Re: WOLREQS: Draft requirements document

It's unfortunate that we didn't have more time to discuss these issues.
I still feel that our other requirements have more of a language impact
than explainability. Since Jos also appears to agree with my concerns
and noone else has expressed an opinion, I think we're best off leaving
it in the set of "other considered requirements" instead of putting it
in the main group.

I've made a (hopefully) final pass on the document in which I've:

- added an appendix to record our other candidate requirements
- moved explainability to this appendix
- put the document back in ASCII text format
- removed the section on requirements from other groups (I think the
merge will happen at, or as a result of, the face-to-face).

You can find this document attached below. This is the version I intend
to send to the Working group by 1pm East Coast time. Please send me any
last-minute comments or quick fixes before then. Please do not edit and
send me the entire document, just send a list of suggested changes.

Deborah, in order to turn on word-wrapping in Netscape Messenger, try:

1) select Edit-Preferences from the menu bar
2) open the "Mail & Newsgroup" category (if not already open)
3) Select "Messages"
4) Under message wrapping, check the box that says "Wrap incoming plain
text messages to window width"


Deborah McGuinness wrote:
> I think we have included a number of things that are requirements for an OWL system not just for pure language.
> r1 for example - shared ontologies - that is a requirement for applications and not for the language,
> r3  ontology evolution - that is a requirement for ontologies over time and not just for the pure language.
> these and explanation have some impact on the language extensions.
> I would not put all of the functions that we are looking for in any of these requirements in the core language.
> for explanation, it is very useful to be able to annotate terms with both if the information is told vs. derived, if told what ontology(ies) it came from, if derived how derived,  if derived, if it was derived using default info
> or "hard/fast" rules.
> if we were considering proof checking, explainability is a necessary precondition to proof checking.
> also, as i input explanation into classic, we found that it was important to know -
> the source of told information (what ontologies things came from)
> if information was told or derived
> if it was derived how it was derived
> we provided a way of rerunning deductions and annotating structures.  It does not have to be done that way, but that was a useful approach.
> in a new world of the web, derived information needs more support since it may be derived exclusively from information from authoritative sources or some from authoritative sources and some from other sources.
> also, it could be derived from default rules or from hard/fast rules....
> and for whether it absolutely necessary for owl, most of our requirements are not absolutely necessary for all use cases but are absolutely necessary for some use cases.
> I would claim that explanation is absolutely necessary for use cases for intelligence analysts, help desk applications, and configuration, among others.
> all of our other requirements  have the same property -
> some use cases will use base ontologies without extension - meaning r2 is not absolutely necessary.
> some will use proprietary ontologies that may not be modified and thus have little need for r3 and no need for r4
> many web ontology uses do not use deduction and thus do not need r5 - inconsistency detection - or just work in such contradictory environments that they do not care about r5
> some applications use small ontologies and have small or medium size kbs and thus do not care at all about r6
> some have trained users or are backend systems for other systems and thus do not need r6 at all
> some have no need for xml and thus have no need for r10
> i doubt if many applications do not need search  but some dont need much other than text search thus having small needs for r12
> many web applications need very simple ontologies thus expressiveness r14 is definitely not absolutely necessary for all applications...
> Also, I am ok listing first the requirements that had consensus  but
> i would not want to take the initial vote with a very heavy weight since i think votes were made before discussion or further description/defense was made for some of the requirements.
> actually i think all of them have validity but the ones that did not get consensus were basically not clearly articulated by themselves or in separation from other requirements.
> since time is short, i am willing to have the separation that we have:
> a list with details of the requirements that have consensus by meeting time.
> a list of requirement descriptions for the other topics so that at least we do not loose the work to date prefixing this list with other desirable requirements.
> i made the third category - other requirements emerging from other groups since we will be responsible ultimately for collecting the requirements and those clearly emerged from that group.
> on format,
> i didnt know how to have the netscape mailer do the wrapping and found it to be a problem a few times when i was trying to print out the documents for reading.
> I assume many will find on the plane that they have printed out things that have been truncated so i wanted to have something that had a good printing format.
> if you want straight text with line breaks that is fine.
> d
WebOnt General Requirements

This document was prepared by the General Requirements Subgroup, which consists of the following individuals:

Jeff Heflin (co-chair)
Deborah McGuinness (co-chair)
Jeremy Carroll        
Dan Connolly          
Jos De Roo            
Pat Hayes             
Ned Smith             
Herman ter Horst      

The purpose of this document is to identify requirements that are too general to result from any single use case area, cut across all use cases areas, or are not directly related to the existing use cases, but nonetheless important.

The following requirements are recommended by the group.

R1. Shared Ontologies

Ontologies are publicly available and different data sources can commit to the same ontology for shared meaning.

Any use case in which distributed data sources use shared terminology.

Interoperability requires agreements on the definitions of terms. Ontologies can provide standard sets of terms and formal definitions of those terms. Data sources that commit to the same ontology explicitly agree to use the same terms with the same meanings.

Although DTDs and XML Schemas can be used to define the syntax of a language, they cannot provide machine-readable semantic definitions for the terms of the language. A web ontology language needs:

1) syntax for defining ontologies
2) syntax for WebOnt documents to commit to one or more ontologies
3) syntax for disambiguating when two or more committed ontologies contain the same term (perhaps by specifying an order of preference for ontologies)

DAML+OIL provides an daml:Ontology element and a number of primitives for defining classes and properties. It uses XML namespaces to identify which set of terms it is using and a daml:imports statement can be used in data documents to effectively commit to the definitions of specific ontologies.

R2. Ontology Extension

Ontologies can be extended by other ontologies in order to provide additional definitions.

Any use case in which the providers of data are decentralized.

Often, shared ontologies are not sufficient. An organization may find that an existing ontology provides 90% of what they need, but the remaining 10% is critical. In such cases, the organization should not have to create a new ontology from scratch, but instead be able to create an ontology which extends an existing ontology and adds any desired terms and definitions.

Although RDF uses XML namespaces to include names from other schemas, there is no discussion of how this relates to the definitions of the names. For example, if definitions for a single name occur in three different schemas, but a document only includes the namespace for one, then it is unclear whether the definition intended by the document is the conjunction of the three schemas, or only the schema which the namespace includes. To explicitly express which definitions are intended, the web ontology language needs syntax for expressing ontology extension.

daml:imports allows an ontology to include the definitions from another ontology.

An important issue is determining the precise semantics of the extension mechanism. Is it equivalent to including the extended ontology in the new document? Does it allow definitions to be refined or restricted?  Also, what happens if the extensions generate incoherent term descriptions?

R3. Ontology Evolution
Ontologies can be changed over time and data sources can specify which version of the ontology they commit to.

Any use case in which the ontology could potentially change, and in particular those in which the owner of the ontology is different from the data providers.

Since the web is constantly growing and changing, we must expect ontologies to change as well. Ontologies may need to change because there were errors in prior versions, a new way of modeling the domain is preferred, or reality has changed (e.g., the addition of new technology). A web ontology language must be able to accommodate ontology revision. Note that ontology evolution is different from R2, because R2 does not change the original ontology. An important issue of revision is whether or not documents that commit to one version of an ontology are compatible with those that commit to another. Both compatible and incompatible revisions should be allowed, but it should be possible to distinguish between the two. Note that it is possible for a revision to change the intended meaning of a term without changing any axioms, thus determining backwards-compatibility requires more than a simple comparison of axioms.

Allowing ontologies to change arbitrarily can have undesirable side effects in documents that committed to prior versions of the ontology. Since these documents are distributed and owned by different parties, it is impossible to coordinate an update to the ontology with updates to all documents that depend on it. One possible solution is:

1) Each revision of an ontology is a separate document that has a unique identity in web space (its own URL).

2) A construct for expressing that an ontology revises a prior version

3) A construct for expressing that a revision is backwards-compatible with one or more prior versions

4) A construct for expressing that certain terms are deprecated, and thus maintained in a revision simply for compatibility with earlier versions. This allows terms to be phased out while retaining backwards-compatibility.

5) A way of committing to a specific version number of an ontology or the latest version of that ontology.

In DAML+OIL, each ontology has its own URL. Each ontology has a daml:versionInfo element that contains a string giving information about the version it represents. However, there is no specified structure for this string, and thus it is of little use for automated software that wishes to determine which ontologies are prior versions of other ontologies. DAML+OIL does not include features for specifying backwards-compatibility or deprecation.

R4. Ontology Interoperability

Different ontologies may model the same concepts in different ways. The language should provide primitives for relating different representations, thus allowing data to be converted to different ontologies, and enabling a "web of ontologies." However, this requirement must be balanced with the need for scalability (R6).

Any use case in which data from different providers with different terminologies must be integrated.

Although shared ontologies (R1) and ontology extension (R2) allow a certain degree of interoperability between different organizations and domains, there are often cases where there are multiple ways to model the same information. This may be due to differences in the perspectives of different organizations, different professions, different nationalities, etc. In order for machines to be able to integrate information that commits to heterogeneous ontologies, there need to be primitives that allow ontologies to map terms to equivalents in other ontologies.

There are many ways that different ontologies can model the same concepts, resulting in different types of heterogeneity. One approach is to have the expressivity of first order logic, which can be used to define articulation axioms and resolve most of the kinds of differences. However, this solution is not in line with the scalability requirement. Below is a list of language features that can be used to map heterogeneous ontologies; the web ontology language should include some (but probably not all of these).

1) subclass/superclass relations
2) inverse relationships
3) equivalence of concepts (for classes, properties, and individuals)
4) logical constructs (implication, conjunction, disjunction)
5) arithmetic functions
6) aggregation (e.g., like SQL GROUP BY)
7) string manipulation
8) procedural attachments (executable code, possibly Java, that can be used to define arbitrarily complex mappings)

Another approach may not to be to include the expressivity of first order logic but instead reduce the expressive power of the language in which the articulation or mapping axioms are expressed.  For example, a limited language might be allowed to state that term A in Ontology 1 is equivalent to term B in Ontology 2 with the following additional restrictions (where those restrictions are not in a language that is as expressive as FOL).

DAML+OIL contains the rdfs:subClassOf and rdfs:subPropertyOf relations for defining taxonomic relations of classes and properties, respectively. It also has the daml:equivalentTo family of properties (daml:sameClassAs, daml:samePropertyAs, and daml:sameIndividualAs) for expressing equivalence of classes, properties, and individuals. There is a daml:inverseOf property for defining inverse relationships. Finally, the description logic primitives allow for mappings similar to some of the logical constructs. However, DAML+OIL does not have features for expressing implication, arithmetic functions, aggregation, string manipulation, or procedural attachments.

R5. Detect Inconsistency
Different ontologies or data sources may be contradictory. It should be possible to detect inconsistencies.

Any use cases in which decentralization of data and lack of controlling authority can lead to conflicts in the data. Any ontology extension task that may result in incoherent descriptions (possibly by extending an ontology in a way that generated an overconstrained term).

The Web is decentralized, allowing any one to say anything. As a result, different viewpoints may be contradictory, or even false information may be provided. In order to prevent agents from combining incompatible data or from taking consistent data and evolving it into an inconsistent state, it is important that inconsistencies can be detected automatically.

First the language must be able to express inconsistent situations. This could be done with a negation operator, disjointness of sets, cardinality restrictions, etc. Second, a reasoning component must be able to detect inconsistencies.  Third, some reporting mechanism must be made available to report inconsistencies possibly along with an explanation of how the inconsistency was generated.

DAML+OIL can express disjoint classes (using daml:disjointWith), cardinality restrictions (with daml:cardinality, daml:minCardinality, and daml:,maxCardinality), and complements (with daml:complementOf). Using a description logic reasoner, inconsistencies within an ontology, between a set of ontologies, and between ontologies and data sources can be detected.

R6. Scalability
The language should be able to be used with large ontologies and large data sets. However, the language must balance this requirement with Expressiveness(R10).

Any use case that uses large ontologies or large data sets.

There are over one billion pages on the Web, and the potential application of the Semantic Web to embedded devices and agents poses even larger amounts of information that must be handled. The web ontology language must support systems that can scale to these sizes.

Many expressive languages are intractable, resulting in them not being scalable. One solution is to restrict the language to features that have efficient algorithms for reasoning. Two candidates for limited reasoning are description logics and datalog.

DAML+OIL is based on  a description logic language for which tractable reasoners may be built.

R7. Ease of Use
The language should provide a low-learning barrier and have clear concepts and meaning. The concepts should be independent from syntax.

Markup and querying of semantic web documents by users, either directly or indirectly.

Although ideally most users will be isolated from the language by front end tools, the basic philosophy of the language must be natural and easy to learn. Furthermore, early adopters, tool developers, and power users may work directly with the syntax, meaning human readable (and writable) syntax is desirable.

Where possible, use concepts and idioms that are familiar to ordinary software engineers and computer scientists. For example, relate ideas to object oriented and or relational databases.  One or more presentation syntaxes could be provided that are natural to users.

DAML+OIL's classes are equivalent to classes in object-oriented terminology. DAML+OIL also incorporates features from frame-based systems and description logics. However, DAML+OIL's syntax has not retained features that have made some of its foundational components such as object-oriented systems and frame systems more usable.  As the language was generated to be compatible with the RDF framework, arguably some of the useful syntax features present in other languages were not preserved.

R8. XML syntax
The language should have an XML serialization.

Exchange of ontologies and data in a standard format.

XML has become widely accepted by industry and numerous tools for processing XML have been developed. If the web ontology language has an XML syntax, then these tools can be extended an reused.

Provide an XML syntax for the language.

DAML+OIL extends RDF and RDF-Schema, which in turn have XML serializations.

There is lack of consensus as to whether the language should also build on RDF and RDF Schema. The arguments for building on RDF are that it is a W3C Recommendation and there exists software for parsing it. Arguments against RDF include that it does not have the widespread acceptance of XML, and trying to fit DAML+OIL into it has occasionally resulted in awkward language constructs.

R9. Ontology-based search
Conceptual search or semantic search -- search exploiting the meaning of terms instead of just the syntax of the search terms.

Any 'find' capability that attempts to exploit term meaning including synonyms, subclass/superclass relationships, relationships between terms, or parametric search.

It is widely recognized that simple statistical information retrieval techniques for search have limitations.  If for example, a user is looking for a car when a web site contains only the word automobile, then no matches will be found even though information exists that would be relevant to the user's query.  Market research shows that search is one of the most important functionalities on web sites ([1] ) and also that sophisticated search is imperative for e-commerce sites ([2]).  Also, as objects become complicated and compositional, it becomes more important to be able to search for relationships between terms (such as the "capital of France" and not just retrieve all documents about France) or "red cashmere sweater with a price under 150 dollars." The latter is typically called parametric search where parameters on the class sweater are restricted.  Parametric search, relational search, and standard conceptual search all rely on term meaning and term composition of search terms and also on the meaning of the terms in the document to be retrieved.

Use background ontologies to provide:
1) query expansion using synonyms and subclasses (and possibly more) from the background ontology
2) understanding of term relationships, e.g., capitals of countries are cities that are located geographically inside the country.
3) identify parameters appropriate to classes and their value restrictions

Background ontologies can be expressed in DAML+OIL.  Also, queries can be formed from DAML expressions.

R10. Expressiveness
The language should be as expressive as possible, given a balance with requirement R6, Scalability. Expressivity determines what can be said in the language, and thus determines its inferential power and what reasoning capabilities should be expected in systems that fully implement it.

Any use case that requires the representation of diverse knowledge.

The degree of semantics that can be expressed by a language depend on the primitives that it provides. An expressive language contains a rich set of primitives that allow a wide variety of knowledge to be formalized. A language with too little expressivity will provide too few reasoning opportunities to be of much use and may not provide any contribution over existing languages.

One possible approach is to base the language on first order logic, but this conflicts with R6, Scalability. Therefore, more restricted, yet more scalable alternatives should be considered. Two such alternatives are description logics and datalog.

DAML+OIL is based on an description logic that has had its constructors chosen to maintain tractability of reasoning.


The following items were discussed as candidate requirements but we were unable to reach consensus as to whether they were actual requirements for WebOnt.

C1. Explainability
 An ontology language and its environment may provide information that is believed to be true. A system should be able to justify its (true and false) beliefs.
A justification should be available for any statement.  A statement such as A is a subclass of B may be justified by a simple reference to where the statement was asserted  (such as told in ontology I) or a more complex statement (such as ?A is stated to be a subclass of C in Ontology I and C is stated to be a subclass of B in Ontology I (and the subclass relationship is transitive?)). While explanations of all beliefs/deductions are important, some of the most important explanations are those that explain contradictions.  Additional support may be provided for this.  For example, if ontology I contains:
   A is disjoint from B
   C is the intersection of A and B
Then it is deducible that C contains no instances and is incoherent.
Explanation has been shown to be critical in user acceptance of systems[3].  If a user is expected to accept an answer from a system, the system should be able to be asked to justify its conclusions.  Similarly if a knowledge base is expected to be evolved, it is imperative for the person modifying the system to be able to query the system about consequences of anticipated and recent updates.  Many applications like customer relation management, help desk, configuration, and expert systems require explanation support in order to be used.
An extensive explanation facility has been designed and implemented for some description logic-based systems utilizing inference rule encodings and pruned, incremental presentations of inference rule applications [4,5]. Simplified versions of this can be done that explain common types of inference (but ignore more atypical and more complicated inferences) in order to get a quickly implementable and simpler explanation capability into a system. At a minimum, a system should be able to point to the source(s) of told information.
A listing of the inference rules for DAML+OIL is available in the axiomatic semantics document. [6]. 
Explainability may be more applicable to the "proof layer" of the Semantic Web than it is to WebOnt. Furthermore, it may have little impact on language design (although it is likely to have significant impact on application design).

C2. Internationalization
The language should support ontologies in multiple languages.

This seems to be covered by the interoperability requirement. Furthermore, character set issues are already handled by XML.  We should embrace whatever internationalization solutions are available from our building blocks of XML and RDF.

C3. Ontology Querying
There are a few possible definitions:
    -  Ability to ask questions about the logical structure of the ontology.
    -  Ability to ask questions about the information that follows from an ontology.
    -  A full fledged query language

This topic needs further clarification.  There is some overlap with ontology-based search that supports parametric search.  There is also some overlap with explainability. A full description of this topic would include all of the work on query languages for ontologies as well as including distinctions from ontology-based search and explanation.

C4. Tagging
It is sometimes useful to attach additional information to a piece of data. For example, the source of the data, a time stamp, or a confidence level. Tagging is the ability associate such information.

This requirement should probably be motivated by specific examples of needs for tagging.

C5. Proof Checking
Proofs can be described in a language and will be checkable. This might be considered a further step to C1: Explainability - which just explains its beliefs but does not include the requirement of a kind of automatic proof checking.

As with Explainability, this may be more appropriate for the "proof layer."

C6. Security 
Ability to specify who can view and modify information.  Ontologies should be able to specify access control information.

The Web typically doesn't allow update (except via file update) and viewing web pages is typically all or nothing.  Filtered viewing is not common.

C7. Trust
How to determine which information is reliable and/or believable. Must be able to know the sources of information and to express what supporting information is needed to believe something.

This is a larger issue which probably belongs in the "proof layer." We can probably reach consensus on the requirement for annotation of authoritative sources vs. other sources of information.

C8. Data Persistence
The Web is constantly changing, so it would be useful to know the lifetime of information. This will be useful for agents to know when they must refresh their knowledge bases.
If applied at the document level, this may be handled by the HTTP expires header. However, we may consider specifying this per fact in a data source, or for certain properties in an ontology.

[1] Laura Koetzle, Paul Hagen, Hillary Drohan, and Moira Dorsey.  "Smarter Sales of Complex Goods".   The Forrester Report. Cambridge, Mass. September 2001.

[2] Paul Hagen, David Weisman, Harley Manning, and Randy Souza.  "Guided Search
for eCommerce ".  The Forrester Report.  Cambridge, Mass. January, 1999.

[3] Deborah L. McGuinness and Jon Wright. ``An Industrial Strength Description Logic-based Configurator Platform''. IEEE Intelligent Systems, Vol. 13, No. 4, July/August 1998, pp. 69-77. 
[4] Deborah L. McGuinness. ``Explaining Reasoning in Description Logics''. Ph.D. Thesis, Rutgers University, 1996. Technical Report LCSR-TR-277. abstract and available from Rutgers Department of Computer Science Technical Report Series,.
[5] Alex Borgida, Enrico Franconi, Ian Horrocks, Deborah L. McGuinness, and Peter F. Patel-Schneider. ``Explaining ALC subsumption'' Proceedings of the International Workshop on Description Logics - DL-99, pp 33-36, Linköping, Sweden, July 1999. 

Received on Monday, 7 January 2002 11:18:58 UTC