There exists at lest one architectural description of the Semantic Web (SW) and that is the one by Tim Berners-Lee. I think that it is a good description, but sometimes a bit technology oriented. A true architecture of a complete system should in my opinion not only include decisions on which technology to use (see Different views of the Semantic Web architecture). But the SW has one concept that its architecture must support and this is the URI concept - since it effectively creates the Web. This means that all the technologies that are eventually chosen must support this concept. I will here present one of my views of the architecture of the SW -- "informational view of the architecture".
All information that will be "understandable" to machines will be in the form of triples - A(O,V). This represents the lowest level of the information, i.e. the smallest entities of information that are used to build higher information levels. Of course, in order to represent triples in an unambiguous and universal way we need technologies (serialization syntaxes and utilities) the support this - something that I have written about in other places. All elements of the triples have to be URIs for a number of reasons (i.e. to make precise statements). These triplets will steam from a number of sources; from annotations of web pages, derived from databases, or conversation between SW agents etc.
The triplets are statements. These statements attach properties to resources. So, for a machine to make more use of the statements it has to "understand" what the properties means. By "understand" I mean that the machine knows a set of formal rules of processing that might be used as a result of this property being applied to the resource. In KR, declarative knowledge representation mean, a bit simplified, that properties are defined by using relationships with other properties or ground the properties to "reality" by using human made "axioms". This would create a net of properties, in other words properties (nodes) and relations ships between them (labeled arcs). But in order to create this sort of properties-network we need a set of atom properties, a set of standard properties that constitute the building blocks for connecting properties (the labeled arcs). (This could be subclass-of, is-a, and similar properties.) Using these constructs to build a network of properties and then use this network to map properties of statements onto would facilitate a lot. If we construct a property-network partial understanding is possible. If a set of statements uses a set of properties (i.e. a set of nodes in the properties network) partial understanding could be possible if another application uses a set of properties that are "near by" or closely related. By following routs in the network there might be a common base property that facilitates the partial understanding. This also means that isolated networks of properties are of lesser use than a big combined network (i.e. do not invent ad hoc properties without first looking for existing ones).
In the former version of this article I one more level. But I think that makes things a bit less understandable.
To summarize: A set of statements (triplets) and the "property network" that the properties (the A in A(O,V)) of the statements maps onto, together makes the information to be at a higher level and gives higher semantic meaning to machines.
Sometimes we need to explicitly declare some logical properties of resources, or we need to give properties a bit more logical substance that facilitated the processing of those resources that has this property. At this level is should be possible to make logical statements in a way that makes the logic valuable to machines.
It would be possible to merge level 2 and level 3, but this would make it to difficult in the beginning.
Not complete, I'll be back
We need some way to say, "I know you said that" instead of "I think you said that". A statement is not necessarily true, or might me true in a certain context. And, in our dark times, people don't care much for truth. Using digital signatures - public key encryption - would make it possible to state that: "this statement was made by somebody that hade the private key". Why this is important and what effect it will have are quite simple, therefore not mentioned.
Not compete, I'll be back
These are only my thoughts at the moment, things change and so does my thoughts. A better and more exhaustive description will appear in my Master's thesis.