Adding Numbers

Group Members

Description

"Adding numbers to this stuff"

We are interested in finding tools that enable the users to quantify a certain relationship in DAML. That is, instead of a relationship being "carved in stone", we want to express a degree of certainty for the truth of this relationship.

In order to do that, we need the tools to support two things:

  1. Allowing / Enabling degrees of belief/trust for a given relationship.
  2. Using the values we assigned in step 1, in order to make inferences and draw conclusions.

Motivation

By every indication Semantic Web will not be a "safe place". So it will be an important task for any reasoner to assess how trustable and reliable its sources of information will be. Let us suppose that we have ways to determine whether or not to trust a source of information (or service provider, or even a specific piece of knowledge described in an ontology). Still we may not be sure if it is relevant.

If we look at all (or a great part of) the human knowledge stored on the Web we will realize that each fact may have different degrees of relevance/significance, depending on the specific domain of interest of the agent that looks at this knowledge base.

Also, for the same source of information S, each different individual/company/agent may assign different values to the degree it prefers S/trusts S/takes S as credible. Even more, if an agent tries to automatically compose a service for some individual, it has to take into account this persons' preferences/personal view of the world. For example I might prefer company X over company Y and so buy my tickets from X even if Y has lower prices and this might be based on my personal experience. Or I prefer listening to the news channel C although channel D may be generally considered more credible. Of course people always take other people's opinion into account and build their model of the world based on subjective conclusion other individuals have drawn.

When quantifying relationships, we are able to express a certain level of belief or trust. These can be conceived as two orthogonal concepts. Besides, sometimes personal preferences must be taken into account.

The tools we reviewed provide only partial support for the aforementioned requirements.

Applications

Adding levels of reliability/credibility is going to make our reasoning even more demanding in terms of time and computational resources needed. We therefore need to discuss the advantages of this approach.

There are two basic reasons for trying to use an epistemic representation by DAML and a lot of applications that elaborate on the fact.

One reason for using different levels of belief is that it is allows us to express the problem in a very natural way. Although at first this may seem unimportant, we will see that this is not the case. In fact, Semantic Web reasoners and agents will have to deal with solving real, complicated problems. In order to express such problems, especially when users preferences are involved, soft constraints are likely to be introduced in one form or the other because it is natural that people will have conflicting requirements For example, a query might be "find me the best ticket at the best value", where the meaning of the "best ticket" might not be the same for all people. Actually, each person might have a distinct version of what "best" means, and his agent should try to come up with a decent solution even though it has to deal with conflicting and fuzzy requirements.

The above example illustrates two facts:

-The preferences a user might have are unclear and so a fuzzy description comes as a good and natural way to represent them.

-Our world is constantly changing and therefore any agent or system should be able to adapt. One way to achieve this is by flexibly adjusting its knowledge and set of inference rules.

Another example of the possible usage of epistemic logic can be in describing relations between corporations. When a group of companies cooperates for a project, partners may be physically located in a separate geographic region, sometimes even in a different country. It is natural for them to communicate through the web. The result of the project is normally vast amounts of information in the form of RFCs, reports, and other documents. A good solution is to use DAML to annotate this knowledge and insure a common understanding of the problem domain.

Depending on the nature of the project different requirements may arise. For example each partner may want to be able to specify how relevant to his part of the project the work of another group is. Or, when things need be double-checked, in what degree this has been completed and how much should the result be trusted. This applies not only in technical projects but in any case a large sum of information is produced by different sources. So for example it might be the case that an operator provides access to third party services , but with modified level of belief.

Tools and/or resources available on the web

Trellis

[WWW]http://excalibur.isi.edu:8888/newTrellis/index.html

Basic Idea

Trellis is a tool that can be used to help users annotate their reasoning and characterize the reliability and credibility of different sources of information. At the same time, it records all the users statements by asserting them in DAML format. Later it uses the acquired semantic knowledge in order to assess the reliability and credibility of each source of information.

This system suggests a set of very interesting ideas. It helps the user organize his points, while at the same time "trains" the system / agent on how it should act. A further example in this direction could be a system providing manual service composing to users that would at the same time try to learn their preferences so that it can propose default values or even eventually automatically create personalized services upon request.

Trellis uses a basic structure of the type :

is {not} likelihood qualifier because according to source-description which is

Basically this signifies that it attaches one reliability-qualifier and one credibility-qualifier to each statement.

We used Trellis to annotate some of the Nature articles and assess the credibility of the facts they stated.

The result is demonstrated below.

Critique

Trellis allows us to quantify only the reliability/credibility of the sources, whereas we will like to be able to quantify each DAML-expressible relationship. We would like to quantify the credibility of an ontology rule, thus being able to express our preference for a service over another.

We would like to make a query like "to what extent X relates to Y?", whereas Trellis only allows us to ask "what is the credibility of source Y, if X1 is based on Y, X2 is based on Y, etc."

The main point is that Trellis uses DAML only as an annotation language, not exploiting its semantic power.

Even so, it's reasoning about the credibility / reliability of a source is not described using DAML, as someone would expect but is provided by an external ad-hoc mathematical formula.

Annotea

[WWW]http://www.w3.org/2001/Annotea/

Is a tool that enables users to annotate URIs or positions within documents, by associating metadata to documents or specific parts of a document. The annotations are described with RDF schema, and consist of resource/property/value triples.

Annotation is actually a generic base class and instances of it (Change, Comment, Explanation) are used to associate a particular meaning to a value. Annotations are stored in RDF databases on specialized servers and are retrieved by the browser, at the time the user browses the annotated document.

Critique

We can use the general Annotea backend mechanism to store belief/trust properties as values, but Annotea doesn't allow us to make any inference, even though a network of related URIs exists.

It does not make use of the DAML semantic. Actually its only relation to DAML is that both use RDF as a "backend".

P-SHOQ(D)

P-SHOQ(D) is a description logic coined up to support probabilistic ontologies and probabilistic default reasoning on the Semantic Web.

SHOQ(D) is an expressive description logic supporting named individuals and concrete datatypes which has the same descriptive power as DAML+OIL, but it does not support inverse roles in order to avoid the very high complexity that results from the unconstrained interaction of inverse roles with nominals and datatypes. So it makes descriptions more suitable for automated reasoning.

P-SHOQ(D) augments SHOQ(D) with the ability to express probabilistic knowledge about concepts and roles concepts or role instances. Yet, reasoning in P-SHOQ(D) is proved to be decidable.

Fuzzy Description Logics

These are fuzzy extensions to DL, enabling fuzzy assertions and fuzzy terminological axioms

Within this formalism we can say that an rdf:description is rdf:about an URI within a degree of certitude. That is, an ontology referring to a BigBall concept could be 0.9 about a BasketBall, 0.6 about SoccerBall and 0.01 about TennisBall.

Entailment in FDL is proved to be PSPACE-complete.

To the best of our knowledge there are no tools for inference in either P-SHOQ(D) and Fuzzy Description Logics, but we would like to adopt some theoretical concepts they introduce in our own tool, that we will develop for assignment 3.

On the Nature Dataset

We created an example showing how Trellis helps users reason about facts stated in Nature articles

Further Links

[FrontPage] [TitleIndex] [WordIndex]