通过共享本体达?语义互?作的讨论
Semantic Interoperability via Shared Ontologies
Our goals and claims for semantically driven interoperability vary in how “magic” we expect interoperability to become. Perhaps we can nail down what we think our approaches can achieve and what we are achieving today. This may also help us be more consistent and clear as to our intent and claims.
The following are some “levels of magic” we have seen presented or suggested.
1. Semantic legacy integration
Through semantic technologies existing interoperability points and data repositories will become automatically federated and interoperable. Existing systems that have not been designed to interact will be able to interact based on semantic analysis performed on either their (1a) Definitions (E.G. Schema/interfaces) or (1b) instance data.
2. Federation of semantic “islands”
Interoperability points (E.G. services) and data repositories that have been independently developed as local Ontologies will become automatically federated and interoperable. Existing systems that have not been designed to interact will be able to interact based on semantic analysis performed on their local Ontologies. The important point here is that no formal or standardized shared Ontologies, specifications or models exist. Legacy systems may become semantically enabled via a human process of “grounding” existing specifications using Ontologies.
3. Federation Via shared Ontologies
Interoperability points and data repositories may be independently developed but to become interoperable must share common, formal definitions of terms and common, reference, mid or upper Ontologies. Semantic integration is accomplished by “tracing the links” between these grounding points using inference. A variation point in this approach is that all or some of the terms must be shared and that there may multiple steps in bridging the Ontologies.
A “mid point” between 2 and 3 (Lets call it 2.5) is that the Ontologies share a human language so that while they are independently developed, the do share terms and term matching can be done using broad-based upper Ontologies (E.G. Sumo) or term/concept maps (E.G. Wordnet). In that there is imprecision in the choice of words, techniques are used to determine context and choice of meaning for a term.
4. Point to point integration
As in (2), independent semantic islands are created and then “maps” are created between them in a point-point fashion.
For all of the above, there is a variation point as to the degree of automation and trust. At one level the integration/federation is fully automated and at the other end of the spectrum it a manual process. A valid mid-point is “smart tools” that assist the human process. This can be summarized as;
a) Fully automatic
b) Human assist
c) Manual
We would be interested in a discussion of what is being achieved and what is achievable in the near to mid term. This can then be compared to an evaluation of the value gained for each level of magic.
Discussion of what is possible and valuable;
In our opinion the landscape looks like this;
#1 – Fully automatic semantic legacy integration is a dream, many specifications do not even contain “words” that be analyzed, making even automation assist difficult. In many cases humans can\’t even achieve interoperability/federation without substantial interaction. Smart tooling can assist in capturing information to “ground” existing specifications and may offer suggestions.
Even if this were to be achieved it would take a long while to trust such a capability, particularly for mission critical capabilities.
#2 – Fully automatic integration of semantic islands, perhaps sharing human vocabulary. A high degree of automation may be possible here, but this is still very much research. With advanced techniques it may be possible to have a reasonable degree of confidence for matching but that some external validation will still be required, thus bringing the human in the loop. In many cases human interpretation will still be required – trust being a central issue.
“,1] ); //–>
A “mid point” between 2 and 3 (Lets call it 2.5) is that the Ontologies share a human language so that while they are independently developed, the do share terms and term matching can be done using broad-based upper Ontologies (E.G. Sumo) or term/concept maps (E.G. Wordnet). In that there is imprecision in the choice of words, techniques are used to determine context and choice of meaning for a term.
4. Point to point integration
As in (2), independent semantic islands are created and then “maps” are created between them in a point-point fashion.
For all of the above, there is a variation point as to the degree of automation and trust. At one level the integration/federation is fully automated and at the other end of the spectrum it a manual process. A valid mid-point is “smart tools” that assist the human process. This can be summarized as;
a) Fully automatic
b) Human assist
c) Manual
We would be interested in a discussion of what is being achieved and what is achievable in the near to mid term. This can then be compared to an evaluation of the value gained for each level of magic.
Discussion of what is possible and valuable;
In our opinion the landscape looks like this;
#1 – Fully automatic semantic legacy integration is a dream, many specifications do not even contain “words” that be analyzed, making even automation assist difficult. In many cases humans can’t even achieve interoperability/federation without substantial interaction. Smart tooling can assist in capturing information to “ground” existing specifications and may offer suggestions.
Even if this were to be achieved it would take a long while to trust such a capability, particularly for mission critical capabilities.
#2 – Fully automatic integration of semantic islands, perhaps sharing human vocabulary. A high degree of automation may be possible here, but this is still very much research. With advanced techniques it may be possible to have a reasonable degree of confidence for matching but that some external validation will still be required, thus bringing the human in the loop. In many cases human interpretation will still be required – trust being a central issue.
#3 – Federation Via shared Ontologies – this seems like the state of the art, particularly where there are “agreed”, “reference” or “mid” or “upper” Ontologies. Sharing via an agreed base ontology is essentially equivalent to creating a shared interface/schema standard, but perhaps with a bit more flexibility. Even at this level human validation will be required to establish trust (Unless the vocabulary is so small and structured as to be uninteresting).
Note that this level does not require one super upper ontology, but that there are, at least, bridges between reference, mid or upper Ontologies.
#4 Point to point integration – is certainly tractable as the integration is “engineered”, this can be the basis for bridging/adapting systems and once set up, can be fully automated. However, point-point integration has a much lower value proposition and has been mostly provided for by existing technologies.
Regards,
Cory Casanave
————– next part ————–
An HTML attachment was scrubbed…
URL: http://colab.cim3.net/forum/sicop-forum/attachments/20051025/6285c55a/attachment.htm
——————————
_________________________________________________________________
Message Archives: http://colab.cim3.net/forum/sicop-forum/
Shared Files: http://colab.cim3.net/file/work/SICoP/
Community Portal: http://colab.cim3.net/
To Post: mailto:sicop-forum@colab.cim3.net
Community Wiki: http://colab.cim3.net/cgi-bin/wiki.pl?SICoP
End of sicop-forum Digest, Vol 13, Issue 9
******************************************
“,0] ); //–>
#3 – Federation Via shared Ontologies – this seems like the state of the art, particularly where there are “agreed”, “reference” or “mid” or “upper” Ontologies. Sharing via an agreed base ontology is essentially equivalent to creating a shared interface/schema standard, but perhaps with a bit more flexibility. Even at this level human validation will be required to establish trust (Unless the vocabulary is so small and structured as to be uninteresting).
Note that this level does not require one super upper ontology, but that there are, at least, bridges between reference, mid or upper Ontologies.
#4 Point to point integration – is certainly tractable as the integration is “engineered”, this can be the basis for bridging/adapting systems and once set up, can be fully automated. However, point-point integration has a much lower value proposition and has been mostly provided for by existing technologies.
Let me respond to your #3 below (Federation Via shared Ontologies):
The best way to achieve semantic Interoperability is when two systems are developed from the same ontology. But this is rarely feasible for the following reasons:
a. Legacy systems are already based on an alternate model and usually can’t be redesigned.
b. A single shared ontology, adequate to cover most domains, would be far too big to work with. An ontology of a 1000-2000 terms concepts might be optimal to work with, but an all-purpose ontology might need to be 100,000 concepts or larger.
c. Conflicting Perspectives: Every ontology reflects a perspective of the domain it conceptualizes. Specific information systems would have trouble utilizing a general purpose ontology, because the perspectives built into the ontology could conflict with its view of the world.
But here’s how a shared upper ontology could possibly solve these problems. I say ‘possibly’ because these ideas still need to be demonstrated and proven to be feasible and scalable.
1) Suppose an upper ontology is adopted by a large organization (e.g. Army, DoD, Federal Government, etc.) This could lead to growing industry adoption and eventually adoption as an international standard. I don’t believe ontologists will ever reach consensus on the best upper ontology (and as Chair of the IEEE Standard Upper Ontology WG, I speak from experience) but some standards are established when a large player adopts a specification and others follow. If a number of programs in this group were to adopt an upper ontology, others might follow, and the momentum could take off toward a de facto standard. Such a market-adopted standard could then be more easily approved by a standards body. It would not be perfect, but it could be good enough for many applications.
2) An upper ontology should probably be between 300 and 3000 concepts. Domain ontologies (each compliant with the shared upper ontology) would provide specific concepts for areas such as Medical, Logistics, C4I, Finance, etc. So, how does a medical ontology semantically interoperate with a Finance ontology? I believe the answer lies in abstracting back to more generic concepts, as humans do when we don\’t know much about a given domain. The medical ontology may need to distinguish between 100 types of surgery, but will probably only need to communicate to the Finance system that a charge is for the general concept of “surgery.” While this is not perfect semantic interoperability, it is probably the best that can be achieved, and is also probably good enough for many applications.
3) Legacy systems will map just once to the upper ontology and again to a small number of domain ontologies. This solves the current n-squared problem, where each new ontology has to map to a growing number of other ontologies to enable interoperability.
If this argument and approach stands up, we should see if any of our programs are willing to select (or help develop) an upper ontology, and then design (or map) to it, then see how well the different systems semantically interoperate. But let\’s first see if this logic is sound. Any comments?
Jim Schoening
U.S. Army Communications-Electronics Life Cycle Management Command
Ft. Monmouth, NJ 07703
732-532-5812 DSN 992-5812
.
_____
From: sicop-forum-bounces@colab.cim3.net [mailto:sicop-forum-bounces@colab.cim3.net] On Behalf Of Cory Casanave
Sent: Thursday, October 20, 2005 7:06 PM
“,1] ); //–>
2) An upper ontology should probably be between 300 and 3000 concepts. Domain ontologies (each compliant with the shared upper ontology) would provide specific concepts for areas such as Medical, Logistics, C4I, Finance, etc. So, how does a medical ontology semantically interoperate with a Finance ontology? I believe the answer lies in abstracting back to more generic concepts, as humans do when we don’t know much about a given domain. The medical ontology may need to distinguish between 100 types of surgery, but will probably only need to communicate to the Finance system that a charge is for the general concept of “surgery.” While this is not perfect semantic interoperability, it is probably the best that can be achieved, and is also probably good enough for many applications.
3) Legacy systems will map just once to the upper ontology and again to a small number of domain ontologies. This solves the current n-squared problem, where each new ontology has to map to a growing number of other ontologies to enable interoperability.
If this argument and approach stands up, we should see if any of our programs are willing to select (or help develop) an upper ontology, and then design (or map) to it, then see how well the different systems semantically interoperate. But let’s first see if this logic is sound. Any comments?
