Monday, December 27, 2004
Real World Semantics
More from Andrew D.M. Smith's PhD Thesis. "In this chapter, I have reviewed many recent simulations of the evolution of language, paying particular attention to their models of meaning representation and meaning creation. All of these models claim to have a ‘semantic’ meaning space, yet on closer inspection, the majority of the models had categories which were innate, pre-specified by the experimenters themselves, and had no reference to any external world at all." This sounds like the objection that Google has against human coded Metadata,"Also, it is interesting to note that metadata efforts have largely failed with web search engines, because any text on the page which is not directly represented to the user is abused to manipulate search engines. There are even numerous companies which specialize in manipulating search engines for profit." Humans often times add metadata that has no reference to the external world. It may be mistaken, incomplete, even deceptive. Google's keyword searches on the other hand are made against features that are actually extracted from the documents themselves. Smith continues, "Many of them, in addition, had no sense relations of even the most basic type; there were no relationships at all between one category and another, which were instead atomic, individual items." Sense relations are probably handled well by semantic web languages, OWL in particular. But as Smith says, "The meanings in these models are not under the control of the agents at all; only the experimenters themselves can create new meanings or delete obsolete ones, and the meanings can only appear or disappear from an agent’s repertoire by ‘magic’. In summary, the semantic models of many language evolution simulations are simply not semantic at all, but are instead merely a rudimentary coding system, which the agents in the experiments use as a template with which to decode items expressed in another medium, namely the signals." And this is true with most, if not all of the semantic web implementations I have seen. The agents are not doing anything semantic at all. The categories are not grounded in features of the world.
"On the other hand, there are a sizeable number of experimenters who have made an effort to incorporate some kind of realistic semantic systems, by including an external world of objects which the categories refer to, and by providing various different methods for creating meaning based on the agents’ experience in this world. Meaning creation in these models is based on the discrimination game, in which an agent’s task is to return a category which describes one particular object, distinguishing it from another set of objects. Experimenters have used both classical categorisation and prototype categorisation successfully for their semantic representations; both have advantages and disadvantages, but for the remainder of this thesis I will use my model of semantic representation and creation, which is explored in more detail in the next chapter."
I think these researchers are on the right track.
"On the other hand, there are a sizeable number of experimenters who have made an effort to incorporate some kind of realistic semantic systems, by including an external world of objects which the categories refer to, and by providing various different methods for creating meaning based on the agents’ experience in this world. Meaning creation in these models is based on the discrimination game, in which an agent’s task is to return a category which describes one particular object, distinguishing it from another set of objects. Experimenters have used both classical categorisation and prototype categorisation successfully for their semantic representations; both have advantages and disadvantages, but for the remainder of this thesis I will use my model of semantic representation and creation, which is explored in more detail in the next chapter."
I think these researchers are on the right track.
