W3C

RIF Production Rule Dialect

W3C Recommendation 22 June 2010

This version:
http://www.w3.org/TR/2010/REC-rif-prd-20100622/
Latest version:
http://www.w3.org/TR/rif-prd/
Previous version:
http://www.w3.org/TR/2010/PR-rif-prd-20100511/ (color-coded diff)
Editors:
Christian de Sainte Marie, IBM/ILOG
Gary Hallmark, Oracle
Adrian Paschke, Freie Universitaet Berlin

Please refer to the errata for this document, which may include some normative corrections.

This document is also available in these non-normative formats: PDF version.

See also translations.


Abstract

This document, developed by the Rule Interchange Format (RIF) Working Group, specifies the production rule dialect of the W3C rule interchange format (RIF-PRD), a standard XML serialization format for production rule languages.

Status of this Document

May Be Superseded

This section describes the status of this document at the time of its publication. Other documents may supersede this document. A list of current W3C publications and the latest revision of this technical report can be found in the W3C technical reports index at http://www.w3.org/TR/.

This document is being published as one of a set of 11 documents:

  1. RIF Overview
  2. RIF Core Dialect
  3. RIF Basic Logic Dialect
  4. RIF Production Rule Dialect (this document)
  5. RIF Framework for Logic Dialects
  6. RIF Datatypes and Built-Ins 1.0
  7. RIF RDF and OWL Compatibility
  8. OWL 2 RL in RIF
  9. RIF Combination with XML data
  10. RIF In RDF
  11. RIF Test Cases

XML Schema Datatypes Dependency

RIF is defined to use datatypes defined in the XML Schema Definition Language (XSD). As of this writing, the latest W3C Recommendation for XSD is version 1.0, with version 1.1 progressing toward Recommendation. RIF has been designed to take advantage of the new datatypes and clearer explanations available in XSD 1.1, but for now those advantages are being partially put on hold. Specifically, until XSD 1.1 becomes a W3C Recommendation, the elements of RIF which are based on it should be considered optional, as detailed in Datatypes and Builtins, section 2.3. Upon the publication of XSD 1.1 as a W3C Recommendation, those elements will cease to be optional and are to be considered required as otherwise specified.

We suggest that for now developers and users follow the XSD 1.1 Last Call Working Draft. Based on discussions between the Schema, RIF and OWL Working Groups, we do not expect any implementation changes will be necessary as XSD 1.1 advances to Recommendation.

Document Unchanged

There have been no changes to the body of this document since the previous version. For details on earlier changes, see the change log.

Please Send Comments

Please send any comments to public-rif-comments@w3.org (public archive). Although work on this document by the Rule Interchange Format (RIF) Working Group is complete, comments may be addressed in the errata or in future revisions. Open discussion among developers is welcome at public-rif-dev@w3.org (public archive).

Endorsed By W3C

This document has been reviewed by W3C Members, by software developers, and by other W3C groups and interested parties, and is endorsed by the Director as a W3C Recommendation. It is a stable document and may be used as reference material or cited from another document. W3C's role in making the Recommendation is to draw attention to the specification and to promote its widespread deployment. This enhances the functionality and interoperability of the Web.

Patents

This document was produced by a group operating under the 5 February 2004 W3C Patent Policy. W3C maintains a public list of any patent disclosures made in connection with the deliverables of the group; that page also includes instructions for disclosing a patent.


Table of Contents

1 Overview

This document specifies the production rule dialect of the W3C rule interchange format (RIF-PRD), a standard XML serialization format for production rule languages.

The production rule dialect is one of a set of rule interchange dialects that also includes the RIF Core dialect ([RIF-Core]) and the RIF basic logic dialect ([RIF-BLD]).

RIF-Core, the core dialect of the W3C rule interchange format, is designed to support the interchange of definite Horn rules without function symbols ("Datalog"). RIF-Core has both a standard first-order semantics and an operational semantics. Syntactically, RIF-Core has a number of extensions of Datalog:

RIF-Core is based on a rich set of datatypes and built-ins that are aligned with Web-aware rule system implementations [RIF-DTB]. In addition, the RIF RDF and OWL Compatibility document [RIF-RDF-OWL] specifies the syntax and semantics of combinations of RIF-Core, RDF, and OWL documents.

RIF-Core is intended to be the common core of all RIF dialects, and it has been designed, in particular, to be a useful common subset of RIF-BLD and RIF-PRD. RIF-PRD includes and extends RIF-Core, and, therefore, RIF-PRD inherits all RIF-Core features. These features make RIF-PRD a Web-aware (even a semantic Web-aware) language. However, it should be kept in mind that RIF is designed to enable interoperability among rule languages in general, and its uses are not limited to the Web.

This document targets designers and developers of RIF-PRD implementations. A RIF-PRD implementation is a software application that serializes production rules as RIF-PRD XML (producer application) and/or that deserializes RIF-PRD XML documents into production rules (consumer application).

1.1 Production rule interchange

Production rules have an if part, or condition, and a then part, or action. The condition is like the condition part of logic rules (as covered by RIF-Core and its basic logic dialect extension, RIF-BLD). The then part contains actions. An action can assert facts, modify facts, retract facts, and have other side-effects. In general, an action is different from the conclusion of a logic rule, which contains only a logical statement. However, the conclusion of rules interchanged using RIF-Core can be interpreted, according to RIF-PRD operational semantics, as actions that assert facts in the knowledge base.

Example 1.1. The following are examples of production rules:

Because RIF-PRD is a production rule interchange format, it specifies an abstract syntax that shares features with concrete production rule languages, and it associates the abstract constructs with normative semantics and a normative XML concrete syntax. Annotations (e.g. rule author) are the only constructs in RIF-PRD without a formal semantics.

The abstract syntax is specified in mathematical English, and the abstract syntactic constructs that are defined in the sections Abstract Syntax of Conditions, Abstract Syntax of Actions and Abstract Syntax of Rules and Rulesets, are mapped into the concrete XML constructs in the section XML syntax. A lightweight notation is used, instead of the XML syntax, to tie the abstract syntax to the specification of the semantics. A more complete presentation syntax is specified using an EBNF in Presentation Syntax. However, only the XML syntax and the associated semantics are normative. The normative XML schema is included in Appendix: XML Schema.

Example 1.2. In RIF-PRD presentation syntax, the first rule in example 1.1. can be represented as follows:

Prefix(ex <http://example.com/2008/prd1#>)
(* ex:rule_1 *)
Forall ?customer ?purchasesYTD (
 If   And( ?customer#ex:Customer
           ?customer[ex:purchasesYTD->?purchasesYTD]
           External(pred:numeric-greater-than(?purchasesYTD 5000)) )
 Then Do( Modify(?customer[ex:status->"Gold"]) ) )

Production rules are statements of programming logic that specify the execution of one or more actions when their conditions are satisfied. Production rules have an operational semantics, that the OMG Production Rule Representation specification [OMG-PRR] summarizes as follows:

  1. Match: the rules are instantiated based on the definition of the rule conditions and the current state of the data source;
  2. Conflict resolution: a decision algorithm, often called the conflict resolution strategy, is applied to select which rule instance will be executed;
  3. Act: the state of the data source is changed, by executing the selected rule instance's actions. If a terminal state has not been reached, the control loops back to the first step (Match).

In the section Operational semantics of rules and rule sets, the semantics for rules and rule sets is specified, accordingly, as a labeled terminal transition system (PLO04), where state transitions result from executing the action part of instantiated rules. When several rules are found to be executable at the same time, during the rule execution process, a conflict resolution strategy is used to select the rule to execute. The section Conflict resolution specifies how a conflict resolution strategy can be attached to a rule set. RIF-PRD defines a default conflict resolution strategy.

In the section Semantics of condition formulas, the semantics of the condition part of rules in RIF-PRD is specified operationally, in terms of matching substitutions. To emphasize the overlap between the rule conditions of RIF-BLD and RIF-PRD, and to share the same RIF definitions for datatypes and built-ins [RIF-DTB], an alternative, and equivalent, specification of the semantics of rule conditions in RIF-PRD, using a model theory, is provided in the appendix Model-theoretic semantics of RIF-PRD condition formulas.

The semantics of condition formulas and the semantics of rules and rule sets make no assumption regarding how condition formulas are evaluated. In particular, they do not require that condition formula be evaluated using pattern matching. However, RIF-PRD conformance, as defined in the section Conformance and interoperability, requires only support for safe rules, that is, forward-chaining rules where the conditions can be evaluated based on pattern matching only.

In the section Operational semantics of actions, the semantics of the action part of rules in RIF-PRD is specified using a transition relation between successive states of the data source, represented by ground condition formulas, thus making the link between the model-theoretic semantics of conditions and the operational semantics of rules and rule sets.

The abstract syntax of RIF-PRD documents, and the semantics of the combination of multiple RIF-PRD documents, is specified in the section Document and imports.

In addition to externally specified functions and predicates, and in particular, in addition to the functions and predicates built-ins defined in [RIF-DTB], RIF-PRD supports externally specified actions, and defines one action built-in, as specified in the section Built-in functions, predicates and actions.


1.2 Running example

The same example rules will be used throughout the document to illustrate the syntax and the semantics of RIF-PRD.

The rules are about the status of customers at a shop, and the discount awarded to them. The rule set contains four rules, to be applied when a customer checks out:

  1. Gold rule: A "Silver" customer with a shopping cart worth at least $2,000 is awarded the "Gold" status.
  2. Discount rule: "Silver" and "Gold" customers are awarded a 5% discount on the total worth of their shopping cart.
  3. New customer and widget rule: A "New" customer who buys a widget is awarded a 10% discount on the total worth of her shopping cart, but she looses any voucher she may have been awarded.
  4. Unknown status rule: A message must be printed, identifying any customer whose status is unknown (that is, neither "New", "Bronze", "Silver" or "Gold"), and the customer must be assigned the status: "New".

The Gold rule must be applied first; that is, e.g., a customer with "Silver" status and a shopping cart worth exactly $2,000 should be promoted to "Gold" status, before being given the 5% discount that would disallow the application of the Gold rule (since the total worth of his shopping cart would then be only $1,900).

In the remainder of this document, the prefix ex1 stands for the fictitious namespace of this example: http://example.com/2009/prd2#.

2 Conditions

This section specifies the syntax and semantics of the condition language of RIF-PRD.

The RIF-PRD condition language specification depends on Section Constants, Symbol Spaces, and Datatypes, in the RIF data types and builtins specification [RIF-DTB].

2.1 Abstract syntax

The alphabet of the RIF-PRD condition language consists of:

For the sake of readability and simplicity, this specification introduces a notation for these constructs. The notation is not intended to be a concrete syntax, so it leaves out many details. The only concrete syntax for RIF-PRD is the XML syntax.

RIF-PRD supports externally defined functions only (including the built-in functions specified in [RIF-DTB]). RIF-PRD, unlike RIF-BLD, does not support uninterpreted function symbols (sometimes called logically defined functions).

RIF-PRD supports a form of negation. Neither RIF-Core nor RIF-BLD support negation, because logic rule languages use many different and incompatible kinds of negation. See also the RIF framework for logic dialects [RIF-FLD].

2.1.1 Terms

The most basic construct in the RIF-PRD condition language is the term. RIF-PRD defines several kinds of term: constants, variables, lists and positional terms.

Definition (Term).

  1. Constants and variables. If tConst or tVar then t is a simple term;
  2. List terms. A list has the form List(t1 ... tm), where m≥0 and t1, ..., tm are ground terms, i.e. without variables. A list of the form List() (i.e., a list in which m=0) is called the empty list;
  3. Positional terms. If tConst and t1, ..., tn, n≥0, are terms then t(t1 ... tn) is a positional term.
    Here, the constant t represents a function and t1, ..., tn represent argument values.   ☐

To emphasize interoperability with RIF-BLD, positional terms may also be written: External(t(t1...tn)).

Example 2.1.

2.1.2 Atomic formulas

Atomic formulas are the basic tests of the RIF-PRD condition language.

Definition (Atomic formula). An atomic formula can have several different forms and is defined as follows:

  1. Positional atomic formulas. If tConst and t1, ..., tn, n≥0, are terms then t(t1 ... tn) is a positional atomic formula (or simply an atom)
  2. Equality atomic formulas. t = s is an equality atomic formula (or simply an equality), if t and s are terms
  3. Class membership atomic formulas. t#s is a membership atomic formula (or simply membership) if t and s are terms. The term t is the object and the term s is the class
  4. Subclass atomic formulas. t##s is a subclass atomic formula (or simply a subclass) if t and s are terms
  5. Frame atomic formulas. t[p1->v1 ... pn->vn] is a frame atomic formula (or simply a frame) if t, p1, ..., pn, v1, ..., vn, n ≥ 0, are terms. The term t is the object of the frame; the pi are the property or attribute names; and the vi are the property or attribute values. In this document, an attribute/value pair is sometimes called a slot
  6. Externally defined atomic formulas. If t is a positional atomic formula then External(t) is an externally defined atomic formula.   ☐

Class membership, subclass, and frame atomic formulas are used to represent classifications, class hierarchies and object-attribute-value relations.

Externally defined atomic formulas are used, in particular, for representing built-in predicates.

In the RIF-BLD specification, as is common practice in logic languages, atomic formulas are also called terms.

Example 2.2.

2.1.3 Formulas

Composite truth-valued constructs are called formulas, in RIF-PRD.

Note that terms (constants, variables, lists and functions) are not formulas.

More general formulas are constructed out of atomic formulas with the help of logical connectives.

Definition (Condition formula). A condition formula can have several different forms and is defined as follows:

  1. Atomic formula: If φ is an atomic formula then it is also a condition formula.
  2. Conjunction: If φ1, ..., φn, n ≥ 0, are condition formulas then so is And(φ1 ... φn), called a conjunctive formula. As a special case, And() is allowed and is treated as a tautology, i.e., a formula that is always true.
  3. Disjunction: If φ1, ..., φn, n ≥ 0, are condition formulas then so is Or(φ1 ... φn), called a disjunctive formula. As a special case, Or() is permitted and is treated as a contradiction, i.e., a formula that is always false.
  4. Negation: If φ is a condition formula, then so is Not(φ), called a negative formula.
  5. Existentials: If φ is a condition formula and ?V1, ..., ?Vn, n>0, are variables then Exists ?V1 ... ?Vn(φ) is an existential formula.   ☐

In the definition of a formula, the component formulas φ and φi are said to be subformulas of the respective condition formulas that are built using these components.

Example 2.3.

And( ?customer # ex1:Customer
     ?customer[ex1:status->"New"]
     Exists ?shoppingCart ?item ( And ( ?customer[ex1:shoppingCart->?shoppingCart]
                                        ?shoppingCart[ex1:containsItem->?item]
                                        ?item # ex1:Widget) )
                                )
    )

The function Var, that maps a term, an atomic formula or a condition formula to the set of its free variables is defined as follows:

Definition (Ground formula). A condition formula φ is a ground formula if and only if Varφ = {} and φ does not contain any existential subformula.   ☐

In other words, a ground formula does not contain any variable term.

2.1.4 Well-formed formulas

Not all formulas are well-formed in RIF-PRD: it is required that no constant appear in more than one context. What this means precisely is explained below.

The set of all constant symbols, Const, is partitioned into the following subsets:

As seen from the following definitions, these subsets are not specified explicitly but, rather, are inferred from the occurrences of the symbols.


Definition (Context of a symbol). The context of an occurrence of a symbol, s∈Const, in a formula, φ, is determined as follows:

Definition (Well-formed formula). A formula φ is well-formed iff:

Definition (RIF-PRD condition language). The RIF-PRD condition language consists of the set of all well-formed formulas.   ☐

2.2 Operational semantics of condition formulas

This section specifies the semantics of the condition formulas in a RIF-PRD document.

Informally, a condition formula is evaluated with respect to a state of facts and it is satisfied, or true, if and only if:

The semantics is specified in terms of matching substitutions in the sections below. The specification makes no assumption regarding how matching substitutions are determined. In particular, it does not require from well-formed condition formulas that they can be evaluated using pattern matching only. However, RIF-PRD requires safeness from well-formed rules, which implies that all the variables in the left-hand side can be bound by pattern matching.

For compatibility with other RIF specifications (in particular, RIF data types and built-ins [RIF-DTB] and RIF RDF and OWL compatibility [RIF-RDF-OWL]), and to make explicit the interoperability with RIF logic dialects (in particular RIF Core [RIF-Core] and RIF-BLD [RIF-BLD]), the semantics of RIF-PRD condition formulas is also specified using model theory, in appendix Model theoretic semantics of RIF-PRD condition formulas.

The two specifications are equivalent and normative.

2.2.1 Matching substitution

Let Term be the set of the terms in the RIF-PRD condition language (as defined in section Terms).

Definition (Substitution). A substitution is a finitely non-identical assignment of terms to variables; i.e., a function σ from Var to Term such that the set {xVar | x ≠ σ(x)} is finite. This set is called the domain of σ and denoted by Dom(σ). Such a substitution is also written as a set such as σ = {ti/xi}i=1..n where Dom(σ) = {xi}i=1..n and σ(xi) = ti, i = 1..n.   ☐

Definition (Ground Substitution). A ground substitution is a substitution σ that assigns only ground terms to the variables in Dom(σ): ∀ xDom(σ), Var(σ(x)) = ∅   ☐

Because RIF-PRD covers only externally defined interpreted functions, a ground functional term can always be replaced by its value. As a consequence, a ground substitution can always be restricted, without loss of generality, to assign only constants to the variable in its domain. In the remainder of this document, it will always be assumed that a ground substitution assigns only constants to the variables in its domain.

If t is a term or a condition formula, and if σ is a ground substitution such that Var(t) ∈ Dom(σ), σ(t) denotes the ground term or the ground condition formula obtained by substituting, in t:

Definition (Matching substitution). Let ψ be a RIF-PRD condition formula; let σ be a ground substitution such that Var(ψ) ⊆ Dom(σ); and let Φ be a set of ground RIF-PRD atomic formulas.

We say that the ground substitution σ matches ψ to Φ if and only if one of the following is true:

2.2.2 Condition satisfaction

We define, now, what it means for a state of the fact base to satisfy a condition formula. The satisfaction of condition formulas in a state of the fact base provides formal underpinning to the operational semantics of rule sets interchanged using RIF-PRD.

Definition (State of the fact base). A state of the fact base, wΦ, is associated to every set of ground atomic formulas, Φ, that satisfies, at least, the following conditions, for all triple of constants c1, c2, c3:

We say that wΦ is represented by Φ; or, equivalently, by the conjunction of all the ground atomic formulas in Φ.   ☐

Each ground atomic formula in Φ represents a single fact, and, often, the ground atomic formulas, themselves, are called facts, as well.

Definition (Condition satisfaction). A RIF-PRD condition formula ψ is satisfied in a state of the fact base, w, if and only if w is represented by a set of ground atomic formulas Φ, and there is a ground substitution σ that matches ψ to Φ.   ☐

Alternative, but equivalent, definitions of a state of the fact base and of the satisfaction of a condition are given in the appendix Model theoretic semantics of RIF-PRD condition formulas: they provide the formal link between the model theory of RIF-PRD condition formulas and the operational semantics of RIF-PRD documents.

3 Actions

This section specifies the syntax and semantics of the RIF-PRD action language. The conclusion of a production rule is often called the action part, the then part, or the right-hand side, or RHS.

The RIF-PRD action language is used to add, delete and modify facts in the fact base. As a rule interchange format, RIF-PRD does not make any assumption regarding the nature of the data sources that the producer or the consumer of a RIF-PRD document uses (e.g. a rule engine's working memory, an external data base, etc). As a consequence, the syntax of the actions that RIF-PRD supports are defined with respect to the RIF-PRD condition formulas that represent the facts that the actions affect. In the same way, the semantics of the actions is specified in terms of how their execution affects the evaluation of rule conditions.

3.1 Abstract syntax

The alphabet of the RIF-PRD action language includes symbols to denote:

3.1.1 Actions

The RIF-PRD action language includes constructs for actions that are atomic, from a transactional point of view, and constructs that represent compounds of atomic actions. Action constructs take constructs from the RIF-PRD condition language as their arguments.

Definition (Atomic action). An atomic action is a construct that represents an atomic transaction. An atomic action can have several different forms and is defined as follows:

  1. Assert fact: If φ is a positional atom, a frame or a membership atomic formula in the RIF-PRD condition language, then Assert(φ) is an atomic action. φ is called the target of the action.
  2. Retract fact: If φ is a positional atom or a frame in the RIF-PRD condition language, then Retract(φ) is an atomic action. φ is called the target of the action.
  3. Retract all slot values: If o and s are terms in the RIF-PRD condition language, then Retract(o s) is an atomic action. The pair (o, s) is called the target of the action.
  4. Retract object: If t is a term in the RIF-PRD condition language, then Retract(t) is an atomic action. t is called the target of the action.
  5. Execute: if φ is a positional atom in the RIF-PRD condition language, then Execute(φ) is an atomic action. φ is called the target of the action.   ☐

Definition (Compound action). A compound action is a construct that can be replaced equivalently by a pre-defined, and fixed, sequence of atomic actions. RIF-PRD defines a single compound action, defined as follows:

  1. Modify fact: if φ is a frame in the RIF-PRD condition language: φ = o[s->v]; then Modify(φ) is a compound action, defined by the sequence: Retract(o s), followed by Assert(φ). φ is called the target of the action.   ☐

Definition (Action). A action is either an atomic action or a compound action.   ☐

Definition (Ground action). An action with target t is a ground action if and only if

  ☐

Example 3.1.

3.1.2 Action blocks

The action block is the top level construct to represent the conclusions of RIF-PRD rules. An action block contains a non-empty sequence of actions. It may also include action variable declarations.

The action variable declaration construct is used to declare variables that are local to the action block, called action variables, and to assign them a value within the action block.

Definition (Action variable declaration). An action variable declaration is a pair, (v p) made of an action variable, v, and an action variable binding (or, simply, binding), p, where p has one of two forms:

  1. frame object declaration: if the action variable, v, is to be assigned the identifier of a new frame, then the action variable binding is a frame object declaration: New(). In that case, the notation for the action variable declaration is: (?o New());
  2. frame slot value: if the action variable, v, is to be assigned the value of a slot of a ground frame, then the action variable binding is a frame: p = o[s->v], where o is a term that represents the identifier of the ground frame and s is a term that represents the name of the slot. The associated notation is: (?value o[s->?value]).   ☐

Definition (Action block). If (v1 p1), ..., (vn pn), n ≥ 0, are action variable declarations, and if a1, ..., am, m ≥ 1, are actions, then Do((v1 p1) ... (vn pn) a1 ... am) denotes an action block.   ☐

Example 3.2. In the following action block, a local variable ?oldValue is bound to a value of the attribute value of the object bound to the variable ?shoppingCart. The ?oldValue is then used to compute a new value, and the Modify action is used to overwrite the old value with the new value in the fact base:

Do( (?oldValue ?shoppingCart[ex1:value->?oldValue])
    Modify( ?shoppingCart[ex1:value->func:numeric-multiply(?oldValue 0.90)] ) )

3.1.3 Well-formed action blocks

Not all action blocks are well-formed in RIF-PRD:

The notion of well-formedness, already defined for condition formulas, is extended to actions, action variable declarations and action blocks.

Definition (Well-formed action). An action α is well-formed if and only if one of the following is true:

Definition (Well-formed action variable declaration). An action variable declaration (?v p) is well-formed if and only if one of the following is true:

For the definition of a well-formed action block, the function Var(f), that has been defined for condition formulas, is extended to actions and frame object declarations as follows:

Definition (Well-formed action block). An action block is well-formed if and only if all of the following are true:

Definition (RIF-PRD action language). The RIF-PRD action language consists of the set of all the well-formed action blocks.   ☐

3.2 Operational semantics of atomic actions

This section specifies the semantics of the atomic actions in a RIF-PRD document.

The effect of the ground atomic actions in the RIF-PRD action language is to modify the state of the fact base, in such a way that it changes the set of conditions that are satisfied before and after each atomic action is performed.

As a consequence, the semantics of the ground atomic actions in the RIF-PRD action language determines a relation, called the RIF-PRD transition relation: →RIF-PRDW × L × W, where W denotes the set of all the states of the fact base, and where L denotes the set of all the ground atomic actions in the RIF-PRD action language.

The semantics of a compound action follows directly from the semantics of the atomic actions that compose it.

Individual states of the fact base are represented by sets of ground atomic formulas (Section Satisfaction of a condition). In the following, the operational semantics of RIF-PRD actions, rules, and rule sets is specified by describing the changes they induce in the fact base.

Definition (RIF-PRD transition relation). The semantics of RIF-PRD atomic actions is specified by the transition relation →RIF-PRDW × L × W. (w, α, w') ∈ →RIF-PRD if and only if wW, w' W, α is a ground atomic action, and one of the following is true:

  1. α is Assert(φ), where φ is a ground atomic formula, and w' = w ∪ {φ};
  2. α is Retract(φ), where φ is a ground atomic formula, and w' = w \ {φ};
  3. α is Retract(o s), where o and s are constants, and w' = (w \ {o[s->v] | for all the values of v});
  4. α is Retract(o), where o is a constant, and w' = w \ {o[s->v] | for all the values of terms s and v} - {o#c | for all the values of term c};
  5. α is Execute(φ), where φ is a ground atomic builtin action, and w' = w.   ☐

Rule 1 says that all the condition formulas that were satisfied before an assertion will be satisfied after, and that, in addition, the condition formulas that are satisfied by the asserted ground formula will be satisfied after the assertion. No other condition formula will be satisfied after the execution of the action.

Rule 2 says that all the condition formulas that were satisfied before a retraction will be satisfied after, except if they are satisfied only by the retracted fact. No other condition formula will be satisfied after the execution of the action.

Rule 3 says that all the condition formulas that were satisfied before the retraction of all the values of a given slot of a given object will be satisfied after, except if they are satisfied only by one of the frame formulas about the object and the slot that are the target of the action, or a conjunction of such formulas. No other condition formula will be satisfied after the execution of the action.

Rule 4 says that all the condition formulas that were satisfied before the removal of a frame object will be satisfied after, except if they are satisfied only by one of the frame or membership formulas about the removed object or a conjunction of such formulas. No other condition formula will be satisfied after the execution of the action.

Rule 5 says that all the condition formulas that were satisfied before the execution of an action builtin will be satisfied after. No other condition formula will be satisfied after the execution of the action.

Example 3.3. Assume an initial state of the fact base that is represented by the following set, w0, of ground atomic formulas, where _c1, _v1 and _s1 denote individuals and where ex1:Customer, ex1:Voucher and ex1:ShoppingCart represent classes:

Initial state:

  1. Assert( _c1[ex1:status->"New"] ) denotes an atomic action that adds to the fact base, a fact that is represented by the ground atomic formula: _c1[ex1:status->"New"]. After the action is executed, the new state of the fact base is represented by
    • w1 = {_c1#ex1:Customer _v1#ex1:Voucher _s1#ex1:ShoppingCart _c1[ex1:voucher->_v1] _c1[ex1:shoppingCart->_s1] _v1[ex1:value->5] _s1[ex1:value->500] _c1[ex1:status->"New"]}
  2. Retract( _c1[ex1:voucher->_v1] ) denotes an atomic action that removes from the fact base, the fact that is represented by the ground atomic formula _c1[ex1:voucher->_v1]. After the action, the new state of the fact base is represenetd by:
    • w2 = {_c1#ex1:Customer _v1#ex1:Voucher _s1#ex1:ShoppingCart _c1[ex1:shoppingCart->_s1] _v1[ex1:value->5] _s1[ex1:value->500] _c1[ex1:status->"New"]}
  3. Retract( _v1 ) denotes an atomic action that removes the individual denoted by the constant _v1 from the fact base. All the class membership and the object-attribute-value facts where _v1 is the object are removed. After the action, the new state of the fact base is represenetd by:
    • w3 = {_c1#ex1:Customer _s1#ex1:ShoppingCart _c1[ex1:shoppingCart->_s1] _s1[ex1:value->500] _c1[ex1:status->"New"]}
  4. Retract( _s1 ex1:value ) denotes an atomic action that removes all the object-attribute-value facts that assign a ex1:value to the ex1:ShoppingCart _s1. After the action, the new state of the fact base is represented by
    • w4 = {_c1#ex1:Customer _s1#ex1:ShoppingCart _c1[ex1:shoppingCart->_s1] _c1[ex1:status->"New"]}
  5. Assert( _s1[ex1:value->450] ) adds in the fact base_the single fact that is represented by the ground frame: <tt>_s1[ex1:value->450]. After the action, the new state of the fact base is represented by:
    • w5 = {_c1#ex1:Customer _s1#ex1:ShoppingCart _c1[ex1:shoppingCart->_s1] _s1[ex1:value->450] _c1[ex1:status->"New"]}
  6. Execute( act:print(func:concat("New customer: " _c1)) ) denotes an action that does not impact the state of the fact base, but that prints a string to an output stream. After the action, the new state of the fact base is represented by:
    • w6 = w5 = {_c1#ex1:Customer _s1#ex1:ShoppingCart _c1[ex1:shoppingCart->_s1] _s1[ex1:value->450] _c1[ex1:status->"New"]}

Notice that steps 4 and 5 can be equivalently replaced by the single compound action:

  ☐

4 Production rules and rule sets

This section specifies the syntax and semantics of RIF-PRD rules and rule sets.

4.1 Abstract syntax

The alphabet of the RIF-PRD rule language includes the alphabets of the RIF-PRD condition language and the RIF-PRD action language and adds symbols for:

4.1.1 Rules

Definition (Rule). A rule can be one of:


Example 4.1. The Gold rule, from the running example: A "Silver" customer with a shopping cart worth at least $2,000 is awarded the "Gold" status, can be represented using the following rule with variable declaration:

Forall ?customer such that And( ?customer # ex1:Customer
                                ?customer[ex1:status->"Silver"] )
  (Forall ?shoppingCart such that And( ?shoppingCart # ex1:ShoppingCart
                                      ?customer[ex1:shoppingCart->?shoppingCart] )
     (If Exists ?value (And( ?shoppingCart[ex1:value->?value]
                             pred:numeric-greater-than-or-equal(?value 2000))
      Then Do( Modify( ?customer[ex1:status->"Gold"] ) ) )

The function Var(f), that has been defined for condition formulas and extended to actions, is further extended to rules, as follows:

4.1.2 Groups

As was already mentioned in the Overview, production rules have an operational semantics that can be described in terms of matching rules against states of the fact base, selecting rule instances to be executed, and executing rule instances' actions to transition to new states of the fact base.

When production rules are interchanged, the intended rule instance selection strategy, often called the conflict resolution strategy, needs to be interchanged along with the rules. In RIF-PRD, the group construct is used to group sets of rules and to associate them with a conflict resolution strategy. Many production rule systems use priorities associated with rules as part of their conflict resolution strategy. In RIF-PRD, the group is also used to carry the priority information that may be associated with the interchanged rules.

Definition (Group). A group consists of a, possibly empty, set of rules and groups, associated with a resolution conflict strategy and, a priority. If strategy is an IRI that identifies a conflict resolution strategy, if priority is an integer, and if each rgj, 0 ≤ j ≤ n, is either a rule or a group, then any of the following represents a group:

If a conflict resolution strategy is not explicitly attached to a group, the strategy defaults to rif:forwardChaining (specified below, in section Conflict resolution).   ☐

4.1.3 Safeness

The definitions in this section are unchanged from the definitions in the section Safeness in [RIF-Core], except for the definition of RIF-PRD rule safeness, that is extended from the definition of RIF-Core rule safeness. The definitions are reproduced for the reader's convenience.

Intuitively, safeness of rules guarantees that all the variables in a rule can be bound, using pattern matching only, before they are used, in a test or in an action.

To define safeness, we need to define, first, the notion of binding patterns for externally defined functions and predicates, as well as under what conditions variables are considered bound.

Definition (Binding pattern). (from [RIF-Core]) Binding patterns for externally defined functions and predicates are lists of the form (p1, ..., pn), such that pi=b or pi=u, for 1 ≤ i ≤ n: b stands for a "bound" and u stands for an "unbound" argument.   ☐

Each external function or predicate has an associated list of valid binding patterns. We define here the binding patterns valid for the functions and predicates defined in [RIF-DTB].

Every function or predicate f defined in [RIF-DTB] has a valid binding pattern for each of its schemas with only the symbol b such that its length is the number of arguments in the schema. In addition,

The functions and predicates defined in [RIF-DTB] have no other valid binding patterns.

To keep the definitions concise and intuitive, boundedness and safeness are defined, in [RIF-Core], for condition formulas in disjunctive normal form, that can be existentially quantified themselves, but that contain, otherwise, no existential sub-formula. The definitions apply to any valid RIF-Core condition formula, because they can always, in principle, be put in that form, by applying the following syntactic transforms, in sequence:

  1. if f contains existential sub-formulas, all the quantified variables are renamed, if necessary, and given a name that is unique in f, and the scope of the quantifiers is extended to f. Assume, for instance, that f has an existential sub-formula, sf = Exists v1...vn (sf'), n ≥ 1, such that the names v1...vn do not occur in f outside of sf. After the transform, f becomes Exists v1...vn (f'), where f' is f with sf replaced by sf' . The transform is applied iteratively to all the existential sub-formulas in f;
  2. the (possibly existentially quantified) resulting formula is rewritten in disjunctive normal form ([Mendelson97], p. 30).

In RIF-PRD, the definitions apply to conditions formulas in the same form as in [RIF-Core], with the exception that, in the disjunctive normal form, negated sub-formulas can be atomic formulas or existential formulas: in the latter case, the existentially quantified formula must be, itself, in disjunctive normal form, and contain no further existential sub-formulas. The definitions apply to any valid RIF-PRD condition formula, because they can always, in principle, be put in that form, by applying the above syntactic transform, modified as follows to take negation into account:

Definition (Boundedness). (from [RIF-Core]) An external term External(f(t1,...,tn)) is bound in a condition formula, if and only if f has a valid binding pattern (p1, ..., pn) and, for all j, 1 ≤ j ≤ n, such that pj=b, tj is bound in the formula.

A variable, v, is bound in an atomic formula, a, if and only if

A variable, v, is bound in a conjunction formula, f = And(c1...cn), n ≥ 1, if and only if, either

A variable, v, is bound in a disjunction formula, if and only if v is bound in every disjunct where it occurs;

A variable, v, is bound in an existential formula, Exists v1,...,vn (f'), n ≥ 1, if and only if v is bound in f'.   ☐

Notice that the variables, v1,...,vn, that are existentially quantified in an existential formula f = Exists v1,...,vn (f'), are bound in any formula, F, that contains f as a sub-formula, if and only if they are bound in f, since they do not exist outside of f.

Definition (Variable safeness). (from [RIF-Core]) A variable, v, is safe in a condition formula, f, if and only if

Notice that the two definitions, above, are not extended for negation and, followingly, that an universally quantified (rule) variable is never bound or safe in a condition formula as a consequence of occurring in a negative formula.

The definition of rule safeness is replaced by the following one, that extends the one for RIF-Core rules.

Definition (RIF-PRD rule safeness). A RIF-PRD rule, r, is safe if and only if

Definition (Group safeness). (from [RIF-Core]) A group, Group (s1...sn), n ≥ 0, is safe if and only if

4.1.4 Well-formed rules and groups

If f is a rule, Var(f) is the set of the free variables in f.

Definition (Well-formed rule). A rule, r, is a well-formed rule if and only if either

Definition (Well-formed group). A group is well-formed group if and only if it is safe and it contains only well-formed groups, g1...gn, n ≥ 0, and well-formed rules, r1...rm, m ≥ 0, such that Var(ri) = ∅ for all i, 0 ≤ i ≤ m.   ☐

The set of the well-formed groups contains all the production rule sets that can be meaningfully interchanged using RIF-PRD.

4.2 Operational semantics of rules and rule sets

4.2.1 Motivation and example

As mentioned in the Overview, the description of a production rule system as a transition system is used to specify the semantics of production rules and rule sets interchanged using RIF-PRD.

The intuition of describing a production rule system as a transition system is that, given a set of production rules RS and a fact base w0, the rules in RS that are satisfied, in some sense, in w0 determine an action a1, whose execution results in a new fact base w1; the rules in RS that are satisfied in w1 determine an action a2 to execute in w1, and so on, until the system reaches a final state and stops. The result is the fact base wn when the system stops.

Example 4.2. The Rif Shop, Inc. is a rif-raf retail chain, with brick and mortar shops all over the world and virtual storefronts in many on-line shops. The Rif Shop, Inc. maintains its customer fidelity management policies in the form of production rule sets. The customer management department uses RIF-PRD to publish rule sets to all the shops and licensees so that everyone uses the latest version of the rules, even though several different rule engines are in use (in fact, some of the smallest shops actually run the rules by hand).

Here is a small rule set that governs discounts and customer status updates at checkout time (to keep the example short, this is a subset of the rules described in the running example):

(* ex1:CheckoutRuleset *)
Group rif:forwardChaining (

  (* ex1:GoldRule *)
  Group 10 (
    Forall ?customer such that (And( ?customer # ex1:Customer
                                     ?customer[ex1:status->"Silver"] ) )
      (Forall ?shoppingCart such that (?customer[ex1:shoppingCart->?shoppingCart])
         (If Exists ?value (And( ?shoppingCart[ex1:value->?value]
                                 pred:numeric-greater-than-or-equal(?value 2000))
          Then Do( Modify( ?customer[ex1:status->"Gold"] ) ) ) )

  (* ex1:DiscountRule *)
  Group (
    Forall ?customer such that (And( ?customer # ex1:Customer ) )
      (If Or (?customer[ex1:status->"Silver"]
              ?customer[ex1:status->"Gold"] )
       Then Do( (?s ?customer[ex1:shoppingCart->?s])
                (?v ?s[ex1:value->?v])
                Modify( ?s[ex1:value->func:numeric-multiply(?v 0.95)] ) ) ) )

To see how the rule set works, consider the case of a shop where the checkout processing of customer John is about to start. The initial state of the fact base can be represented as follows:

w0 = {_john#ex1:Customer _john[ex1:status->"Silver"] _s1#ex1:ShoppingCart _john[ex1:shoppingCart->_s1] _s1[ex1:value->2000]}

When instantiated against w0, the first pattern in the "Gold rule", And( ?customer#ex1:Customer ?customer[ex1:status->"Silver"] ), yields the single matching substitution: {(_john/?customer)}. The second pattern in the same rule also yields a single matching substitution: {(_john/?customer)(_s1/?shoppingCart)}, for which the existential condition is satisfied.

Likewise, the instantiation of the "Discount rule" yields a single matching substitution that satisfies the condition: {(_john/?customer)}. The conflict set is:
{ex1:GoldRule/{(_john/?customer)(_s1/?shoppingCart)}, ex1:DiscountRule/{(_john/?customer)}}

The instance ex1:GoldRule/{(_john/?customer)(_s1/?shoppingCart)} is selected because of its higher priority. The ground compound action: Modify(_john[ex1:status->"Gold"]), is executed, resulting in a new state of the fact base, represented as follows:

w1 = {_john#ex1:Customer _john[ex1:status->"Gold"] _s1#ex1:ShoppingCart _john[ex1:shoppingCart->_s1] _s1[ex1:value->2000]}

In the next cycle, there is no substitution for the rule variable ?customer that matches the pattern to the state of the fact base, and the only matching rule instance is: ex1:DiscountRule/{(_john/?customer)}, which is selected for execution. The action variables ?s and ?v are bound, based on the state of the fact base, to _s1 and 200, respectively, and the ground compound action, Modify(_s1[ex1:value->1900]), is executed, resulting in a new state of the fact base:

w2 = {_john#ex1:Customer _john[ex1:status->"Gold"] _s1#ex1:ShoppingCart _john[ex1:shoppingCart->_s1] _s1[ex1:value->1900]}

In w2, the only matching rule instance is, again: ex1:DiscountRule/{(_john/?customer)}. However, that same instance has already been selected and the corresponding action has been executed. Nothing has changed in the state of the fact base that would justify that the rule instance be selected gain. The principle of refraction applies, and the rule instance is removed from consideration.

This leaves the conflict set empty, and the system, having detected a final state, stops.

The result of the execution of the system is w2.   ☐

4.2.2 Rules normalization

A rule, R, whose condition, rewritten in disjunctive normal form as described in section Safeness, consists of more than one disjunct, is equivalent, logically as well as operationally, to a set (or conjunction) of rules that have, all, the same conclusion as R, and each rule has one of the disjuncts as its condition: the rule R: If C1 Or ... Or Cn Then A is equivalent to the set of rules {ri=0..n| ri: If Ci Then A}.

Without loss of generality, and to keep the specification as simple and intuitive as possible, the operational semantics of production rules and rule sets is specified, in the following sections, for rules and rule sets that have been normalized as follows:

  1. All the rules are rewritten in disjunctive normal form as described in section Safeness;
  2. Each rule is replaced by a group of rules

In the same way, without loss of generality, and to keep the specification as simple and intuitive as possible, the operational semantics of production rules and rule sets is specified, in the following sections, for rules and rule sets where all the compound actions have been replaced by the equivalent sequences of atomic actions.

4.2.3 Definitions and notational conventions

Formally, a production rule system is defined as a labeled terminal transition system (e.g. PLO04), for the purpose of specifying the semantics of a RIF-PRD rule or group of rules.

Definition (labeled terminal transition system): A labeled terminal transition system is a structure {C, L, →, T}, where

For many purposes, a representation of the states of the fact base is an appropriate representation of the states of a production rule system seen as a transition system. However, the most widely used conflict resolution strategies require information about the history of the system, in particular with respect to the rule instances that have been selected for execution in previous states. Therefore, each state of the transition system used to represent a production rule system must keep a memory of the previous states and of the rule instances that where selected and that triggered the transition in those states.

Here, a rule instance is defined as the result of the substitution of constants for all the rule variables in a rule.

Let R denote the set of all the rules in the rule language under consideration.

Definition (Rule instance). Given a rule, r ∈ R, and a ground substitution, σ, such that CVar(r) ⊆ Dom(σ), where CVar(r) denotes the set of the rule variables in r, the result, ri = σ(r), of the substitution of the constant σ(?x) for each variable ?x ∈ CVar(r) is a rule instance (or, simply, an instance) of r.  ☐

Given a rule instance ri, let rule(ri) identify the rule from which ri is derived by substitution of constants for the rule variables, and let substitution(ri) denote the substitution by which ri is derived from rule(ri).

In the following, two rule instances ri1 and ri2 of a same rule r will be considered different if and only if substitution(ri1) and substitution(ri2) substitute a different constant for at least one of the rule variables in CVar(r).

A rule instance, ri, is said to match a state of a fact base, w, if its defining substitution, substitution(ri), matches the RIF-PRD condition formula that represents the condition of the instantiated rule, rule(ri), to the set of ground atomic formulas that represents the state of facts w.

Let W denote the set of all the possible states of a fact base.

Definition (Matching rule instance). Given a rule instance, ri, and a state of the fact base, w ∈ W, ri is said to match w if and only if one of the following is true:

Definition (Action instance). Given a state of the fact base, w ∈ W, given a rule instance, ri, of a rule in a rule set, RS, and given the action block in the action part of the rule rule(ri): Do((v1 p1)...(vn pn) a1...am), n ≥ 0, m ≥ 1, where the (v1 p1), 0 ≤ i ≤ n, represent the action variable declarations and the aj, 1 ≤ j ≤ m, represent the sequence of atomic actions in the action block; if ri matches w, the substitution σ = substitution(ri) is extended to the action variables v1...vn, n ≥ 0, in the following way:

The sequence of ground atomic actions that is the result of substituting a constant for each variable in the atomic actions of the action block of the rule instance, ri, according to the extended substitution, is the action instance associated to ri.   ☐

Let actions(ri) denote the action instance that is associated to a rule instance ri. By extension, given an ordered set of rule instances, ori, actions(ori) denotes the sequence of ground atomic actions that is the concatenation, preserving the order in ori, of the action instances associated to the rule instances in ori.

The components of the states of a production rule system seen as a transition system can now be defined more precisely. To avoid confusion between the states of the fact base and the states of the transition system, the latter will be called production rule system states.

Definition (Production rule system state). A production rule system state (or, simply, a system state) is either a system cycle state or a system transitional state. Every production rule system state, s, cycle or transitional, is characterized by

In the following, we will write previous(s) = NIL to denote that a system state s is the initial state.

Definition (Conflict set). Given a rule set, RS ⊆ R, and a system state, s, the conflict set determined by RS in s is the set, conflictSet(RS, s) of all the different instances of the rules in RS that match the state of the fact base, facts(s) ∈ W.   ☐

The rule instances that are in the conflict set are, sometimes, said to be fireable.

In each non-final cycle state, s, of a production rule system, a subset, picked(s), of the rule instances in the conflict set is selected and ordered; their action parts are instantiated, and the resulting sequence of ground atomic actions is executed. This is sometimes called: firing the selected instances.

4.2.4 Operational semantics of a production rule system

All the elements that are required to define a production rule system as a labeled terminal transition system have now been defined.

Definition (RIF-PRD Production Rule System). A RIF-PRD production rule system is defined as a labeled terminal transition system PRS = {S, A, →PRS, T}, where :

Intuitively, the first condition in the definition of the transition relation →PRS states that a production rule system can transition from one system cycle state to another only if the state of facts in the latter system cycle state can be reached from the state of facts in the former by performing a sequence of ground atomic actions supported by RIF-PRD, according to the semantics of the atomic actions.

The second condition states that the allowed paths out of any given system cycle state are determined only by how rule instances are picked for execution, from the conflict set, by the conflict resolution strategy.

Given a rule set RS ⊆ R, the associated conflict resolution strategy, LS, and halting test, H, and an initial state of the fact base, w ∈ W, the input function to a RIF-PRD production rule system is defined as:
Eval(RS, LS, H, w) →PRS s ∈ S, such that facts(s) = w and previous(s) = NIL.
Using *PRS to denote the transitive closure of the transition relation PRS, there are zero, one or more final states of the system, s' ∈ T, such that:
Eval(RS, LS, H, w) →*PRS s'.

The execution of a rule set, RS, in a state, w, of a fact base, may result in zero, one or more final state of the fact base, w' = facts(s'), depending on the conflict resolution strategy and the set of final system states.

Therefore, the behavior of a RIF-PRD production rule system also depends on:

  1. the conflict resolution strategy, that is, how rule instances are precisely selected for execution from the rule instances that match a given state of the fact base, and
  2. how the set T of final system states is precisely defined.

4.2.5 Conflict resolution

The process of selecting one or more rule instances from the conflict set for firing is often called: conflict resolution.

In RIF-PRD the conflict resolution algorithm (or conflict resolution strategy) that is intended for a set of rules is denoted by a keyword or a set of keywords that is attached to the rule set. In this version of the RIF-PRD specification, a single conflict resolution strategy is specified normatively: it is denoted by the keyword rif:forwardChaining (a constant of type rif:IRI), because it accounts for a common conflict resolution strategy used in most forward-chaining production rule systems. That conflict resolution strategy selects a single rule instance for execution.

Future versions of the RIF-PRD specification may specify normatively the intended conflict resolution strategies to be attached to additional keywords. In addition, RIF-PRD documents may include non-standard keywords: it is the responsibility of the producers and consumers of such document to agree on the intended conflict resolution strategies that are denoted by such non-standard keywords. Future or non-standard conflict resolution strategies may select an ordered set of rule instances for execution, instead of a single one: the functions picked and actions, in the previous section, have been defined to take this case into account.

Conflict resolution strategy: rif:forwardChaining

Most existing production rule systems implement conflict resolution algorithms that are a combination of the following elements (under these or other, idiosyncratic names; and possibly combined with additional, idiosyncratic rules):

Many existing production rule systems implement also some kind of fire the most specific rule first strategy, in combination with the above. However, whereas they agree on the definition of refraction and the priority or recency ordering, existing production rule systems vary widely on the precise definition of the specificity ordering. As a consequence, rule instance specificity was not included in the basic conflict resolution strategy that RIF-PRD specifies normatively.

The RIF-PRD keyword rif:forwardChaining denotes the common conflict resolution strategy that can be summarized as follows: given a conflict set

  1. Refraction is applied to the conflict set, that is, all the refracted rule instances are removed from further consideration;
  2. The remaining rule instances are ordered by decreasing priority, and only the rule instances with the highest priority are kept for further consideration;
  3. The remaining rule instances are ordered by decreasing recency, and only the most recent rule instances are kept for further consideration;
  4. Any remaining tie is broken is some way, and a single rule instance is kept for firing.

As specified earlier, picked(s) denotes the ordered list of the rule instances that were picked in a system state, s. Under the conflict resolution strategy denoted by rif:forwardChaining, for any given system cycle state, s, the list denoted by picked(s) contains a single rule instance. By definiiton, if s is a system transitional state, picked(s) is the empty set.

Given a system state, s, a rule set, RS, and a rule instance, ri ∈ conflictSet(RS, s), let recency(ri, s) denote the number of system states before s, in which ri has been continuously a matching instance: if s is the current system state, recency(ri, s) provides a measure of the recency of the rule instance ri. recency(ri, s) is specified recursively as follows:

In the same way, given a rule instance, ri, and a system state, s, let lastPicked(ri, s) denote the number of system states before s, since ri has been last fired. lastPicked(ri, s) is specified recursively as follows:

Given a rule instance, ri, let priority(ri) denote the priority that is associated to rule(ri), or zero, if no priority is associated to rule(ri). If rule(ri) is inside nested Groups, priority(ri) denotes the priority that is associated with the innermost Group to which a priority is explicitly associated, or zero.

Example 4.3. Consider the following RIF-PRD document:

Document (
  Prefix( ex2 <http://example.com/2009/prd3#> )
  (* ex2:ExampleRuleSet *)
  Group (
    (* ex2:Rule_1 *) Forall ...
    (* ex2:HighPriorityRules *)
    Group 10 (
      (* ex2:Rule_2 *) Forall ...
      (* ex2:Rule_3 *) 
      Group 9 (Forall ... ) )
    (* ex2:NoPriorityRules *)
    Group (
      (* ex2:Rule_4 *) Forall ...
      (* ex2:Rule_5 *) Forall ... )
 )

No conflict resolution strategy is identified explicitly, so the default strategy rif:forwardChaining is used.

Because the ex2:ExampleRuleSet group does not specify a priority, the default priority 0 is used. Rule 1, not being in any other group, inherits its priority, 0, from the top-level group.

Rule 2 inherits its priority, 10, from the enclosing group, identified as ex2:HighPriorityRules. Rule 3 specifies its own, lower, priority: 9.

Since neither Rule 4 nor Rule 5 specify a priority, they inherit their priority from the enclosing group ex2:NoPriorityRules, which does not specify one either, and, thus, they inherit 0 from the top-level group, ex2:ExampleRuleSet.   ☐

Given a set of rule instances, cs, the conflict resolution strategy rif:forwardChaining can now be described with the help of four rules, where ri and ri' are rule instances:

  1. Refraction rule: if ri ∈ cs and lastPicked(ri, s) < recency(ri, s), then cs = cs - ri;
  2. Priority rule: if ri ∈ cs and ri' ∈ cs and priority(ri) < priority(ri'), then cs = cs - ri;
  3. Recency rule: if ri ∈ cs and ri' ∈ cs and recency(ri, s) > recency(ri', s), then cs = cs - ri;
  4. Tie-break rule: if ri ∈ cs, then cs = {ri}. RIF-PRD does not specify the tie-break rule more precisely: how a single instance is selected from the remaining set is implementation specific.

The refraction rule removes the instances that have been in the conflict set in all the system states at least since they were last fired; the priority rule removes the instances such that there is at least one instance with a higher priority; the recency rule removes the instances such that there is at least one instance that is more recent; and the tie-break rule keeps one rule from the set.

To select the singleton rule instance, picked(s), to be fired in a system state, s, given a rule set, RS, the conflict resolution strategy denoted by the keyword rif:forwardChaining consists of the following sequence of steps:

  1. initialize picked(s) with the conflict set, that a rule set RS determines in a system state s: picked(s) = conflictSet(RS, s);
  2. apply the refraction rule to all the rule instances in picked(s);
  3. then apply the priority rule to all the remaining instances in picked(s);
  4. then apply the recency rule to all the remaining instances in picked(s);
  5. then apply the tie-break rule to the remaing instance in picked(s);
  6. return picked(s).

Example 4.4. Consider, from example 4.2, the conflict set that the rule set ex1:CheckoutRuleset determines in the system state, s2, that corresponds to the state w2 = facts(s2) of the fact base, and use it to initialize the set of rule instance considered for firing, picked(s2):

conflictSet(ex1:CheckoutRuleset, s2) = { ex1:DiscountRule/{(_john/?customer)} } = picked(s2)

The single rule instance in the conflict set, ri = ex1:DiscountRule/{(_john/?customer)}, did already belong to the conflict sets in the two previous states, conflictSet(ex1:CheckoutRuleset, s1) and conflictSet(ex1:CheckoutRuleset, s0); so that its recency in s2 is: recency(ri, s2) = 3.

On the other hand, that rule instance was fired in system state s1: picked(s1) = (ex1:DiscountRule/{(_john/?customer)}); so that, in s2, it has been last fired one cycle before: lastPicked(ri, s2) = 1.

Therefore, lastPicked(ri, s2) < recency(ri, s2), and ri is removed from picked(s2) by refraction, leaving picked(s2) empty.   ☐

4.2.6 Halting test

By default, a system state is final, given a rule set, RS, and a conflict resolution strategy, LS, if there is no rule instance available for firing after application of the conflict resolution strategy.

For the conflict resolution strategy identified by the RIF-PRD keyword rif:forwardChaining, a system state, s, is final given a rule set, RS if and only if the remaining conflict set is empty after application of the refraction rule to all the rule instances in conflictSet(RS, s). In particular, all the system states, s, such that conflictSet(RS, s) = ∅ are final.

5 Document and imports

This section specifies the structure of a RIF-PRD document and its semantics when it includes import directives.

5.1 Abstract syntax

In addition to the language of conditions, actions, and rules, RIF-PRD provides a construct to denote the import of a RIF or non-RIF document. Import enables the modular interchange of RIF documents, and the interchange of combinations of multiple RIF and non-RIF documents.

5.1.1 Import directive

Definition (Import directive). An import directive consists of:

RIF-PRD gives meaning to one-argument import directives only. Such directives can be used to import other RIF-PRD and RIF-Core documents. Two-argument import directives are provided to enable import of other types of documents, and their semantics is covered by other specifications. For example, the syntax and semantics of the import of RDF and OWL documents, and their combination with a RIF document, is specified in [RIF-RDF-OWL].

5.1.2 RIF-PRD document

Definition (RIF-PRD document). A RIF-PRD document consists of zero or more import directives, and zero or one group.   ☐

Definition (Imported document). A document is said to be directly imported by a RIF document, D, if and only if it is identified by the locator IRI in an import directive in D. A document is said to be imported by a RIF document, D, if it is directly imported by D, or if it is imported, directly or not, by a RIF document that is directly imported by D.   ☐

Definition (Document safeness). (from [RIF-Core]) A document is safe if and only if it

5.1.3 Well-formed documents

Definition (Conflict resolution strategy associated with a document). A conflict resolution strategy is associated with a RIF-PRD document, D, if and only if

Definition (Well-formed RIF-PRD document). A RIF-PRD document, D, is well-formed if and only if it satisfies all the following conditions:

The last condition in the above definition makes the intent behind the rif:local constants clear: occurrences of such constants in different documents can be interpreted differently even if they have the same name. Therefore, each document can choose the names for the rif:local constants freely and without regard to the names of such constants used in the imported documents.

5.2 Operational semantics of RIF-PRD documents

The semantics of a well-formed RIF-PRD document that contains no import directive is the semantics of the rule set that is represented by the top-level group in the document, evaluated with the conflict resolution strategy that is associated to the document, and the default halting test, as specified above, in section Halting test.

The semantics of a well-formed RIF-PRD document, D, that imports the well-formed RIF-PRD documents D1, ..., Dn, n ≥ 1, is the semantics of the rule set that is the union of the rule sets represented by the top-level groups in D and the imported documents, with the rif:local constants renamed to ensure that the same symbol does not occur in two different component rule sets, and evaluated with the conflict resolution strategy that is associated to the document, and the default halting test.

6 Built-in functions, predicates and actions

In addition to externally specified functions and predicates, and in particular, in addition to the functions and predicates built-ins defined in [RIF-DTB], RIF-PRD supports externally specified actions, and defines action built-ins.

The syntax and semantics of action built-ins are specified like for the other buit-ins, as described in the section Syntax and Semantics of Built-ins in [RIF-DTB]. However, their formal semantics is trivial: action built-ins behave like predicates that are always true, since action built-ins, in RIF-PRD, MUST NOT affect the semantics of the rules.

Although they must not affect the semantics of the rules, action built-ins may have other side effects.

RIF action built-ins are defined in the namespace: http://www.w3.org/2007/rif-builtin-action#. In this document, we will use the prefix: act: to denote the RIF action built-ins namespace.

6.1 Built-in actions

6.1.1 act:print

7 Conformance and interoperability

7.1 Semantics-preserving transformations

RIF-PRD conformance is described partially in terms of semantics-preserving transformations.

The intuitive idea is that, for any initial state of facts, the conformant consumer of a conformant RIF-PRD document must reach at least one of the final state of facts intended by the conformant producer of the document, and that it must never reach any final state of facts that was not intended by the producer. That is:

Let Τ be a set of datatypes and symbol spaces that includes the datatypes specified in [RIF-DTB] and the symbol spaces rif:iri and rif:local. Suppose also that Ε is a set of external predicates and functions that includes the built-ins listed in [RIF-DTB] and in the section Built-in actions. We say that a rule r is a RIF-PRDΤ,Ε rule if and only if

Suppose, further, that C is a set of conflict resolution strategies that includes the one specified in section Conflict resolution, and that H is a set of halting tests that includes the one specified in section Halting test: we say that a rule set , R, is a RIF-PRDΤ,Ε,C,H rule set if and only if

Given a RIF-PRDΤ,Ε,C,H rule set, R, an initial state of the fact base, w, a conflict resolution strategy c ∈ C and a halting test h ∈ H, let FR,w,c,h denote the set of all the sets, f, of RIF-PRD ground atomic formulas that represent final states of the fact base, w' , according to the operational semantics of a RIF-PRD production rule system, that is: f ∈ FR,w,c,h if and only if there is a state, s' , of the system, such that Eval(R, c, h, w) →*PRS s' and w' = facts(s') and f is a representation of w' .

In addition, given a rule language, L, a rule set expressed in L, RL, a conflict resolution strategy, c, a halting test, h, and an initial state of the fact base, w, let FL,RL, c, h, w denote the set of all the formulas in L that represent a final state of the fact base that an L processor can possibly reach.

Definition (Semantics preserving mapping).

7.2 Conformance Clauses

Definition (RIF-PRD conformance).

In addition, conformant RIF-PRD producers and consumers SHOULD preserve annotations.

7.3 Interoperability

[RIF-Core] is specified as a specialization of RIF-PRD: all valid [RIF-Core] documents are valid RIF-PRD documents and must be accepted by any conformant RIF-PRD consumer.

Conversely, it is desirable that any valid RIF-PRD document that uses only abstract syntax that is defined in [RIF-Core] be a valid [RIF-Core] document as well. For some abstract constructs that are defined in both RIF-Core and RIF-PRD, RIF-PRD defines alternative XML syntax that is not valid RIF-Core XML syntax. For example, an action block that contains no action variable declaration and only assert atomic actions can be expressed in RIF-PRD using the XML elements Do or And. Only the latter option is valid RIF-Core XML syntax.

To maximize interoperability with RIF-Core and its non-RIF-PRD extensions, a conformant RIF-PRD consumer SHOULD produce valid [RIF-Core] documents whenever possible. Specifically, a conformant RIF-PRD producer SHOULD use only valid [RIF-Core] XML syntax to serialize a rule set that satisfies all of the following:

When processing a rule set that satisfies all the above conditions, a RIF-PRD producer is guaranteed to produce a valid [RIF-Core] XML document by applying the following rules recursively:

  1. Remove redundant information. The behavior role element and all its sub-elements should be omitted in the RIF-PRD XML document;
  2. Remove nested rule variable declarations. If the rule inside a rule with variable delcaration, r1, is also a rule with variable declaration, r2, all the rule variable delarations and all the patterns that occur in r1 (and not in r2) should be added to the rule variable declarations and the patterns that occur in r2, and, after the transform, r1 should be replaced by r2, in the rule set;
  3. Remove patterns. If a pattern occurs in a rule with variable declaration, r1:
  4. Convert action blocks. The action block, in each rule, should be replaced by a conjunction, and, inside the conjunction, each assert action should be replaced by its target atomic formula.

Example 7.1. Consider the following rule, R, derived from the Gold rule, in the running example, to have only assertions in the action part:

R: Forall ?customer such that (And( ?customer # ex1:Customer
                                    ?customer[status->"Silver"] ) )
      (Forall ?shoppingCart such that (?customer[shoppingCart->?shoppingCart])
         (If Exists ?value (And( ?shoppingCart[value->?value]
                                 pred:numeric-greater-than-or-equal(?value 2000))
          Then Do( Assert(ex1:Foo(?customer))
                   Assert(ex1:Bar(?shoppingCart)) ) ) )

The serialization of R in the following RIF-Core conformant XML form does not impacts its semantics (see example 8.12 for another valid RIF-PRD XML serialization, that is not RIF-Core conformant):

<Forall>
   <declare><Var>?customer</Var></declare>
   <declare><Var>?shoppingCart</Var></declare>
   <formula>
      <Implies>
         <if>
            <And>
               <formula>   <!-- first pattern -->
                  <And>
                     <formula><Member> ... </Member></formula>
                     <formula><Frame> ... </Frame></formula>
                  </And>
               </formula>
               <formula>   <!-- second pattern -->
                  <Member> ... </Member>
               </formula>
               <formula>   <!-- original existential condition -->
                  ...
               </formula>
           </And>
        </if>
        <then>
           <And>
              <formula>   <!-- serialization of ex1:Foo(?customer) -->
                 ...
              </formula>
              <formula>   <!-- serialization of ex1:Bar(?shoppingCart) -->
                 ...
              </formula>
        </then>
     </Implies>
  </formula>
</Forall>

8 XML Syntax

This section specifies the concrete XML syntax of RIF-PRD. The concrete syntax is derived from the abstract syntax defined in sections 2.1, 3.1 and 4.1 using simple mappings. The semantics of the concrete syntax is the same as the semantics of the abstract syntax.

8.1 Notational conventions

8.1.1 Namespaces

Throughout this document, the xsd: prefix stands for the XML Schema namespace URI http://www.w3.org/2001/XMLSchema#, the rdf: prefix stands for http://www.w3.org/1999/02/22-rdf-syntax-ns#, and rif: stands for the URI of the RIF namespace, http://www.w3.org/2007/rif#.

Syntax such as xsd:string should be understood as a compact URI [CURIE] -- a macro that expands to a concatenation of the character sequence denoted by the prefix xsd and the string string. The compact URI notation is used for brevity only, and xsd:string should be understood, in this document, as an abbreviation for http://www.w3.org/2001/XMLSchema#string.

8.1.2 BNF pseudo-schemas

The XML syntax of RIF-PRD is specified for each component as a pseudo-schema, as part of the description of the component. The pseudo-schemas use BNF-style conventions for attributes and elements: "?" denotes optionality (i.e. zero or one occurrences), "*" denotes zero or more occurrences, "+" one or more occurrences, "[" and "]" are used to form groups, and "|" represents choice. Attributes are conventionally assigned a value which corresponds to their type, as defined in the normative schema. Elements are conventionally assigned a value which is the name of the syntactic class of their content, as defined in the normative schema.

<!-- sample pseudo-schema -->
    <defined_element
          required_attribute_of_type_string="xs:string"
          optional_attribute_of_type_int="xs:int"? >
      <required_element />
      <optional_element />?
      <one_or_more_of_these_elements />+
      [ <choice_1 /> | <choice_2 /> ]*
    </defined_element>

8.1.3 Syntactic components

Three kinds of syntactic components are used to specify RIF-PRD:

8.2 Relative IRIs and XML base

Relative IRIs are allowed in RIF-PRD XML syntax, anywhere IRIs are allowed, including constant types, symbol spaces, location, and profile. The attribute xml:base [XML-Base] is used to make them absolute.

8.3 Conditions

This section specifies the XML constructs that are used in RIF-PRD to serialize condition formulas.

8.3.1 TERM

The TERM class of constructs is used to serialize terms, be they simple terms, that is, constants and variables; lists; or positional terms, the latter being, per the definition of a well-formed formula, representations of externally defined functions.

As an abstract class, TERM is not associated with specific XML markup in RIF-PRD instance documents.

    [ Const | Var | List | External ]

8.3.1.1 Const

In RIF, the Const element is used to serialize a constant.

The Const element has a required type attribute and an optional xml:lang attribute:

The content of the Const element is the constant's literal, which can be any Unicode character string.

    <Const type=rif:iri [xml:lang=xsd:language]? >
        Any Unicode string
    </Const>

Example 8.1.

a. A constant with built-in type xsd:integer and value 2,000:

<Const type="xsd:integer">2000</Const>

b. The Customer class, in the running example, is identified by a constant of type rif:iri, in the namespace http://example.com/2009/prd2#:

<Const type="rif:iri">
   http://example.com/2009/prd2#Customer
</Const>

8.3.1.2 Var

In RIF, the Var element is used to serialize a variable.

The content of the Var element is the variable's name, serialized as an NCName.

    <Var> xsd:NCName </Var>

8.3.1.3 List

In RIF, the List element is used to serialize a list.

The List element contains one items element, that contains zero or more TERMs (without variables) that serialize the elements of the list. The order of the sub-elements is significant and MUST be preserved. This is emphasized by the fixed value "yes" of the mandatory attribute ordered in the items element.

    <List>
      <items ordered="yes">
        GROUNDTERM*
      </items>
    </List>

Example 8.2.

<List>
  <items ordered="yes">
    <Const type="xsd:string> New </Const>
    <Const type="xsd:string> Bronze </Const>
    <Const type="xsd:string> Silver </Const>
    <Const type="xsd:string> Gold </Const>
  </items>
</List>
<List>
  <items ordered="yes">
  </items>
</List>

8.3.1.4 External

As a TERM, the External element is used to serialize a positional term. In RIF-PRD, a positional term represents always a call to an externally defined function, e.g. a built-in, a user-defined function, a query to an external data source, etc.

The External element contains one content element, which in turn contains one Expr element that contains one op element, followed zero or one args element:

    <External>
       <content>
          <Expr>
             <op> Const </op>
             <args ordered="yes"> TERM+ </args>?
          </Expr>
       </content>
    </External>

Example 8.3. The example shows one way to serialize, in RIF-PRD, the product of a variable named ?value and the xsd:decimal value 0.9, where the operation conforms to the specification of the built-in func:numeric-multiply, as specified in [RIF-DTB].

RIF built-in functions are associated with the namespace http://www.w3.org/2007/rif-builtin-function#.

<External>
   <content>
      <Expr>    
         <op> <Const type="rif:iri"> http://www.w3.org/2007/rif-builtin-function#numeric-multiply </Const> </op>
         <args ordered="yes"> 
            <Var> ?value </Var>
            <Const type="xsd:decimal"> 0.9 </Const>
         </args>
      </Expr>
   </content>
</External>

8.3.2 ATOMIC

The ATOMIC class is used to serialize atomic formulas: positional atoms, equality, membership and subclass atomic formulas, frame atomic formulas and externally defined atomic formulas.

As an abstract class, ATOMIC is not associated with specific XML markup in RIF-PRD instance documents.

    [ Atom | Equal | Member | Subclass | Frame | External ]

8.3.2.1 Atom

In RIF, the Atom element is used to serialize a positional atomic formula.

The Atom element contains one op element, followed by zero or one args element:

    <Atom>
       <op> Const </op>
       <args ordered="yes"> TERM+ </args>?
    </Atom>

Example 8.4. The example shows the RIF XML serialization of the positional atom ex1:gold(?customer), where the predicate symbol gold is defined in the example namespace http://example.com/2009/prd2#.

<Atom>
   <op>
      <Const type="rif:iri">
         http://example.com/2009/prd2#gold
      </Const>
   </op>
   <args ordered="yes">
      <Var> ?customer </Var>
   </args>
</Atom>

8.3.2.2 Equal

In RIF, the Equal element is used to serialize equality atomic formulas.

The Equal element must contain one left sub-element and one right sub-element. The content of the left and right elements must be a construct from the TERM abstract class, that serialize the terms of the equality. The order of the sub-elements is not significant.

    <Equal>
       <left> TERM </left>
       <right> TERM </right>
    </Equal>

8.3.2.3 Member

In RIF, the Member element is used to serialize membership atomic formulas.

The Member element contains two required sub-elements:

    <Member>
       <instance> TERM </instance>
       <class> TERM </class>
    </Member>

Example 8.5. The example shows the RIF XML serialization of class membership atom that tests whether a variable named ?customer belongs to a class identified by the name Customer in the namespace http://example.com/2009/prd2#

<Member>
   <instance> <Var> ?customer </Var> </instance>
   <class>
      <Const type="rif:iri">
         http://example.com/2009/prd2#Customer
      </Const>
   </class>
</Member>

8.3.2.4 Subclass

In RIF, the Subclass element is used to serialize subclass atomic formulas.

The Subclass element contains two required sub-elements:

    <Subclass>
       <sub> TERM </sub>
       <super> TERM </super>
    </Subclass>

8.3.2.5 Frame

In RIF, the Frame element is used to serialize frame atomic formulas.

Accordingly, a Frame element must contain:

    <Frame>
       <object> TERM </object>
       <slot ordered="yes"> TERM TERM </slot>*
    </Frame>

Example 8.6. The example shows the RIF XML serialization of an expression that states that the object denoted by the variable ?customer has the value denoted by the string "Gold" for the property identified by the symbol status that is defined in the example namespace http://example.com/2009/prd2#

<Frame>
   <object> <Var> ?customer </Var> </object>
   <slot ordered="yes"> 
      <Const type="rif:iri">
         http://example.com/2009/prd2#status
      </Const>
      <Const type="xsd:string> Gold </Const>
   </slot>
</Frame>

8.3.2.6 External

In RIF-PRD, the External element is also used to serialize an externally defined atomic formula, in addition to serializing externally defined functions.

When it is an ATOMIC (as opposed to a TERM; that is, in particular, when it appears in a place where an ATOMIC is expected, and not a TERM), the External element contains one content element that contains one Atom element. The Atom element serializes the externally defined atom properly said:

The op Const in the Atom element must be a symbol of type rif:iri that must uniquely identify the externally defined predicate to be applied to the args TERMs.

    <External>
       <content>
          Atom
       </content>
    </External>

Example 8.7. The example below shows the RIF XML serialization of an externally defined atomic formula that tests whether the value denoted by the variable named ?value is greater than or equal to the integer value 2000, where the test is intended to behave like the built-in predicate pred:numeric-greater-than-or-equal as specified in [RIF-DTB]:

RIF built-in predicates are associated with the namespace http://www.w3.org/2007/rif-builtin-predicate#.

<External>
   <content>
      <Atom>    
         <op> <Const type="rif:iri"> http://www.w3.org/2007/rif-builtin-predicate#numeric-greater-than-or-equal </Const> </op>
         <args ordered="yes">
            <Var> ?value </Var>
            <Const type="xsd:integer"> 2000 </Const>
         </args>
      </Atom>
   </content>
</External>

8.3.3 FORMULA

The FORMULA class is used to serialize condition formulas, that is, atomic formulas, conjunctions, disjunctions, negations and existentials.

As an abstract class, FORMULA is not associated with specific XML markup in RIF-PRD instance documents.

    [ ATOMIC | And | Or | INeg | Exists ]

8.3.3.1 ATOMIC

An atomic formula is serialized using a single ATOMIC statement. See specification of ATOMIC, above.

8.3.3.2 And

A conjunction is serialized using the And element.

The And element contains zero or more formula sub-elements, each containing an element of the FORMULA group, that serializes one of the conjuncts.

    <And>
       <formula> FORMULA </formula>*
    </And>

8.3.3.3 Or

A disjunction is serialized using the Or element.

The Or element contains zero or more formula sub-elements, each containing an element of the FORMULA group, that serializes one of the disjuncts.

    <Or>
       <formula> FORMULA </formula>*
    </Or>

8.3.3.4 INeg

The kind of negation that is used in RIF-PRD is serialized using the INeg element.

The Negate element contains exactly one formula sub-element. The formula element contains an element of the FORMULA group, that serializes the negated statement.

    <INeg>
       <formula> FORMULA </formula>
    </INeg>

8.3.3.5 Exists

An existentially quantified formula is serialized using the Exists element.

The Exists element contains:

    <Exists>
       <declare> Var </declare>+
       <formula> FORMULA </formula>
    </Exists>

Example 8.8. The example shows the RIF XML serialization of a condition on the existence of a value greater than or equal to 2.000, in the Gold rule of the running example, as represented in example 4.2.

<Exists>
   <declare> <Var> ?value </Var> </declare>
   <formula>
      <And>
         <Frame>
            <object> <Var> ?shoppingCart </Var> </object>
            <slot ordered="yes"> 
               <Const type="rif:iri">
                  http://example.com/2009/prd2#value
               </Const>
               <Var> ?value </Var>
            </slot>
         </Frame>
         <External>
            <content>
               <Atom>    
                  <op> <Const type="rif:iri"> http://www.w3.org/2007/rif-builtin-predicate#numeric-greater-than-or-equal </Const> </op>
                  <args ordered="yes">
                     <Var> ?value </Var>
                     <Const type="xsd:integer"> 2000 </Const>
                  </args>
               </Atom>
            </content>
         </External>
      </And>
   </formula>
</Exists>

8.4 Actions

This section specifies the XML syntax that is used to serialize the action part of a rule supported by RIF-PRD.

8.4.1 ACTION

The ACTION class of elements is used to serialize the actions: assert, retract, modify and execute.

As an abstract class, ACTION is not associated with specific XML markup in RIF-PRD instance documents.

    [ Assert | Retract | Modify | Execute ]

8.4.1.1 Assert

An atomic assertion action is serialized using the Assert element.

The Assert element has one target sub-element that contains an Atom, a Frame or a Member element that represents the target of the action.


    <Assert>
         <target> [ Atom | Frame | Member ] </target>
    </Assert>

8.4.1.2 Retract

The Retract construct is used to serialize retract atomic actions.

The Retract element has one target sub-element that contains an Atom, a Frame, a TERM, or a pair of TERM constructs that represent the target of the action. The target element has an optional attribute, ordered, that MUST be present when the element contains two TERM sub-elements: the order of the sub-elements is significant and MUST be preserved. This is emphasized by the required value "yes" of the attribute.

    <Retract>
       <target ordered="yes"?> 
          [ Atom | Frame | TERM | TERM TERM ]
       </target>
    </Retract>

8.4.1.3 Modify

A compound modification is serialized using the Modify element.

The Modify element has one target sub-element that contains one Frame that represents the target of the action.

    <Modify>
         <target> Frame </target>
    </Modify>


Example 8.9. The example shows the RIF XML representation of the action that updates the status of a customer, in the Gold rule, in the running example, as represented in example 4.2: Modify(?customer[status->"Gold"])

<Modify>
   <target>
      <Frame>
         <object>
            <Var> ?customer </Var>
         </object>
         <slot ordered="yes">
            <Const type="rif:iri"> http://example.com/2009/prd2#status </Const>
            <Const type="xsd:string"> Gold </Const>
         </slot>
      </Frame>
   </target>
</Modify>

The action could be equivalently serialized as the sequence of a Retract and an Assert atomic actions:

<Retract>
   <target ordered="yes">
      <Var> ?customer </Var>
      <Const type="rif:iri"> http://example.com/2009/prd2#status </Const>
   </target>
</Retract>
<Assert>
   <target>
      <Frame>
         <object>
            <Var> ?customer </Var>
         </object>
         <slot ordered="yes">
            <Const type="rif:iri"> http://example.com/2009/prd2#status </Const>
            <Const type="xsd:string"> Gold </Const>
         </slot>
      </Frame>
   </target>
</Assert>

8.4.1.4 Execute

The execution of an externally defined action is serialized using the Execute element.

The Execute element has one target sub-element that contains an Atom, that represents the externally defined action to be executed.

The op Const in the Atom element must be a symbol of type rif:iri that must uniquely identify the externally defined action to be applied to the args TERMs.

    <Execute>
         <target> Atom </target>
    </Execute>

Example 8.10. The example shows the RIF XML serialization of the message printing action, in the Unknonw status rule, in the running example, using the act:print action built-in.

The namespace for RIF-PRD action built-ins is http://www.w3.org/2007/rif-builtin-action#.

<Execute>
  <target>
    <Atom>
      <op>
        <Constant type="rif:iri"> http://www.w3.org/2007/rif-builtin-action#print </Const>
      </op>
      <args ordered="yes">
        <External>
          <content>
            <Expr>
              <op>
                <Constant type="rif:iri"> http://www.w3.org/2007/rif-builtin-function#concat </Const>
              </op>
              <args ordered="yes">
                <Const type="xsd:string> New customer: </Const>
                 ?customer 
              </args>
            </Expr>
          </content>
        </External>
      </args>
    </Atom>
  </target>
</Execute>

8.4.2 ACTION_BLOCK

The ACTION_BLOCK class of constructs is used to represent the conclusion, or action part, of a production rule serialized using RIF-PRD.

If action variables are declared in the action part of a rule, or if some actions are not assertions, the conclusion must be serialized as a full action block, using the Do element. However, simple action blocks that contain only one or more assert actions SHOULD be serialized like the conclusions of logic rules using RIF-Core or RIF-BLD, that is, as a single asserted Atom or Frame, or as a conjunction of the asserted facts, using the And element.

In the latter case, to conform with the definition of an action block well-formedness, the formulas that serialize the individual conjuncts MUST be atomic Atoms and/or Frames.

As an abstract class, ACTION_BLOCK is not associated with specific XML markup in RIF-PRD instance documents.

    [ Do | And | Atom | Frame ]
8.4.2.1 New

The New element is used to serialize the construct used to create a new frame identifer, in an action variable declaration.

The New element is always empty.

    <New />
8.4.2.2 Do

An action block is serialized using the Do element.

A Do element contains:

    <Do>
       <actionVar ordered="yes">
          Var
          [ New | Frame ]
       </actionVar>*
       <actions ordered="yes">
          ACTION+
       </actions>
    </Do>

Example 8.11. The example shows the RIF XML serialization of an action block that asserts that a customer gets a new $5 voucher.

<Do>
   <actionVar ordered="yes">
      <Var>?voucher</Var>
      <New />
   </actionVar>
   <actions  ordered="yes">
      <Assert>
         <target>
            <Member>
               <instance><Var>?voucher</Var></instance>
               <class>
                  <Const type="rif:iri">http://example.com/2009/prd2#Voucher</Const>
               </class>
            </Member>
         </target>
      </Assert>
      <Assert>
         <target>
            <Frame>
               <object><Var>?voucher</Var></object>
               <slot ordered="yes">
                  <Const type="rif:iri">http://example.com/2009/prd2#value</Const>
                  <Const type="xsd:integer">5</Const>
               </slot>
            </Frame>
         </target>
      </Assert>
      <Assert>
         <target>
            <Frame>
               <object><Var>?customer</Var></object>
               <slot ordered="yes">
                  <Const type="rif:iri">http://example.com/2009/prd2#voucher</Const>
                  <Var>?voucher</Var>
               </slot>
            </Frame>
         </target>
      </Assert>
   </actions>
</Do>

8.5 Rules and Groups

This section specifies the XML constructs that are used, in RIR-PRD, to serialize rules and groups.

8.5.1 RULE

In RIF-PRD, the RULE class of constructs is used to serialize rules, that is, unconditional as well as conditional actions, or rules with bound variables.

As an abstract class, RULE is not associated with specific XML markup in RIF-PRD instance documents.

    [ Implies | Forall | ACTION_BLOCK ]

8.5.1.1 ACTION_BLOCK

An unconditional action block is serialized, in RIF-PRD XML, using the ACTION_BLOCK class of construct.

8.5.1.2 Implies

Conditional actions are serialized, in RIF-PRD, using the XML element Implies.

The Implies element contains an if sub-element and a then sub-element:

    <Implies>
       <if> FORMULA </if>
       <then> ACTION_BLOCK </then>
    </Implies>

8.5.1.3 Forall

The Forall construct is used, in RIF-PRD, to represent rules with bound variables.

The Forall element contains:

    <Forall>
       <declare> Var </declare>+
       <pattern> FORMULA </pattern>*
       <formula> RULE </formula>
    </Forall>

Example 8.12. The example shows the rule variables declaration part of the Gold rule, from the running example, as represented in example 4.2.

<Forall>
   <declare><Var>?customer</Var></declare>
   <pattern>
      <And>
         <formula><Member> ... </Member></formula>
         <formula><Frame> ... </Frame></formula>
      </And>
   </pattern>
   <formula>
      <Forall>
         <declare><Var>?shoppingCart</Var></declare>
         <pattern><Member> ... </Member></pattern>
         <formula>
            <Implies> ... </Implies>
         </formula>
      </Forall>
   </formula>
</Forall>

8.5.2 Group

The Group construct is used to serialize a group.

The Group element has zero or one behavior sub-element and zero or more sentence sub-elements:

    <Group>
       <behavior>
          <ConflictResolution> xsd:anyURI </ConflictResolution>?
          <Priority> -10,000 ≤ xsd:int ≤ 10,000 </Priority>?
       </behavior>?
       <sentence> [ RULE | Group ] </sentence>*
    </Group>

8.6 Document and directives

8.6.1 Import

The Import directive is used to serialize the reference to an RDF graph or an OWL ontology to be combined with a RIF document. The Import directive is inherited from [RIF-Core]. Its abstract syntax and its semantics are specified in [RIF-RDF-OWL].

The Import directive contains:

    <Import>
       <location> xsd:anyURI </location>
       <profile> xsd:anyURI </profile>?
    </Import>

8.6.2 Document

The Document is the root element of any RIF-PRD instance document.

The Document contains zero or more directive sub-elements, each containing an Import directive, and zero or one payload sub-element, that must contain a Group element.

    <Document>
       <directive> Import </Import>*
       <payload> Group </payload>?
    </Document>

The semantics of a document that imports RDF and/or OWL documents is specified in [RIF-RDF-OWL] and [RIF-BLD]. The semantics of a document that does not import other documents is the semantics of the rule set that is serialised by the Group in the document's payload sub-element, if any.


8.7 Constructs carrying no semantics

8.7.1 Annotation

Annotations can be associated with any concrete class element in RIF-PRD: those are the elements with a CamelCase tagname starting with an upper-case character:

    CLASSELT = [ TERM | ATOMIC | FORMULA | ACTION | ACTION_BLOCK | New | RULE | Group | Document | Import ]

An identifier can be associated to any instance element of the abstract CLASSELT class of constructs, as an optional id sub-element that MUST contain a Const of type rif:iri.

Annotations can be included in any instance of a concrete class element using the meta sub-element.

The Frame construct is used to serialize annotations: the content of the Frame's object sub-element identifies the object to which the annotation is associated:, and the Frame's slots represent the annotation properly said as property-value pairs.

If all the annotations are related to the same object, the meta element can contain a single Frame sub-element. If annotations related to several different objects need be serialized, the meta role element can contain an And element with zero or more formula sub-elements, each containing one Frame element, that serializes the annotations relative to one identified object.

   <any concrete element in CLASSELT>
       <id> Const </id>?
       <meta>
          [ Frame 
            |
            <And>
               <formula> Frame </formula>*
            </And>
          ]
       </meta>?
       other CLASSELT content
    </any concrete element in CLASSELT>

Notice that the content of the meta sub-element of an instance of a RIF-PRD class element is not necessarily associated to that same instance element: only the content of the object sub-element of the Frame that represents the annotations specifies what the annotations are about, not where it is included in the instance RIF document.

It is suggested to use Dublin Core, RDFS, and OWL properties for annotations, along the lines of http://www.w3.org/TR/owl-ref/#Annotations -- specifically owl:versionInfo, rdfs:label, rdfs:comment, rdfs:seeAlso, rdfs:isDefinedBy, dc:creator, dc:description, dc:date, and foaf:maker.

Example 8.13. The example shows the structure of the document that contains the runnig example rule set, as represented in example 4.2, including annotations such as rule set and rule names.

<Document>
  <payload>
    <Group>
      <id><Const type="rif:iri">http://example.com/2009/prd2#CheckoutRuleSet</Const></id>
      <meta>
        <Frame>
          <object><Const type="rif:iri">http://example.com/2009/prd2#CheckoutRuleSet</Const></object>
          <slot ordered="yes">
            <Const type="rif:iri">http://dublincore.org/documents/dcmi-namespace/#creator</Const>
            <Const type="xsd:string>W3C RIF WG</Const>
          </slot>
          <slot>
            <Const type="rif:iri">http://dublincore.org/documents/dcmi-namespace/#description</Const>
            <Const type="xsd:string">Running example rule set from the RIF-PRD specification</Const>
          </slot>
        </Frame>
      </meta>
      <behavior> ... </behavior>
      <sentence>
        <Group>
          <id><Const type="rif:iri">http://example.com/2009/prd2#GoldRule</Const></id>
          <behavior> ... </behavior>
          <sentence><Forall> ... </Forall></sentence>
        </Group>
      </sentence>
      <sentence>
        <Group>
          <id><Const type="rif:iri">http://example.com/2009/prd2#DiscountRule</Const></id>
          <sentence><Forall> ... </Forall></sentence>
        </Group>
      </sentence>
    </Group>
  </payload>
</Document>

9 Presentation syntax (Informative)

To make it easier to read, a non-normative, lightweight notation was introduced to complement the mathematical English specification of the abstract syntax and the semantics of RIF-PRD. This section specifies a presentation syntax for RIF-PRD, that extends that notation. The presentation syntax is not normative. However, it may help implementers by providing a more succinct overview of RIF-PRD syntax.

The EBNF for the RIF-PRD presentation syntax is given as follows. For convenience of reading we show the entire EBNF in its four parts (rules, conditions, actions, and annotations).

Rule Language:

  Document       ::= IRIMETA? 'Document' '(' Base? Prefix* Import* Group? ')'
  Base           ::= 'Base' '(' ANGLEBRACKIRI ')'
  Prefix         ::= 'Prefix' '(' Name ANGLEBRACKIRI ')'
  Import         ::= IRIMETA? 'Import' '(' LOCATOR PROFILE? ')'
  Group          ::= IRIMETA? 'Group' Strategy? Priority? '(' (RULE | Group)* ')'
  Strategy       ::= Const
  Priority       ::= Const
  RULE           ::= (IRIMETA? 'Forall' Var+ (' such that ' FORMULA+)?  '(' RULE ')') | CLAUSE
  CLAUSE         ::= Implies | ACTION_BLOCK 
  Implies        ::= IRIMETA? 'If' FORMULA 'Then' ACTION_BLOCK
  LOCATOR        ::= ANGLEBRACKIRI
  PROFILE        ::= ANGLEBRACKIRI

Action Language:

  ACTION  ::= IRIMETA? (Assert | Retract | Modify | Execute )
  Assert         ::= 'Assert' '(' IRIMETA? (Atom | Frame | Member) ')'
  Retract        ::= 'Retract' '(' ( IRIMETA? (Atom | Frame) | TERM | TERM TERM ) ')'
  Modify         ::= 'Modify'  '(' IRIMETA? Frame ')'
  Execute        ::= 'Execute' '(' IRIMETA? Atom ')' 
  ACTION_BLOCK   ::= IRIMETA? ('Do (' ('(' IRIMETA? Var IRIMETA? (Frame | 'New()') ')')* ACTION+  ')' |
                     'And (' ( IRIMETA? (Atom | Frame) )* ')' | Atom | Frame)

Condition Language:

  FORMULA        ::= IRIMETA? 'And' '(' FORMULA* ')' |
                     IRIMETA? 'Or' '(' FORMULA* ')' |
                     IRIMETA? 'Exists' (IRIMETA? Var)+ '(' FORMULA ')' |
                     ATOMIC |
                     IRIMETA? NEGATEDFORMULA  |
                     IRIMETA? Equal |
                     IRIMETA? Member |
                     IRIMETA? Subclass | 
                     IRIMETA? 'External' '(' IRIMETA? Atom ')'
  ATOMIC         ::= IRIMETA? (Atom | Frame)
  Atom           ::= UNITERM
  UNITERM        ::= (IRIMETA? Const) '(' (TERM* ')'
  GROUNDUNITERM  ::= (IRIMETA? Const) '(' (GROUNDTERM* ')'
  NEGATEDFORMULA ::= 'Not' '(' FORMULA ')' | 'INeg' '(' FORMULA ')' 
  Equal          ::= TERM '=' TERM
  Member         ::= TERM '#' TERM
  Subclass       ::= TERM '##' TERM
  Frame          ::= TERM '[' (TERM '->' TERM)* ']'
  TERM           ::= IRIMETA? (Const | Var | List | 'External' '(' Expr ')')
  GROUNDTERM     ::= IRIMETA? (Const | List | 'External' '(' GROUNDUNITERM ')')
  Expr           ::= UNITERM
  List           ::= 'List' '(' GROUNDTERM* ')'
  Const          ::= '"' UNICODESTRING '"^^' SYMSPACE | CONSTSHORT
  Var            ::= '?' Name
  Name           ::= NCName
  SYMSPACE       ::= ANGLEBRACKIRI | CURIE


Annotations:

  IRIMETA        ::= '(*' IRICONST? (Frame | 'And' '(' Frame* ')')? '*)'

A NEGATEDFORMULA can be written using either Not or INeg. INeg is short for inflationary negation and is preferred over 'Not' to avoid ambiguity about the semantics of the negation.

The RIF-PRD presentation syntax does not commit to any particular vocabulary and permits arbitrary Unicode strings in constant symbols, argument names, and variables. Constant symbols can have this form: "UNICODESTRING"^^SYMSPACE, where SYMSPACE is an ANGLEBRACKIRI or CURIE that represents the identifier of the symbol space of the constant, and UNICODESTRING is a Unicode string from the lexical space of that symbol space. ANGLEBRACKIRI and CURIE are defined in the section Shortcuts for Constants in RIF's Presentation Syntax in [RIF-DTB]. Constant symbols can also have several shortcut forms, which are represented by the non-terminal CONSTSHORT. These shortcuts are also defined in the same section of [RIF-DTB]. One of them is the CURIE shortcut, which is extensively used in the examples in this document. Names are XML NCNames. Variables are composed of NCName symbols prefixed with a ?-sign.

Example 9.1. Here is the transcription, in the RIF-PRD presentation syntax, of the complete RIF-PRD document corresponding to the running example:

Document(

 Prefix( ex1 <http://example.com/2009/prd2> )

 (* ex1:CheckoutRuleset *)
 Group rif:forwardChaining (  
   (* ex1:GoldRule *)
   Group 10 (
     Forall ?customer such that (And( ?customer # ex1:Customer
                                      ?customer[status->"Silver"] )) 
       (Forall ?shoppingCart such that (?customer[shoppingCart->?shoppingCart])
          (If Exists ?value (And( ?shoppingCart[value->?value]
                                  pred:numeric-greater-than-or-equal(?value 2000))
           Then Do( Modify( ?customer[status->"Gold"] ) ) ) ) ) )

   (* ex1:DiscountRule *)
   Group (
     Forall ?customer such that (And( ?customer # ex1:Customer ) )
       (If Or( ?customer[status->"Silver"]
               ?customer[status->"Gold"] )
        Then Do( (?s ?customer[shoppingCart->?s])
                 (?v ?s[value->?v])
                 Modify( ?s[value->func:numeric-multiply(?v 0.95)] ) ) ) )

   (* ex1:NewCustomerAndWidgetRule *)
   Group
     Forall ?customer such that (And( ?customer # ex1:Customer
                                      ?customer[status->"New"] ) )
       (If Exists ?shoppingCart ?item
                  ( And ( ?customer[shoppingCart->?shoppingCart]
                          ?shoppingCart[containsItem->?item]
                          ?item # ex1:Widget ) ) )
        Then Do( (?s ?customer[shoppingCart->?s])
                 (?val ?s[value->?val])
                 (?voucher ?customer[voucher->?voucher])
                 Retract( ?customer[voucher->?voucher] )
                 Retract( ?voucher )
                 Modify( ?s[value->func:numeric-multiply(?val 0.90)] ) ) ) )

   (* ex1:UnknownStatusRule *)
   Group (
     Forall ?customer such that ( ?customer # ex1:Customer )
       (If Not(Exists ?status
                      (And( ?customer[status->?status]
                            External(pred:list-contains(List("New", "Bronze", "Silver", "Gold"), ?status)) )))
        Then Do( Execute( act:print(func:concat("New customer: " ?customer)) )
                 Assert( ?customer[status->"New"] ) ) ) ) )
)

10 Acknowledgements

This document is the product of the Rules Interchange Format (RIF) Working Group (see below) whose members deserve recognition for their time and commitment. The editors extend special thanks to Harold Boley and Changhai Ke for their thorough reviews and insightful discussions; the working group chairs, Chris Welty and Christian de Sainte Marie, for their invaluable technical help and inspirational leadership; and W3C staff contact Sandro Hawke, a constant source of ideas, help, and feedback.


The regular attendees at meetings of the Rule Interchange Format (RIF) Working Group at the time of the publication were: Adrian Paschke (Freie Universitaet Berlin), Axel Polleres (DERI), Changhai Ke (IBM), Chris Welty (IBM), Christian de Sainte Marie (IBM), Dave Reynolds (HP), Gary Hallmark (ORACLE), Harold Boley (NRC), Hassan Aït-Kaci (IBM), Jos de Bruijn (FUB), Leora Morgenstern (IBM), Michael Kifer (Stony Brook), Mike Dean (BBN), Sandro Hawke (W3C/MIT), and Stella Mitchell (IBM).

11 References

11.1 Normative references

[OMG-PRR]
Production Rule Representation (PRR), OMG specification, version 1.0, 2007.

[RDF-CONCEPTS]
Resource Description Framework (RDF): Concepts and Abstract Syntax, Klyne G., Carroll J. (Editors), W3C Recommendation, 10 February 2004, http://www.w3.org/TR/2004/REC-rdf-concepts-20040210/. Latest version available at http://www.w3.org/TR/rdf-concepts/.

[RDF-SCHEMA]
RDF Vocabulary Description Language 1.0: RDF Schema, Brian McBride, Editor, W3C Recommendation 10 February 2004, http://www.w3.org/TR/2004/REC-rdf-schema-20040210/. Latest version available at http://www.w3.org/TR/rdf-schema/.

[RFC-3066]
RFC 3066 - Tags for the Identification of Languages, H. Alvestrand, IETF, January 2001, at http://www.ietf.org/rfc/rfc3066.

[RFC-3987]
RFC 3987 - Internationalized Resource Identifiers (IRIs), M. Duerst and M. Suignard, IETF, January 2005, http://www.ietf.org/rfc/rfc3987.txt.

[RIF-BLD]
RIF Basic Logic Dialect Harold Boley, Michael Kifer, eds. W3C Recommendation, 22 June 2010, http://www.w3.org/TR/2010/REC-rif-bld-20100622/. Latest version available at http://www.w3.org/TR/rif-bld/.

[RIF-Core]
RIF Core Dialect Harold Boley, Gary Hallmark, Michael Kifer, Adrian Paschke, Axel Polleres, Dave Reynolds, eds. W3C Recommendation, 22 June 2010, http://www.w3.org/TR/2010/REC-rif-core-20100622/. Latest version available at http://www.w3.org/TR/rif-core/.

[RIF-DTB]
RIF Datatypes and Built-Ins 1.0 Axel Polleres, Harold Boley, Michael Kifer, eds. W3C Recommendation, 22 June 2010, http://www.w3.org/TR/2010/REC-rif-dtb-20100622/. Latest version available at http://www.w3.org/TR/rif-dtb/.

[RIF-FLD]
RIF Framework for Logic Dialects Harold Boley, Michael Kifer, eds. W3C Recommendation, 22 June 2010, http://www.w3.org/TR/2010/REC-rif-fld-20100622/. Latest version available at http://www.w3.org/TR/rif-fld/.

[RIF-RDF-OWL]
RIF RDF and OWL Compatibility Jos de Bruijn, editor. W3C Recommendation, 22 June 2010, http://www.w3.org/TR/2010/REC-rif-rdf-owl-20100622/. Latest version available at http://www.w3.org/TR/rif-rdf-owl/.

[XDM]
XQuery 1.0 and XPath 2.0 Data Model (XDM), W3C Recommendation, World Wide Web Consortium, 23 January 2007. This version is http://www.w3.org/TR/2007/REC-xpath-datamodel-20070123/. Latest version available at http://www.w3.org/TR/xpath-datamodel/.

[XML-Base]
XML Base (second edition), W3C Recommendation, World Wide Web Consortium, 28 January 2009, http://www.w3.org/TR/2009/REC-xmlbase-20090128/. Latest version available at http://www.w3.org/TR/xmlbase/.

[XML-SCHEMA2]
XML Schema Part 2: Datatypes Second Edition, W3C Recommendation, World Wide Web Consortium, 28 October 2004, http://www.w3.org/TR/2004/REC-xmlschema-2-20041028/. Latest version available at http://www.w3.org/TR/xmlschema-2/.

[XPath-Functions]
XQuery 1.0 and XPath 2.0 Functions and Operators, W3C Recommendation, World Wide Web Consortium, 23 January 2007, http://www.w3.org/TR/2007/REC-xpath-functions-20070123/. Latest version available at http://www.w3.org/TR/xpath-functions/.

11.2 Informational references

[CIR04]
Production Systems and Rete Algorithm Formalisation, Cirstea H., Kirchner C., Moossen M., Moreau P.-E. Rapport de recherche n° inria-00280938 - version 1 (2004).

[CURIE]
CURIE Syntax 1.0, M. Birbeck, S. McCarron, W3C Candidate Recommendation, 16 January 2009, http://www.w3.org/TR/2009/CR-curie-20090116. Latest version available at http://www.w3.org/TR/curie/.

[Enderton01]
A Mathematical Introduction to Logic, Second Edition, H. B. Enderton. Academic Press, 2001.

[FIT02]
Fixpoint Semantics for Logic Programming: A Survey, Melvin Fitting, Theoretical Computer Science. Vol. 278, no. 1-2, pp. 25-51. 6 May 2002.

[HAK07]
Data Models as Constraint Systems: A Key to the Semantic Web, Hassan Ait-Kaci, Constraint Programming Letters, 1:33--88, 2007.

[KLW95]
Logical foundations of object-oriented and frame-based languages, M. Kifer, G. Lausen, J. Wu. Journal of ACM, July 1995, pp. 741--843.

[Mendelson97]
Introduction to Mathematical Logic, Fourth Edition, E. Mendelson. Chapman & Hall, 1997.

[PLO04]
A Structural Approach to Operational Semantics, Gordon D. Plotkin, Journal of Logic and Algebraic Programming, Volumes 60-61, Pages 17-139 (July - December 2004).

12 Appendix: Model-theoretic semantics of RIF-PRD condition formulas

This appendix provides an alternative specification of the Semantics of condition formulas, and it is also normative.

This alternative specification is provided for the convenience of the reader, for compatibility with other RIF specifications, such as [RIF-DTB] and [RIF-RDF-OWL], and to make explicit the interoperability with RIF logic dialects, in particular [RIF-Core] and [RIF-BLD].

12.1 Semantic structures

The key concept in a model-theoretic semantics of a logic language is the notion of a semantic structure [Enderton01, Mendelson97].

Definition (Semantic structure). A semantic structure, I, is a tuple of the form <TV, DTS, D, Dind, Dfunc, IC, IV, Ilist, Itail, IP, Iframe, Isub, Iisa, I=, Iexternal, Itruth>. Here D is a non-empty set of elements called the Herbrand domain of I, that is, the set of all ground terms which can be formed by using the elements of Const. Dind, Dfunc are nonempty subsets of D. Dind is used to interpret the elements of Const that are individuals and Dfunc is used to interpret the elements of Const that are function symbols. Const denotes the set of all constant symbols and Var the set of all variable symbols. TV denotes the set of truth values that the semantic structure uses and DTS is a set of identifiers for primitive datatypes (please refer to Section Datatypes in the RIF data types and builtins specification [RIF-DTB] for the semantics of datatypes).

As far as the assignment of a standard meaning to formulas in the RIF-PRD condition language is concerned, the set TV of truth values consists of just two values, t and f.

The other components of I are total mappings defined as follows:

  1. IC maps Const to D.

    This mapping interprets constant symbols. In addition:

  2. IV maps Var to Dind.

    This mapping interprets variable symbols.

  3. Ilist and Itail are used to interpret lists. They are mappings of the following form:

    In addition, these mappings are required to satisfy the following conditions:

    Note that the last condition above restricts Itail only when its last argument is in Ilist(Dind), the image of Ilist. If the last argument of Itail is not in Ilist(Dind), then the list is malformed and there are no restrictions on the value of Itail except that it must be in Dind.

  4. IP maps D to functions D*indD (here D*ind is a set of all sequences of any finite length over the domain Dind).

    This mapping interprets positional terms atoms.

  5. Iframe maps Dind to total functions of the form SetOfFiniteBags(Dind × Dind) → D.

    This mapping interprets frame terms. An argument, d ? Dind, to Iframe represents an object and the finite bag {<a1,v1>, ..., <ak,vk>} represents a bag of attribute-value pairs for d. We will see shortly how Iframe is used to determine the truth valuation of frame terms.

    Bags (multi-sets) are used here because the order of the attribute/value pairs in a frame is immaterial and pairs may repeat. Such repetitions arise naturally when variables are instantiated with constants. For instance, o[?A->?B ?C->?D] becomes o[a->b a->b] if variables ?A and ?C are instantiated with the symbol a while ?B and ?D are instantiated with b. (We shall see later that o[a->b a->b] is equivalent to o[a->b].)

  6. Isub gives meaning to the subclass relationship. It is a mapping of the form Dind × DindD.

    The operator ## is required to be transitive, i.e., c1 ## c2 and c2 ## c3 must imply c1 ## c3. TThis is ensured by a restriction in Section Interpretation of condition formulas;

  7. Iisa gives meaning to class membership. It is a mapping of the form Dind × DindD.

    The relationships # and ## are required to have the usual property that all members of a subclass are also members of the superclass, i.e., o # cl and cl ## scl must imply o # scl. This is ensured by a restriction in Section Interpretation of condition formulas;

  8. I= is a mapping of the form Dind × DindD.

    It gives meaning to the equality operator.

  9. Itruth is a mapping of the form DTV.

    It is used to define truth valuation for formulas.

  10. Iexternal is a mapping from the coherent set of schemas for externally defined functions to total functions D* → D. For each external schema σ = (?X1 ... ?Xn; τ) in the coherent set of external schemas associated with the language, Iexternal(σ) is a function of the form DnD.

    For every external schema, σ, associated with the language, Iexternal(σ) is assumed to be specified externally in some document (hence the name external schema). In particular, if σ is a schema of a RIF built-in predicate, function or action, Iexternal(σ) is specified so that:

For convenience, we also define the following mapping I from terms to D:

The effect of datatypes. The set DTS must include the datatypes described in Section Primitive Datatypes of the RIF data types and builtins specification [RIF-DTB].

The datatype identifiers in DTS impose the following restrictions. Given dtDTS, let LSdt denote the lexical space of dt, VSdt denote its value space, and Ldt: LSdtVSdt the lexical-to-value-space mapping (for the definitions of these concepts, see Section Primitive Datatypes of the RIF data types and builtins specification [RIF-DTB]. Then the following must hold:

That is, IC must map the constants of a datatype dt in accordance with Ldt.

RIF-PRD does not impose restrictions on IC for constants in symbol spaces that are not datatypes included in DTS.   ☐

12.2 Interpretation of condition formulas

This section defines how a semantic structure, I, determines the truth value TValI(φ) of a condition formula, φ.

We define a mapping, TValI, from the set of all condition formulas to TV. Note that the definition implies that TValI(φ) is defined only if the set DTS of the datatypes of I includes all the datatypes mentioned in φ and Iexternal is defined on all externally defined functions and predicates in φ.

Definition (Truth valuation). Truth valuation for well-formed condition formulas in RIF-PRD is determined using the following function, denoted TValI:

12.3 Condition satisfaction

We define, now, what it means for a state of the fact base to satisfy a condition formula. The satisfaction of condition formulas in a state of the fact base provides formal underpinning to the operational semantics of rule sets interchanged using RIF-PRD.

Definition (Models). A semantic structure I is a model of a condition formula, φ, written as I |= φ, iff TValI(φ) = t.   ☐


Definition (Herbrand interpretation). Given a non-empty set of constants, Const, the Herbrand domain is the set of all the ground terms that can be formed using the elements of Const, and the Herbrand base is the set of all the well-formed ground atomic formulas that can be formed with the elements in the Herbrand domain.

A semantic structure, I, is a Herbrand interpretation, if the set of all the ground formulas which are true with respect to I (that is, of which I is a model), is a subset of the corresponding Herbrand base, BI.   ☐

In RIF-PRD, the semantics of condition formulas is defined with respect to semantic structures where the domain, D is the Herbrand domain that is determined by the set of all the constants, Const; that is, with respect to Herbrand interpretations.

Definition (State of the fact base). To every Herbrand interpretation I, we associate a state of the fact base, wI, that is represented by the subset of the Herbrand base that contains all the ground atomic formulas of which I is a model; or, equivalently, by the conjunction of all these ground atomic formulas.   ☐

Definition (Condition satisfaction). A RIF-PRD condition formula φ is satisfied in a state of the fact base, wI, if and only if I is a model of φ.   ☐

At the syntactic level, the interpretation of the variables by a valuation function IV is realized by a substitution. As a consequence, a ground substitution σ matches a condition formula ψ to a set of ground atomic formulas Φ if and only if σ realizes the valuation function IV of a semantics structure I that is a model of ψ and Φ is a representation of a state of the fact base, wI (as defined above), that is associated to I; that is, if and only if ψ is satisfied in wI (as defined above).

This provides the formal link between the satisfaction of a condition formula, as defined above, and a matching substitution, and, followingly, between the alternative definitions of a state of facts and the satisfaction of a condition, here and in section Semantics of condition formulas.


13 Appendix: XML schema

The RIF-PRD XML Schema is specified below as a redefinition and an extension of the RIF-Core XML Schema [RIF-Core] and is also available at http://www.w3.org/2010/rif-schema/prd/.


<?xml version="1.0" encoding="UTF-8"?>

<xs:schema
    targetNamespace="http://www.w3.org/2007/rif#"
    xmlns:xs="http://www.w3.org/2001/XMLSchema"
    xmlns:xml="http://www.w3.org/XML/1998/namespace"
    xmlns="http://www.w3.org/2007/rif#"
    elementFormDefault="qualified">

<xs:import namespace='http://www.w3.org/XML/1998/namespace'
           schemaLocation='http://www.w3.org/2001/xml.xsd'/>


<!-- ================================================== -->
<!-- Redefine some elements in the Core Conditions      -->
<!-- Extension of the choice                            -->
<!-- ================================================== -->

   <xs:group name="ATOMIC">
      <xs:choice>
         <xs:element ref="Atom"/>
         <xs:element ref="Frame"/>
         <xs:element ref="Member"/>
         <xs:element ref="Equal"/>
         <xs:element ref="Subclass"/> <!-- Subclass is not in RIF-Core -->
         <xs:element name="External" type="External-FORMULA.type"/>
      </xs:choice>
   </xs:group>

   <xs:group name="FORMULA">
      <xs:choice>
         <xs:group   ref="ATOMIC"/> 
         <xs:element ref="And"/>
         <xs:element ref="Or"/>
         <xs:element ref="Exists"/>
         <xs:element ref="INeg"/> <!-- INeg is nt in RIF-Core -->
      </xs:choice>
   </xs:group>

<!-- ================================================== -->
<!-- Additional elements to the Core Condition schema   -->
<!-- ================================================== -->

   <xs:element name="Subclass">
<!--                                                    -->
<!--     <Subclass>                                     -->
<!--        <sub> TERM </sub>                           -->
<!--        <super> TERM </super>                       -->
<!--     </Subclass>                                    -->
<!--                                                    -->
      <xs:complexType>
         <xs:sequence>
            <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
            <xs:element name="sub">
               <xs:complexType>
                  <xs:group ref="TERM" minOccurs="1" maxOccurs="1"/>
               </xs:complexType>
            </xs:element>
            <xs:element name="super">
               <xs:complexType>
                  <xs:group ref="TERM" minOccurs="1" maxOccurs="1"/>
               </xs:complexType>
            </xs:element>
         </xs:sequence>
      </xs:complexType>
   </xs:element>

   <xs:element name="INeg">
<!--                                                    -->
<!--     <INeg>                                         -->
<!--        <formula> FORMULA </formula>                -->
<!--     </INeg>                                        -->
<!--                                                    -->
      <xs:complexType>
        <xs:sequence>
          <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
          <xs:element ref="formula" minOccurs="1" maxOccurs="1"/>
        </xs:sequence>
      </xs:complexType>
   </xs:element>

<!-- ============================================ -->
<!-- CoreCond.xsd starts here                     -->
<!-- ============================================ -->

  <xs:complexType name="External-FORMULA.type">
    <!-- sensitive to FORMULA (Atom) context-->
    <xs:sequence>
      <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      <xs:element name="content" type="content-FORMULA.type"/>
    </xs:sequence>
  </xs:complexType>
  
  <xs:complexType name="content-FORMULA.type">
    <!-- sensitive to FORMULA (Atom) context-->
    <xs:sequence>
      <xs:element ref="Atom"/>
    </xs:sequence>
  </xs:complexType>
 
  <xs:element name="And">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="formula" minOccurs="0" maxOccurs="unbounded"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="Or">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="formula" minOccurs="0" maxOccurs="unbounded"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="Exists">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="declare" minOccurs="1" maxOccurs="unbounded"/>
        <xs:element ref="formula"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="formula">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="FORMULA"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="declare">
    <xs:complexType>
      <xs:sequence>
        <xs:element ref="Var"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="Atom">
    <!--
  Atom           ::= UNITERM
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="UNITERM"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>  
  
  <xs:group name="UNITERM">
    <!--
  UNITERM        ::= Const '(' (TERM* ')'
    -->
    <xs:sequence>
      <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      <xs:element ref="op"/>
      <xs:element name="args" type="args-UNITERM.type" minOccurs="0" maxOccurs="1"/>
    </xs:sequence>
  </xs:group>
 
  <xs:group name="GROUNDUNITERM">
    <!-- sensitive to ground terms
  GROUNDUNITERM        ::= Const '(' (GROUNDTERM* ')'
    -->
    <xs:sequence>
      <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      <xs:element ref="op"/>
      <xs:element name="args" type="args-GROUNDUNITERM.type" minOccurs="0" maxOccurs="1"/>
    </xs:sequence>
  </xs:group>


  <xs:element name="op">
    <xs:complexType>
      <xs:sequence>
        <xs:element ref="Const"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:complexType name="args-UNITERM.type">
    <!-- sensitive to UNITERM (TERM) context-->
    <xs:sequence>
      <xs:group ref="TERM" minOccurs="1" maxOccurs="unbounded"/>
    </xs:sequence>
    <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
  </xs:complexType>


  <xs:complexType name="args-GROUNDUNITERM.type">
    <!-- sensitive to GROUNDUNITERM (TERM) context-->
    <xs:sequence>
      <xs:group ref="GROUNDTERM" minOccurs="1" maxOccurs="unbounded"/>
    </xs:sequence>
    <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
  </xs:complexType>

  <xs:element name="Equal">
    <!--
  Equal          ::= TERM '=' ( TERM | IRIMETA? 'External' '(' Expr ')' )
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="left"/>
        <xs:element ref="right"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="left">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="TERM"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="right">
    <xs:complexType>
        <xs:group ref="TERM"/>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="Member">
    <!--
  Member         ::= TERM '#' TERM
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="instance"/>
        <xs:element ref="class"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>

  <xs:element name="instance">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="TERM"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="class">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="TERM"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="Frame">
    <!--
  Frame          ::= TERM '[' (TERM '->' TERM)* ']'
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="object"/>
        <xs:element name="slot" type="slot-Frame.type" minOccurs="0" maxOccurs="unbounded"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="object">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="TERM"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:complexType name="slot-Frame.type">
    <!-- sensitive to Frame (TERM) context-->
    <xs:sequence>
      <xs:group ref="TERM"/>
      <xs:group ref="TERM"/>
    </xs:sequence>
    <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
  </xs:complexType>
 
  <xs:group name="TERM">  
    <!--
  TERM           ::= IRIMETA? (Const | Var | External | List )
    -->
      <xs:choice>
         <xs:element ref="Const"/>
         <xs:element ref="Var"/>
         <xs:element name="External" type="External-TERM.type"/>
         <xs:element ref="List"/>
      </xs:choice>
  </xs:group>

  <xs:group name="GROUNDTERM">  
    <!--
  GROUNDTERM           ::= IRIMETA? (Const | List | 'External' '(' 'Expr' '(' GROUNDUNITERM ')' ')')
    -->
      <xs:choice>
         <xs:element ref="Const"/>
         <xs:element ref="List"/>
         <xs:element name="External" type="External-GROUNDUNITERM.type"/>
      </xs:choice>
  </xs:group>
  
  <xs:element name="List">
    <!--
      List           ::= 'List' '(' GROUNDTERM* ')'
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/> 
        <xs:element name="items"> 
          <xs:complexType> 
            <xs:sequence>
              <xs:group ref="GROUNDTERM" minOccurs="0" maxOccurs="unbounded"/>
            </xs:sequence>
            <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
          </xs:complexType>
        </xs:element> 
      </xs:sequence>
    </xs:complexType>
  </xs:element> 
  
  <xs:complexType name="External-TERM.type">
    <!-- sensitive to TERM (Expr) context-->
    <xs:sequence>
      <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      <xs:element name="content" type="content-TERM.type"/>
    </xs:sequence>
  </xs:complexType>
  

  <xs:complexType name="External-GROUNDUNITERM.type">
    <!-- sensitive to GROUNDTERM (Expr) context-->
    <xs:sequence>
      <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      <xs:element name="content" type="content-GROUNDUNITERM.type"/>
    </xs:sequence>
  </xs:complexType>


  <xs:complexType name="content-TERM.type">
    <!-- sensitive to TERM (Expr) context-->
    <xs:sequence>
      <xs:element ref="Expr"/>
    </xs:sequence>
  </xs:complexType>

   
   <xs:complexType name="content-GROUNDUNITERM.type">
    <!-- sensitive to GROUNDTERM (Expr) context-->
    <xs:sequence>
      <xs:element name="Expr" type="content-GROUNDEXPR.type"/>
    </xs:sequence>
  </xs:complexType>

  <xs:complexType name="content-GROUNDEXPR.type">
    <!-- sensitive to GROUNDEXPR context-->
    <xs:sequence>
      <xs:group ref="GROUNDUNITERM"/>
    </xs:sequence>
  </xs:complexType>
  <xs:element name="Expr">
    <!--
  Expr           ::= Const '(' (TERM | IRIMETA? 'External' '(' Expr ')')* ')'
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:element ref="op"/>
        <xs:element name="args" type="args-Expr.type" minOccurs="0" maxOccurs="1"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>

  <xs:complexType name="args-Expr.type">
    <!-- sensitive to Expr (TERM) context-->
    <xs:choice minOccurs="1" maxOccurs="unbounded">
      <xs:group ref="TERM"/>
    </xs:choice>
    <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
  </xs:complexType>

  <xs:element name="Const">
    <!--
  Const          ::= '"' UNICODESTRING '"^^' SYMSPACE | CONSTSHORT
    -->
    <xs:complexType mixed="true">
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      </xs:sequence>
      <xs:attribute name="type" type="xs:anyURI" use="required"/>
      <xs:attribute ref="xml:lang"/>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="Var">
    <!--
  Var            ::= '?' NCName
    -->
    <xs:complexType mixed="true">
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:group name="IRIMETA">
    <!--
  IRIMETA   ::= '(*' IRICONST? (Frame | 'And' '(' Frame* ')')? '*)'
    -->
    <xs:sequence>
      <xs:element ref="id" minOccurs="0" maxOccurs="1"/>
      <xs:element ref="meta" minOccurs="0" maxOccurs="1"/>
    </xs:sequence>
  </xs:group>
 
  <xs:element name="id">
    <xs:complexType>
      <xs:sequence>
        <xs:element name="Const" type="IRICONST.type"/>   <!-- type="&rif;iri" -->
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="meta">
    <xs:complexType>
     <xs:choice>
       <xs:element ref="Frame"/>
       <xs:element name="And" type="And-meta.type"/>
     </xs:choice>
    </xs:complexType>
  </xs:element>
  
  <xs:complexType name="And-meta.type">
  <!-- sensitive to meta (Frame) context-->
    <xs:sequence>
      <xs:element name="formula" type="formula-meta.type" minOccurs="0" maxOccurs="unbounded"/>
    </xs:sequence>
  </xs:complexType>
 
  <xs:complexType name="formula-meta.type">
    <!-- sensitive to meta (Frame) context-->
    <xs:sequence>
      <xs:element ref="Frame"/>
    </xs:sequence>
  </xs:complexType>
  
  <xs:complexType name="IRICONST.type" mixed="true">
    <!-- sensitive to location/id context-->
    <xs:sequence/>
    <xs:attribute name="type" type="xs:anyURI" use="required" fixed="http://www.w3.org/2007/rif#iri"/>
  </xs:complexType>

<!-- ============================================ -->
<!-- Definition of the actions (not in RIF-Core)  -->
<!-- ============================================ -->

   <xs:element name="New">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>


   <xs:group name="INITIALIZATION">
      <xs:choice>
         <xs:element ref="New"/>
         <xs:element ref="Frame"/>
      </xs:choice>
   </xs:group>

   <xs:element name="Do">
<!--                                                    -->
<!--     <Do>                                           -->
<!--        <actionVar ordered="yes">                   -->
<!--          Var                                       -->
<!--          INITIALIZATION                            -->
<!--        </actionVar>*                               -->
<!--        <actions ordered="yes">                     -->
<!--          ACTION+                            -->
<!--        </actions>                                  -->
<!--     </Do>                                          -->
<!--                                                    -->
      <xs:complexType>
         <xs:sequence>
            <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
            <xs:element name="actionVar" minOccurs="0" maxOccurs="unbounded">
               <xs:complexType>
                  <xs:sequence>
                     <xs:element ref="Var" minOccurs="1" maxOccurs="1"/>
                     <xs:group ref="INITIALIZATION" minOccurs="1" maxOccurs="1"/>
                  </xs:sequence>
                  <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
               </xs:complexType>
            </xs:element>
            <xs:element name="actions" minOccurs="1" maxOccurs="1">
               <xs:complexType>
                  <xs:sequence>
                     <xs:group ref="ACTION" minOccurs="1" maxOccurs="unbounded"/>
                  </xs:sequence>
                  <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
               </xs:complexType>
            </xs:element>
         </xs:sequence>
      </xs:complexType>
   </xs:element>

   <xs:group name="ACTION">
      <xs:choice>
         <xs:element ref="Assert"/>
         <xs:element ref="Retract"/>
         <xs:element ref="Modify"/>
         <xs:element ref="Execute"/>
      </xs:choice>	
   </xs:group>

   <xs:element name="Assert">
<!--                                                    -->
<!--     <Assert>                                       -->
<!--        <target> [ Atom                             -->
<!--                 | Frame                            -->
<!--                 | Member ]                         -->
<!--        </target>                                   -->
<!--     </Assert>                                      -->
<!--                                                    -->
      <xs:complexType>
         <xs:sequence>
            <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
            <xs:element name="target" minOccurs="1" maxOccurs="1">
               <xs:complexType>
                  <xs:choice>
                     <xs:element ref="Atom"/>
                     <xs:element ref="Frame"/>
                     <xs:element ref="Member"/>
                  </xs:choice>
               </xs:complexType>
	    </xs:element>
	 </xs:sequence>
     </xs:complexType>
   </xs:element>

   <xs:element name="Retract">
<!--                                                    -->
<!--     <Retract>                                      -->
<!--        <target ordered="yes"?>                     -->
<!--           [ Atom                                   -->
<!--             | Frame                                -->
<!--             | TERM                                 -->
<!--             | TERM TERM ]                          -->
<!--        </target>                                   -->
<!--     </Assert>                                      -->
<!--                                                    -->
      <xs:complexType>
         <xs:sequence>
            <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
            <xs:element name="target" minOccurs="1" maxOccurs="1">
               <xs:complexType>
                  <xs:choice>
                     <xs:element ref="Atom"/>
                     <xs:element ref="Frame"/>
                     <xs:group ref="TERM"/>
                     <xs:sequence>
                        <xs:group ref="TERM"/>
                        <xs:group ref="TERM"/>
                     </xs:sequence>
                  </xs:choice>
                  <xs:attribute name="ordered" type="xs:string" fixed="yes"/>
               </xs:complexType>
            </xs:element>
         </xs:sequence>
      </xs:complexType>
   </xs:element>

   <xs:element name="Modify">
<!--                                                    -->
<!--     <Modify>                                       -->
<!--        <target> Frame </target>                    -->
<!--     </Modify>                                      -->
<!--                                                    -->
     <xs:complexType>
      <xs:sequence>
       <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
       <xs:element name="target" minOccurs="1" maxOccurs="1">
        <xs:complexType>
         <xs:sequence>
          <xs:element ref="Frame"/>
         </xs:sequence>  
        </xs:complexType>
       </xs:element>
      </xs:sequence> 
     </xs:complexType>           
   </xs:element>
   
   <xs:element name="Execute">
<!--                                                    -->
<!--     <Execute>                                      -->
<!--        <target> Atom </target>                     -->
<!--     </Execute>                                     -->
<!--                                                    -->
     <xs:complexType>
      <xs:sequence>
       <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
       <xs:element name="target" minOccurs="1" maxOccurs="1">
        <xs:complexType>
         <xs:sequence>
          <xs:element ref="Atom"/>
         </xs:sequence>  
        </xs:complexType>
       </xs:element>
      </xs:sequence> 
     </xs:complexType>           
   </xs:element>    

<!-- ================================================== -->
<!-- Redefine Group related Core construct              -->
<!-- ================================================== -->

   <xs:complexType name="Group-contents">
      <xs:sequence>
         <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
         <xs:element ref="behavior" minOccurs="0" maxOccurs="1"/>
                     <!-- behavior in not in RIF-Core -->
         <xs:element ref="sentence" minOccurs="0" maxOccurs="unbounded"/>          
      </xs:sequence>
   </xs:complexType>

<!-- ================================================== -->
<!-- Group related addition                             -->
<!-- ================================================== -->

   <xs:element name="behavior">
<!--                                                    -->
<!--     <behavior>                                     -->
<!--        <ConflictResolution>                        -->
<!--           xsd:anyURI                               -->
<!--        <ConflictResolution>?                       -->
<!--        <Priority>                                  -->
<!--           -10,000 ≤ xsd:int ≤ 10,000         -->
<!--        </Priority>?                                -->
<!--     </behavior>                                    -->
<!--                                                    -->
      <xs:complexType>
         <xs:sequence>
            <xs:element name="ConflictResolution" minOccurs="0" maxOccurs="1" type="xs:anyURI"/>
            <xs:element name="Priority" minOccurs="0" maxOccurs="1">
               <xs:simpleType>
                  <xs:restriction base="xs:int">
                     <xs:minInclusive value="-10000"/>
                     <xs:maxInclusive value="10000"/>
                  </xs:restriction>
               </xs:simpleType>
            </xs:element>
         </xs:sequence>
      </xs:complexType>
   </xs:element>

<!-- ================================================== -->
<!-- Redefine rule related Core constructs              -->
<!-- ================================================== -->

  <xs:group name="RULE">
    <!--
  RULE      ::= (IRIMETA? 'Forall' Var+ '(' CLAUSE ')') | CLAUSE
    -->
    <xs:choice>
       <xs:element name="Forall" type="Forall-premises"/>
       <xs:group   ref="CLAUSE"/> 
    </xs:choice>
  </xs:group>

  <xs:complexType name="Forall-premises">
    <xs:sequence>
      <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
      <xs:element ref="declare" minOccurs="1" maxOccurs="unbounded"/>
      <xs:element name="pattern" minOccurs="0" maxOccurs="unbounded">
        <xs:complexType>
          <xs:sequence>
            <xs:group ref="FORMULA"/>
          </xs:sequence>
        </xs:complexType>
      </xs:element> 
      <!-- different from formula in And, Or and Exists -->
      <xs:element name="formula">
        <xs:complexType>
          <xs:group ref="RULE"/>
        </xs:complexType>
      </xs:element>
    </xs:sequence>
  </xs:complexType>

  <xs:group name="CLAUSE">  
    <!--
  CLAUSE   ::= Implies | ACTION_BLOCK
    -->
    <xs:choice>
      <xs:element ref="Implies"/>
      <xs:group ref="ACTION_BLOCK"/> 
    </xs:choice>
  </xs:group>

   <xs:complexType name="then-part">
            <xs:group ref="ACTION_BLOCK" minOccurs="1" maxOccurs="1"/>
   </xs:complexType>

<!-- ================================================== -->
<!-- Rule related additions                             -->
<!-- ================================================== -->
  
  <xs:group name="ACTION_BLOCK">
    <!-- 
  ACTION_BLOCK   ::= 'Do (' (Var (Frame | 'New'))* ACTION+  ')' |
                     'And (' (Atom | Frame)* ')' | Atom | Frame 
    -->                   
    <xs:choice>
     <xs:element  ref="Do"/>
     <xs:element  name="And" type="And-then.type"/>
     <xs:element  ref="Atom"/>
     <xs:element  ref="Frame"/>
    </xs:choice>
  </xs:group>

<!-- ================================================== -->
<!-- CoreRule.xsd starts here                           -->
<!-- ================================================== -->
 
  <xs:element name="Document">
    <!--
  Document  ::= IRIMETA? 'Document' '(' Base? Prefix* Import* Group? ')'
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="directive" minOccurs="0" maxOccurs="unbounded"/>
        <xs:element ref="payload" minOccurs="0" maxOccurs="1"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="directive">
    <!--
  Base and Prefix represented directly in XML
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:element ref="Import"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="payload">
    <xs:complexType>
      <xs:sequence>
        <xs:element name="Group" type="Group-contents"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
  
  <xs:element name="Import">
    <!--
  Import    ::= IRIMETA? 'Import' '(' IRICONST PROFILE? ')'
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/> 
        <xs:element ref="location"/>
        <xs:element ref="profile" minOccurs="0" maxOccurs="1"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="location" type="xs:anyURI"/>
 
  <xs:element name="profile" type="xs:anyURI"/>

  <xs:element name="sentence">
   <xs:complexType>
     <xs:choice>
       <xs:element name="Group" type="Group-contents"/>
       <xs:group ref="RULE"/>
       </xs:choice>
   </xs:complexType>
  </xs:element>
      
  <xs:element name="Implies">
    <!--
  Implies   ::= IRIMETA? ATOMIC  ':-' FORMULA
    -->
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="IRIMETA" minOccurs="0" maxOccurs="1"/>
        <xs:element ref="if"/>
        <xs:element name="then" type="then-part"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>
 
  <xs:element name="if">
    <xs:complexType>
      <xs:sequence>
        <xs:group ref="FORMULA"/>
      </xs:sequence>
    </xs:complexType>
  </xs:element>

  <xs:complexType name="And-then.type">
    <!-- sensitive to then (And) context-->
    <xs:sequence>
      <xs:element name="formula" type="formula-then.type" minOccurs="0" maxOccurs="unbounded"/>
    </xs:sequence>
  </xs:complexType>
 
  <xs:complexType name="formula-then.type">
    <!-- sensitive to then (And) context-->
    <xs:choice>
     <xs:element  ref="Atom"/>
     <xs:element  ref="Frame"/>
    </xs:choice>
  </xs:complexType>

 </xs:schema> 

14 Appendix: Change Log (Informative)

This appendix summarizes the main changes to this document since its publication as a Candidate Recommendation.

Changes since the draft of October 1st, 2009:

Changes since the draft of February 12th, 2010: