations of it, made in the interior space of an agent, permit discrimination of figures out of it's ground of the representation of the world. Such discriminations are concepts. But the map is not the territory, and concepts are not the world. And for this reason they are always approximate and in a sense will always fail to correctly represent the world. Conceptualizations do not replicate the world.
Conceptualizations form discrimination trees. But they are static. Nothing is identified, denoted or defined until an agent performs an act of discrimination using that conceptualization. Therefore, we should speak of identification and denotation events. The valuable thing about an identifier is that it provides to an agent a means of immediate discrimination of an entity from all other entities. URIs, words, symbols or signs can be mapped to conceptual discrimination trees. When this is done, a symbol can be used to select a tree that can be used to discriminate an entity within the representation of the world from the rest of the representation. This is a semantic event. Note that these conditions are necessary. Most machines do not have representations of the world, thus most machines cannot participate in semantic events. But they are also sufficient, in particular, full intelligence is not required. A bounded interior, a representation of the exterior world, a set of conceptual discrimination trees that ranges over the representation and a mapping of trees to symbols are the conditions of semantic agency. And it can be implemented by machines. This is important because if the semantics of the Semantic Web are only usable by humans, we would be better off using natural language, perhaps enhanced with global database schemas. In order for semantics to be used by machinery, these conditions for semantic agency must met.
Conceptualizations are of limited utility. They are always special purpose. No general or upper ontology will ever gain universal acceptance. In fact, even a single agent may not be able to do all it wants to do with a single upper ontology. Agents will need multiple ontologies, different ones for different goals. This makes it implausible to expect that by simply duplicating an ontology in another agent, we can prepare for frictionless communication. If this were true, the IT problem of application integration would be solved by duplicating the database schemas of each application in the database of the other application. And it would have given rise to a global marketplace for shared database schemas.
It often seems that the Semantic Web is viewed as a single, global agent with which all other agents are communicating. On this view, the Semantic Web is sort of a global mind, or universal conceptualization (or at least a global database) accessible to all. Everyone contributes, somehow it all get reconciled, and all can benefit from it.
There is also a growing number of proponents of another view, one which recognizes that agents probably won't ever and do not in fact require a shared conceptualization in order to participate in semantic activities with each other. On this view, the focus is on negotiation, coordination and iterative games to reconcile different mappings to different conceptualizations and arrive at a shared vocabulary that works for the conversation between the two agents.
after Luc Steels
Conceptualizations form discrimination trees. But they are static. Nothing is identified, denoted or defined until an agent performs an act of discrimination using that conceptualization. Therefore, we should speak of identification and denotation events. The valuable thing about an identifier is that it provides to an agent a means of immediate discrimination of an entity from all other entities. URIs, words, symbols or signs can be mapped to conceptual discrimination trees. When this is done, a symbol can be used to select a tree that can be used to discriminate an entity within the representation of the world from the rest of the representation. This is a semantic event. Note that these conditions are necessary. Most machines do not have representations of the world, thus most machines cannot participate in semantic events. But they are also sufficient, in particular, full intelligence is not required. A bounded interior, a representation of the exterior world, a set of conceptual discrimination trees that ranges over the representation and a mapping of trees to symbols are the conditions of semantic agency. And it can be implemented by machines. This is important because if the semantics of the Semantic Web are only usable by humans, we would be better off using natural language, perhaps enhanced with global database schemas. In order for semantics to be used by machinery, these conditions for semantic agency must met.
Conceptualizations are of limited utility. They are always special purpose. No general or upper ontology will ever gain universal acceptance. In fact, even a single agent may not be able to do all it wants to do with a single upper ontology. Agents will need multiple ontologies, different ones for different goals. This makes it implausible to expect that by simply duplicating an ontology in another agent, we can prepare for frictionless communication. If this were true, the IT problem of application integration would be solved by duplicating the database schemas of each application in the database of the other application. And it would have given rise to a global marketplace for shared database schemas.
It often seems that the Semantic Web is viewed as a single, global agent with which all other agents are communicating. On this view, the Semantic Web is sort of a global mind, or universal conceptualization (or at least a global database) accessible to all. Everyone contributes, somehow it all get reconciled, and all can benefit from it.
There is also a growing number of proponents of another view, one which recognizes that agents probably won't ever and do not in fact require a shared conceptualization in order to participate in semantic activities with each other. On this view, the focus is on negotiation, coordination and iterative games to reconcile different mappings to different conceptualizations and arrive at a shared vocabulary that works for the conversation between the two agents.
after Luc Steels
