## Notes on Semantic Structures

**Rationale**: There are several equivalent ways to define first-order semantic structures. The one we adopted has the advantage that it generalizes to rule sets with negation as failure (NAF) and to logics for dealing with uncertainty and inconsistency. The difficulty is that some popular theories for NAF, such as the *well-founded semantics*, are based on three-valued semantic structures. Some popular ways to handle uncertain or inconsistent information (which is certainly important in the Web environment) rely on four-valued and other multi-valued logics. Therefore, following M. Fitting, Fixpoint Semantics for Logic Programming A Survey, Theoretical Computer Science, 1999, we build our definitions to be compatible with future RIF dialects, which will be based on multivalued logics.

**Truth values**: Some RIF dialects will have additional truth values. For instance, some versions of NAF use three truth values: **t**, **f**, and **u** (undefined). Handling of contradictions and uncertainty requires at least four truth values: **t**, **u**, **f**, and **i** (inconsistent).

**Ordering truth values**: In the well-founded semantics for NAF, **f** <_{t} **u** <_{t} **t**, and it is again a total order. But in four-valued logics, which are often used for dealing with uncertain or inconsistent information, the truth order is partial: **f** <_{t} **u** <_{t} **t** and **f** <_{t} **i** <_{t} **t**. Such logics also have another partial order, called the * knowledge order* <

_{k}:

**u**<

_{k}

**t**<

_{k}

**i**; and

**u**<

_{k}

**f**<

_{k}

**i**. Under the knowledge order, true and false are incomparable, and facts that are both true and false receive the truth value

**i**, which is the least upper bound of

**f**and

**t**in the knowledge order.

**Intended models for rule sets**: The notion of a model is only the basic ingredient in the definition of a semantics of a rule set. In general, a **semantics of a rule set****R** is the set of its * intended models* (see Y. Shoham. Nonmonotonic logics: meaning and utility. In: Proc. 10th International Joint Conference on Artificial Intelligence, Morgan Kaufmann, pp. 388--393, 1987). There are different theories of what the intended sets of models are supposed to look like depending on the features of the particular rule sets.

For the basic logic dialect, which is based on Horn rules, the set of intended models of **R** is commonly agreed upon: it is the unique minimal model of **R**.

However, when rule bodies contain literals negated with the negation-as-failure connective `naf`, then only *some* of the minimal models of a rule set are viewed as intended. This issue will be addressed in the appropriate dialects of RIF. The two most common theories of intended models are based on the so called * well-founded models* and

*. Here we will just illustrate the problem with a simple example.*

**stable models**Suppose **R** consists of a single rule `p :- naf q`. If `naf` were interpreted as classical negation, `not`, then this rule would be simply equivalent to `p \/ q`, and so it would have two kinds of models: one in which `p` is true and one where `q` is true. In contrast to first-order logic, most rule-based systems do not consider `p` and `q` symmetrically. Instead, they view the rule `p :- naf q` as a statement that `p` must be true if it is not possible to establish the truth of `q`. Since it is, indeed, impossible to establish the truth of `q`, such theories will derive `p` even though it does not logically follow from `p \/ q`. The logic underlying rule-based systems also assumes that only the *minimal* models are intended (minimality here is with respect to the set of true facts). Therefore, the intended models of the above rule set **R** must have the property that not only `p` is true but also that `q` is false.