Monday, June 13, 2005

Emergent Semantics in peer-to-peer networks 

13.4.3 Emergent Semantics in peer-to-peer networks
"The mechanisms being developed and studied in the ECAgent projects have the potential to solve an important problem in current distributed information technology, namely the exchange of information in peer-to-peer networks, more specifically the problem of semantic interoperability. Instead of imposing a universal pre-defined ontology over universally defined conceptual schemata, the techniques of emergent ontologies and languages potentially enable each agent to develop a repertoire of grounded categories and labels for these categories and negotiate their use and semantics with other agents. The communication system as well as its semantics is hence emergent and adaptive instead of predefined,
leading to a Self-organisation Approach to Semantic Interoperability (SASI)."
.......
"There are two key problems in peer-to-peer information exchange. The first one is that the data and names used in the taxonomies of one peer (the client) are typically different from those used by another peer (the server) and so the client’s owner cannot know how to formulate the query nor can the server’s owner or the server itself know how to respond if the query is not formulated according to its taxonomy. This is a real problem for users of currently operational peer-to-peer systems. For example, in a music file sharing network, users must try to guess the titles of data and the meaning of names given to the folders and subfolders.

The second problem is that the conceptual schemata used for storing data and meta-data in each information system may be very different, particularly if the meta-data is itself open-ended. Even a simple incompatibility such as usage of different languages can be a problem. For example, the client may have a meta-datum ’country(Belgium)’ whereas the server may have ’pays(Belgique)’. Without semantic knowledge, information systems cannot know how the two meta-data map onto each other, and so a client cannot simply formulate a query for a server using his own meta-data.

Both problems are instances of the so called semantic interoperability problem."
.......
"Alternatively, it is possible that each peer has its own local taxonomy, and its own conceptual schema but that these are translated into a global ontology and conceptual schema which is used for querying and information exchange and
thus exacts as an Interlingua between peers. The translations could be based on defining as much as possible the semantics of the names in the taxonomies. For example, if a user has a sub-folder in his music file system with songs by the Beatles, then the semantics of the implied category is translated in a query over meta-data: ’performed-by(TheBeatles)’. This query can then be used (possibly after translation by a mediator) into a query over the meta-data of a peer. This is the approach currently being explored by the Semantic Web initiative (Berners-Lee et al., 2001), w.r.t. web information systems, and, more generally, by ‘universal’ ontologies such as advocated by CYC or Wordnet (Lenat et al., 1995). It has lead to extensive efforts to develop common ontologies, support systems for defining these ontologies, ways for mapping local schemas into global schemas, and mechanisms to use ontologies in information retrieval, i.e. for mapping categories to data (Davies et al., 2002).

However a consensus is growing that this approach has several major drawbacks (Aberer et al., 2004; Steels, 1997a) as well:

1. The semantic web which relies on universal ontologies just pushes the problem of semantic interoperability to another level. It still requires standardisation based on universal ontologies. It is hard to imagine that a world-wide consensus is reachable and enforcable in every domain of human activity for which information systems are currently in use. Even in restricted domains this is hard because of an increasingly interconnected global world.

2. Human activity and the information systems built for them are open systems. They cannot be defined once and for all but must be adapted to new needs.

3. Peer-to-peer information systems are distributed systems. There is no central control point and so it is not possible to control them centrally.
4. Many information systems already exist and ways should be found to enable their participation in peer-to-peer networks.

An alternative approach to semantic interoperability is to extend information systems with components so that peers can develop and negotiate their own communication protocols in interaction with the data world and the world of human users. So the agents autonomously create an Interlingua which they can each locally interpret. Just as in human natural languages, the consensus will be for ever emergent, adaptive and local. This approach is one of the ways to achieve emergent semantics and is the one that we see as major application area for ECAgents.

The technical solutions that rely on techniques drawn from recent work on language games for robot-robot and robot-human communication (Steels, 1998b), (Steels, 2003a), as further developed in the ECAgents project, have to be expamaknded and changed to make them applicable to the current task. Earlier work in this area has been reported in (Steels, 1997a) and (Avesani and Agostini, 2003) but large-scale application has so far not yet been attempted.

In this approach, Semantic interoperability is seen as a coordination problem between the world, information systems, and human users. A particular kind of ‘semiotic dynamics’ is defined so that both the labels used in peer-to-peer communication and the categories the agents use to interpret these labels become aligned as a side-effect of peer-to-peer information exchange. The labels used in information exchange as well as the semantics of the labels is emergent and the conceptual schemata used for the meta-data in each peer are local and extensable. We note that the Interlingua emerging through agent interactions will never be static and may be locally specialised among a group of peers. The categories defining the ontology of each agent are defined purely in terms of local meta-data and so they are not uniform either."

The emphasis is mine

0 Comments:

Post a Comment

This page is powered by Blogger. Isn't yours?