Kashori

There is still controversy, but most agree that big globally accepted ontologies are not likely in the near future, if ever. Critics of the Semantic Web (upper case) say such ontologies are required and that this will be the downfall of the technology, since the required global agreements will never happen in the real, factious world. Semantic Web proponents come back and say this criticism is due to a misunderstanding of the way the technology is designed to work. They claim the semantic web (lower case) is designed for the organic growth of a beautiful fractal tangle of interlocking local ontologies. But proponents usually finish with a mumbled bit about agreement, 'there must be some", if just to avoid a cacophony of chaotic islands of solipsisms. So what gives? What sort of agreement, if any, is going to be necessary, accepting that it will not be the one big, global, uber-ontology?

In Symbol Grounding and the Symbolic Theft Hypothesis the authors, Angelo Cangelosi, Alberto Greco and Stevan Harnad, write, "There are two opposite ways of acquiring categories. First, we can use “sensorimotor toil”, in which new categories are acquired through real-time, feedback-corrected, trial and error experience. Secondly, we can use “symbolic theft”, in which new categories are acquired through language, based on hearsay from propositions (e.g., through boolean combinations of symbols describing them). In competition, symbolic theft always outperforms sensorimotor toil. It is more efficient than toil because only one propositional description of a new category is enough to learn it. In contrast, repeated experience is required to learn a category by sensorimotor toil. Due to this significant advantage, it has been hypothesized that symbolic theft is the basis of the adaptive advantage of language (Harnad, 1996). However, some basic categories must still be learned by toil to avoid an infinite regress in the symbol grounding problem."

I believe such symbolic theft (which is a victimless 'crime') is a critical part of the semantic web. The reason is that most of the semantics in the semantic web comes from natural languages like English. Most URIs have recognizable English words embedded in them and that gives them the same shared semantics as the English words themselves. In other words, most URI ride on the back of the common knowledge of natural language words for their meaning. There is no need to come to agreement about such a URI, not because common knowledge is not necessary, but because it already exists about the embedded natural language words that compose the URI. In short, there is no need for a globally agreed upon ontology because the semantic web is grounded in natural language, which is already a widely accepted, if not a globally agreed upon, uber-something, with ontological elements. So I propose ubiquitous copying, or symbolic theft, of word meaning is the actual route to common knowledge, rather than global agreement.

But there is another way that symbolic theft can serve the semantic web. That is through the copying and reuse of URI. The more a URI is copied and reused, the more common knowledge of it accumulates in the communities that use it. If this is true, then the semantic web depends on programmers, database designers, and authors getting into the habit of selecting their URIs for things from a common naming source - the web. And they will do such stealing when it becomes easier and cheaper to copy than it is to invent new terms. When all the creators of data copy the same name for an element up front, there will be no need to translate, merge, integrate, or otherwise massage data to work together. Like in natural languages, these copied terms will form the very essence of communication. As the semantic web really takes off, programmers should and will be able to steal variable and data element names, relational database table and column names, xml schema element names, object attribute names, and more from the web. This will be something available in Eclipse and in other editors where developers actually need to create the names of data elements.

But what about the Symbol Grounding Problem? It is real. It is what makes computers spit out errors so preposterous that even most three year olds could never make them. In persons, it is what causes superstitions, prejudice, and the madness of crowds. In societies, it leads to nationalism, ethnic cleansing, and genocide. Ungrounded symbols run amok, in humans as well as machines. For humans, the main antidote is science and the rational investigation of experience, i.e., sensorimotor toil. Is there a solution to the symbol grounding problem for computing machinery that can be used in the semantic web? That is one of the continuing topics of this blog.

Labels: common knowledge, evolution, language, semantics