Martin Ph. (2003). Knowledge Representation, Sharing and Retrieval on the Web.
Chapter of a book titled "Web Intelligence", (Eds.: N. Zhong, J. Liu, Y. Yao), Springer-Verlag, January 2003. This book includes extended versions of a few selected papers/invited talks from the WI/IAT'01 conference, as well as a few invited papers. More details on "Web Intelligence" can be found on the Web Intelligence Consortium (WIC) site.

PDF version (35 pages; my formatting).     Postscript version (35 pages; exactly as in the book).    
In this article, I mention the idea of "knowledge servers replicating knowledge from each each other" as a way to "combine both the advantages of centralization and distribution". Additional details on this potential "Web of knowledge servers" are given below.

This book chapter is an extension of an article (in HTML), Manageable Approaches to the Semantic Web, presented at the "Practice & Experience" alternate track of the WWW 2002, 11th International World Wide Web Conference, Honolulu, Hawaii, USA, May 7-11, 2002.

A shorter article (15 pages), Large-scale cooperatively-built heterogeneous KBs, focuses on the mechanisms used in WebKB-2 to permit Web users to search, add and annotate knowledge into a same knowledge base. It was presented at ICCS 2001, 9th International Conference on Conceptual Structures.

Another article, Knowledge representation in CGLF, CGIF, KIF, Frame-CG and Formalized-English, focuses on knowledge representation cases and notations. It was presented at ICCS 2002, 10th International Conference on Conceptual Structures.

====================== Web of knowledge servers =====================

Regarding scalability, a solution that combine the advantages of knowledge servers such as WebKB-2 (i.e. the advantages of centralization and the WebKB-2 protocols that encourage re-uses/connections but do not force the users to agree or discuss) and those of distributed Web files (i.e. the advantages of distribution) would be the case where each server
1) periodically checks other related servers (more general servers, competing servers and slightly more specialized servers),
2) integrates (and hence mirrors) all the objects (categories and statements) more general than the objects in a reference collection that it uses to define its "domain" (if this is a general server, this collection is reduced to pm#thing),
3) integrates either all the objects that are more specialized than the objects in the reference collection, or if a certain depth of specialization is fixed, associates to its most specialized objects the URLs of the servers that can provide specializations for them (note: classic server classifications according to general domains are far too coarse to index/retrieve knowledge from distributed knowledge servers; a classification by objects is necessary), and
4) also associates the URL of the more general servers to the direct specializations of the generalizations of the objects in a reference collection (because the specializations of some of these specializations do not generalize nor generalize the objects in the reference collection).
Then, a Web user could first try to use (query or update) a general server (any of them, it would not matter) and this server can, if necessary, redirect her to use a more specialized knowledge server (and so on, recursively, but in theory only one of competing servers has to be tried since they mirror each other). If a Web user directly tries a specialized server, thanks to Point 4 (and further automatic exploration in other servers), this server may redirect her to use a more appropriate server or indicate which other servers may provide more information for her query (or directly forward this query to these other servers).
Of course, integrating knowledge from other servers may not be obvious but it is much easier than trying to re-use/integrate dozens/hundreds/thousands of semi-independently designed small ontologies (hence poorly inter-related and often poorly designed ontologies).