CycL is Cyc's knowledge representation language. However, CycL is really two separate sublanguages, the EL (Epistemological Level) and the HL (Heuristic Level). This section will explain the motivation for this distinction and how it affects ontology development.
6.1.1. What is the motivation for the EL/HL distinction? Why isn't one knowledge representation language good enough?
The two most important criteria for a knowledge representation language are expressiveness and efficiency. Expressiveness is the most important criterion for entering knowledge. Efficiency is the most important criterion for making inferences over that knowledge. Expressiveness in a representation language is the ability to express facts, rules, and relationships quickly, easily, and in a principled and factored way. Efficiency in a representation language is the ability to conclude new facts from existing facts as quickly and easily as possible, with as much coverage as possible.
Most programming languages, for example C++, would have a high efficiency but a low expressiveness if used as a knowledge representation language. The efficiency would be high because there would be a tight coupling between the knowledge and the program inferencing over that knowledge, and the knowledge would be in an easily machine-readable form. The expressiveness would be low because all ontologists or knowledge enterers would have to be C++ programmers, and they would all be constrained to express only those assertions which can be programmed in C++.
Most natural languages, for example English, would have a high expressiveness but a low efficiency if used as a knowledge representation language. The expressiveness would be high because it is quick and easy to express facts, rules, and relationships in English. The efficiency would be low because it would be almost impossible to reason about English sentences to conclude new English sentences, since natural language is so complex, ambiguous, and computationally inefficient.
The following is a graphical representation of the tradeoff between expressiveness and efficiency.
If we chose any one language as Cyc's representation language, we would be committing to a particular tradeoff between expressiveness and efficiency. However, if we choose two knowledge representation languages, one with a high expressiveness (the EL) and the other with a high efficiency (the HL), we can have both expressiveness and efficiency, and we don't have to trade one for the other. Separating the EL from the HL benefits both the ontologists, who need only learn a stable representation language, and the system designers, who are free to modify the internal data structures as needed without changing the EL and thus without disrupting the efforts of the ontologists.
The EL, or Epistemological Level, is the knowledge representation language optimized for entering knowledge into Cyc and presenting it to users. It is what you see when you are using the Cyc Browser or most other modes of interaction with Cyc. It is the "face" of CycL that you are already familiar with from earlier sections of this document.
The HL, or Heuristic Level, differs from the EL in the following ways:
- Reifiable NAUTs (non-atomic reified terms) are converted to NARTs (non-atomic reified terms).
- HL assertions are efficiently indexed.
- Non-embedded quantifiers and logical operators are eliminated, and replaced with a (possibly skolemized) list of CNF (conjuctive normal form) clauses.
- Arguments to commutative relations are sorted into a canonical order.
- ...and various other differences, many of which are idiosyncratic to Cyc.
The goal of the HL is maximum efficiency without losing important information. Unimportant information, like the order of arguments to commutative relations, is lost with the conversion from EL to HL. See the section on the Canonicalizer for more information on the difference between EL and HL.
note: This should eventually replace all but the two introductory paragraphs of http://localhost/cycdoc/ref/canon.html
The Canonicalizer is the module in Cyc which converts from EL to HL. Having two separate representation languages means that we have to have a way to convert between the two. The Canonicalizer converts from EL to HL, and the Uncanonicalizer converts back from HL to EL. The Canonicalizer is important because it makes the EL/HL distinction possible. Without the Canonicalizer, we would be stuck with choosing a particular tradeoff between expressiveness and efficiency. With the Canonicalizer, we can have both. See the section on the EL/HL Distinction for more information on EL and HL, and the motivation for making the separation.
The diagram above shows how all access to the KB, input and output, is mitigated by the Inference Engine, specifically the Canonicalizer.
- compression:mapping logically equivalent EL expressions into a common HL construct
- correctness:only equivalent EL expressions are mapped into a common HL construct
The raison d'etre for the Canonicalizer is to translate soundly equivalent EL expressions into a single HL construct. This avoids redundantly adding one expression that is simply a rephrasing of another expression. For example, imagine CycL contains the predicate #$sociallyDominates. The following are all equivalent EL sentences that map into a single HL sentence:
(implies (and (isa ?X Dog) (isa ?Y Dog) (different ?X ?Y)) (or (sociallyDominates ?X ?Y) (sociallyDominates ?Y ?X))) (implies (and (isa ?X Dog) (isa ?Y Dog) (different ?Y ?X)) (or (sociallyDominates ?X ?Y) (sociallyDominates ?Y ?X))) (implies (and (isa ?Y Dog) (isa ?X Dog) (different ?X ?Y)) (or (sociallyDominates ?X ?Y) (sociallyDominates ?Y ?X))) (implies (and (isa ?Y Dog) (isa ?X Dog) (different ?Y ?X)) (or (sociallyDominates ?Y ?X) (sociallyDominates ?X ?Y))) (or (sociallyDominates ?X ?Y) (sociallyDominates ?Y ?X) (not (isa ?X Dog)) (not (isa ?Y Dog)) (not (different ?X ?Y)))
The second differs from the first only in the order of the arguments to the symmetric predicate #$different. The third differs from the first only in the order of the antecedent literals. This also involves canonicalizing the order of arguments to a commutative relation, in this case #$and. The fourth differs from the first only by variable naming. The fifth differs from the first by a few applications of standard logical transformations (e.g., DeMorgan's Law). Attempting to add all five EL sentences to the Cyc KB results in only one HL construct.
An important secondary objective of the Canonicalizer is speed; it is not uncommon for users to load transcripts containing thousands of EL sentences, each of which must be canonicalized. Therefore, canonicalization must be as fast as possible. However, the goal of efficient translation is directly at odds with the goal of detecting and removing duplication (which can involve arbitrary theorem proving). To balance this trade-off, the Canonicalizer uses a limited set of rules for simplifying and transforming sentences that has been tuned with hundreds of thousands of examples; not all equivalent EL expressions simplify to the same HL expression. These rules reflect a pragmatic compromise between efficiently translating EL expressions into HL constructs and detecting and eliminating duplication, and they are truth-preserving. Thus the Canonicalizer does not (cannot) guarantee to map all equivalent EL sentences into a single HL construct; however, it can guarantee soundness (only truly equivalent sentences are mapped into a common HL construct), and it can guarantee a relatively fast response.
Most of the time, it doesn't affect it at all, because no information is lost or altered in the conversion from EL to HL and back to EL. However, sometimes a knowledge enterer will assert something which canonicalizes in an unexpected way, and it's good to know a little bit about the workings of the Canonicalizer so as to understand why and how it's doing what it's doing.
Another important thing to remember about the Canonicalizer is that its behavior is KB-dependent, as well as microtheory-dependent. The Canonicalizer uses existing assertions in the KB to determine how to canonicalize new inputs. For example, if you created a predicate #$foo, and you asserted the following two EL sentences in order:
(isa foo CommutativeRelation) (foo 3 2 1)
then the second one would canonicalize as
(foo 1 2 3)
If you did not assert the commutativity sentence, or asserted it in a microtheory inaccessible to the second assertion, then the Canonicalizer would not reorder the arguments of #$foo.
Here is an overview of what the canonicalizer does:
- Well-formedness checking (wff-checking)
- Expands EL relations
- Evaluates EL-evaluatable relations (see #$evaluateAtEL)
- Converts #$exceptFor and #$exceptWhen into #$abnormal
- Expands implications
- Pushes negations inwards
- Standardizes variables
- Converts existentials to Skolem functions
- Removes leading universals
- Pushes disjunctions inwards (or conjunctions, if a query)
- Clause Canonicalization
- Polycanonicalizes if necessary
- Sorts arguments of commutative relations, including #$and and #$or
- Converts some EL objects to HL objects, e.g. reifiable nauts and EL variables
- Adds #$termOfUnit and #$evaluate literals when necessary
- Simplification is performed at various points during canonicalization
Clausification is a standard logical algorithm implemented by the Canonicalizer. However, special support for particular syntactic constructs require special processing. The precanonicalizer handles those special transformations that are easiest to accomplish prior to conversion to clausal normal form; the clause canonicalizer handles those special transformations easiest to accomplish after conversion to clausal normal form.
Many of these operations have perfect inverses performed by the Uncanonicalizer, so they are invisible at the EL, assuming that both the Canonicalizer and Uncanonicalizer are in working order. But some of the transformations do have a projection onto the EL, so those will be covered in the following sections.
The wff-checker imposes user-defined constraints on all input to Cyc, including assertions and queries. Tighter constraints are imposed on assertions than queries. This is important from the OE perspective because the canonicalizer will not allow you to assert ill-formed CycL.
The input in question must be syntactically well-formed, i.e. it must be a CycL sentence. What it means to be a CycL sentence has already been covered in the CycL section.
The CycL sentence and all of its subexpressions must meet all applicable arity constraints. Relations in CycL are declared to be of either fixed or variable arity. The arity of fixed arity relations is declared and constant (but may be microtheory specific). Variable arity relations (e.g., #$and, #$different) accept zero or more arguments, but may be constrained by the predicates #$arityMin and #$arityMax.
The wff-checker ensures the semantic constraint that each relation referenced in a formula has a declared arity (variable or a non-negative integer) and that the number of arguments satisfies the arity.
If the sentence is being asked as a query, then these constraints are all that are imposed. If the sentence is being asserted, then the following constraints are also imposed:
The CycL sentence and all of its subexpressions must meet all applicable argument constraints. All predicates used to state argument constraints are instances of #$ArgConstraintPredicate. Two important things to remember are that argument constraints conjoin, and that they are inherited from genlPreds. However, it is good OE practice to state the exact argument constraints for every predicate, and not to depend on this inheritance. This is for modularity's sake in case the genlPreds hierarchy changes. Also, it has the effect that you only have to look in one place to see the constraints; you don't also have to look at all the genlPreds.
An important distinction in wff-checking is whether the term in question must provably satisfy the constraint, or whether it must merely be consistent with the constraint (e.g. not disjoint). Generally the decision of which criteria to apply depends on the type of term being checked; whether it is a fort or a naut. Quoted from #$ArgumentTypeChecking-FORTSvsNonReifiedNATS:
"Arg-type checking imposes somewhat weaker constraints on non-reified function terms than it does for forts (first order reified terms). Specifically, non-reified terms are required only to be consistent with arg-isa and arg-genls constraints using disjointness reasoning constraint is satisfied when no disjointness violation can be found. This is significantly weaker than the constraints imposed on forts: forts must provably satisfy applicable arg-isa and arg-genl constraints using isa and genls reasoning."
The predicates that are currently (August 2001) supported in code are:
- #$arg?Isa, #$arg?Genl, #$arg?GenlAttribute
- #$argAndRestIsa, #$argAndRestGenl, #$argsIsa, #$argsGenl
These are discussed in more detail in their comments in the KB.
Predicates currently (August 2001) unenforced by code include:
- #$argNotIsa, #$argNotGenl
- #$interArgNotIsa*, #$interArgNotGenl*
- #$interArg?ResultIsa, #$interArg?ResultGenls
You can use these as statements of intent, but they will not be enforced by the wff-checker.
Another topic relevant to arg-type constraints is defns. The predicates #$defnIff, #$defnNecessary, and #$defnSufficient, collectively known as "defns", can be used to determine collection membership via a piece of code rather than explicit assertions in the KB. A good example of a collection defined this way is #$PositiveInteger. For example, if you try to assert
(fanOutArg covers-Dustlike "Brazil")
in the BaseKB, the wff-checker will respond:
Term "Brazil" violates defns of arg-isa #$PositiveInteger applicable to argument 2 of relation #$fanOutArg in mt #$BaseKB.
because it can determine at runtime that the string "Brazil" is not a positive integer. This gives us the power to algorithmically define certain collections instead of reifying their extent, which would be especially problematic in the case of collections of infinite cardinality.
For more detailed information on arg-type constraints, see the comments on #$NoteAboutELVersusHLVersusCycL, #$NoteOnArgumentTypingAndPropertiesOfRelations, and #$NoteAboutArgIsaCycIndexedTerm.
For information on special arg-type constraints that affect the behavior of the canonicalizer, see #$SharedNoteAboutCanonicalizerDirectiveArgConstraints.
The CycL sentence must not violate monotonically true SBHL knowledge in the KB. (SBHL assertions are assertions using predicates like #$isa, #$genls, and $#genlAttributes.) For example, if you attempt to assert the CycL sentence
(genls Collection Individual)
in the BaseKB, the wff-checker will give you an error something like:
Formula (genls Collection Individual) was not well formed because: Collection is known not to be a spec of Individual in mt BaseKB.
If you try to assert this assertion, your agenda would halt and you would get the following message:
Last Agenda operation: (FI-ASSERT '(genls Collection Individual) 'BaseKB ':DEFAULT) Agenda halted due to: sbhl conflict: (genls Collection Individual) :TRUE BaseKB because: (genls Collection SetOrCollection) :TRUE (disjointWith SetOrCollection Individual) :TRUE
The CycL sentence must be assertible. For example, making assertions which directly reference a Skolem function (e.g. #$SKF-8675309) is prohibited. See #$notAssertible and #$notAssertibleCollection for more details.
There are various other idiosyncrasies of certain predicates which are enforced by the wff-checker. A detailed explanation of the intricacies of the wff-checker is beyond the scope of this document.
Precanonicalizations can cause terms that you mentioned in your EL input to completely disappear at the HL. A consequence of this is that if you are looking for an assertion via its index on that term, you will not find the assertion there because the term in question was precanonicalized away. If you are cognizant of precanonicalizations, you stand a chance of not being flummoxed by such missing terms.
6.2.3.c. Expands EL relations
#$ELRelations are relations that exist only at the EL and not at the HL. A good example of this is #$lessThan. If you try to assert, for example,
(lessThan 1 2)
then that will be canonicalized as
(greaterThan 2 1)
Instead of having two HL predicates #$lessThan and #$greaterThan, thus doubling the number of defining rules or HL modules needed to support them, we can avoid this entirely by automatically canonicalizing sentences about #$lessThan into sentences about #$greaterThan. Note that this precanonicalization is only performed when the predicate is actually used as a predicate, i.e. in the arg0 position. If the predicate is used as a term, e.g.
(isa lessThan BinaryPredicate)
then no such transformation is performed.
6.2.3.d. Evaluates EL-evaluatable relations (see #$evaluateAtEL)
Evaluatable relations (q.v.) can be marked with the meta-predicate #$evaluateAtEL, which means that the canonicalizer will evaluate them if possible. Suppose that
were asserted in the BaseKB. Then if you asserted
(isa (PlusFn 1 2) PositiveInteger)
in the BaseKB, this would canonicalize as
(isa 3 PositiveInteger)
#$evaluateAtEL makes strong semantic assumptions about the relation to which it is applied. For example, it assumes that the result of the evaluation is unchanging and independent of other knowledge in the KB. Furthermore, it assumes that you will never want to refer to the functional term (PlusFn 1 2). Use it with ontological caution.
6.2.3.e. Converts #$exceptFor and #$exceptWhen into #$abnormal
#$exceptFor and #$exceptWhen are like #$ELRelations that have special support in code. For example
(exceptWhen <sentence-1> <sentence-2>)
denotes that <sentence-2> is default true, but is not warranted when <sentence-1> holds; this sentence is transformed by the precanonicalizer into
(implies <sentence-1> (abnormal <variables> <kb-assertion>))
where <variables> are the EL variables appearing in <sentence-2>, and <kb-assertion> is the HL construct, already asserted in the Cyc knowledge base, derived from canonicalizing <sentence-2>.
See #$EmptinessOfAbnormalityNote and the section on exceptions in inference for more information.
Clausification is why something you asserted might appear in a logical form which is quite different from what you asserted. For example, if you asserted
(implies (hasAttributes ?PERSON Single) (not (#$hasAttributes ?PERSON Married)))
it would canonicalize as
(or (not (hasAttributes ?PERSON Single)) (not (hasAttributes ?PERSON Married)))
which would uncanonicalize (i.e. display in the Cyc Browser) as
(not (and (hasAttributes ?PERSON Single) (hasAttributes ?PERSON Married)))
The following sections will explain what the Canonicalizer does with logical operators and quantifiers: #$and, #$or, #$not, #$implies, #$thereExists, and #$forAll. This explanation is not particularly Cyc-specific. Any decent book on first-order logic (e.g. Logical Foundations of Artificial Intelligence by Genesereth and Nilsson) , or even a web search on "conjunctive normal form" or "skolemization" should have a reasonable explanation of the standard conversion of arbitrary FOL sentences into conjunctive normal form (CNF). The Canonicalizer basically follows the standard approach for clausification.
The canonicalizer eliminates all implications by converting them to an equivalent logical form, by repeatedly performing the following transformation:
(implies <sentence-1> <sentence-2>) -> (or (not <sentence-1>) <sentence-2>)
Pushes negations inwards
Using DeMorgan's Law, double negation elimination, and other logical transformations, the Canonicalizer pushes all negations inwards until all negations apply directly to atomic formulas.
(not (#$and <sentence-1> ...)) -> (or (not <sentence-1>) ...) (not (or <sentence-1> ...)) -> (and (not <sentence-1>) ...) (not (forAll <var> <sentence>)) -> (thereExists <var> (not <sentence>)) (not (thereExists <var> <sentence>)) -> (forAll <var> (not <sentence>))
Standardizes variables It is possible in CycL to have two different variables with the same name. This is usually an OE mistake, but the Canonicalizer will catch this mistake and rename such variables. For example, if you tried to assert
(forAll ?X (elementOf ?X (TheSetOf ?X (isa ?X Thing))))
This would canonicalize into something like
(forAll ?X (elementOf ?X (TheSetOf ?Y (isa ?Y Thing))))
because #$TheSetOf is a #$ScopingRelation with #$scopingArg 1, which means that the variable in argument position 1 of the relation #$TheSetOf is a scoped variable. Scoped variables cannot be referenced outside of their scoping expressions, in this case the #$TheSetOf formula. Thus the ?X in #$forAll ?X and the ?X in (#$isa ?X #$Thing) are two different variables.
Converts existentials to Skolem functions
This section will not provide a detailed discussion of Skolemization since this can be found in textbooks on formal logic, but it give a brief overview, and then explain how Skolemization works in Cyc.
Basically, Skolemization is a method of eliminating existential quantification by introducing new functions. For example, the CycL sentence
(implies (isa ?PER Person) (thereExists ?MOM (mother ?PER ?MOM)))
would skolemize into something like
(implies (isa ?PER Person) (mother ?PER (SKF-8675309 ?PER)))
In this particular case, #$SKF-8675309 is equivalent to #$MotherFn, the function that, for any given person, denotes his or her mother. Skolem functions are just like any other denotational functions in Cyc, except that they are automatically generated by the Canonicalizer any time a new assertion is made using an existential quantifier (e.g. #$thereExists). They are named #$SKF-n, for some arbitrarily chosen n. The arity, arg constraints, and result types are automatically determined by the Canonicalizer.
Skolem constants are implemented by zero-arity skolem functions, the result of which is effectively a Skolem constant. For example, #$SKF-8675309 might be a zero-arity Skolem function, but (#$SKF-8675309), the result of applying that Skolem function to zero arguments, would effectively be a Skolem constant.
Skolem functions have one or more defining assertions; they are only meaningful in the context of those assertions. Therefore, it is impermissible to make non-meta assertions about existing Skolem functions. You can make meta-assertions about them (e.g. #$comment), but no more. If you want to change the definition of a Skolem function, you must change its defining assertions, which is tantamount to killing the Skolem function and creating a new one.
Removes leading universals
Since all existentially quantified variables have been removed by Skolemization, all remaining free variables must be universally quantified. The redundant #$forAll wrappers are stripped from the CycL sentence before storing it in the KB. The Uncanonicalizer could recompute the #$forAll wrappers if it chose to, but there is a convention in CycL that free variables are implicitly universally quantified at the top level.
Pushes disjunctions inwards (or conjunctions, if a query)
The Canonicalizer uses the transformation
(or P (and Q R)) -> (and (or P Q) (or P R))
to transform the sentence into a conjunction of disjunctions of literals, if the sentence is being asserted into the KB. If the sentence is being asked as a query, then the dual transformation is applied, and the sentence is converted to disjunctive normal form (DNF) instead of CNF.
6.2.3.g. Clause Canonicalization
Polycanonicalizes if necessary
The entire Cyc KB can be viewed as a huge conjunction of CycL assertions. (Of course, they would need to be decontextualized via #$ist in order to not lose the distinction between different microtheories.) Hence, if you assert an EL sentence of the form (#$and P Q), this will result in two assertions in the KB. Sometimes this result is not obvious, for instance a rule like
(implies (or P Q) R)
would be clausified into
(and (or (not P) R) (or (not Q) R))
which would also result in two assertions being added to the KB. When a single EL sentence results in more than one assertion being added to the KB, this is called polycanonicalization. The Uncanonicalizer is not aware of polycanonicalization (except for Skolem functions) and so it will not reconstruct the original EL sentence as you entered it, but will instead reconstruct two (or more) EL sentences, each of which will be a logical piece of the original input sentence.
Sorts arguments of commutative relations, including #$and and #$or
The Canonicalizer will sort the arguments of commutative relations into a canonical order. #$and and #$or are always commutative. A user-defined relation is commutative iff it is an instance of #$CommutativeRelation in a microtheory relevant to the given canonicalization.
The effect of this is that the assertions in the KB may have a different argument order than you originally wrote, and the order of the literals may be different. This has absolutely no bearing on inference efficiency; the inference engine will choose which literal to evaluate based on its own heuristics; not the order chosen by the Canonicalizer.
Implementation note: canonicalizing the order of arguments to commutative relations obviates rules that implement commutativity, such as:
(implies (different ?X ?Y) (different ?Y ?X))
This requires that every component of the inference engine (e.g. the KB indexing mechanisms and HL modules) assume responsibility to correctly handle commutativity.
Converts some EL objects to HL objects, e.g. reifiable nauts and EL variables
In Cyc parlance, reifiable nauts are canonicalized into narts. A nart is a non-atomic reified term, and a naut is a non-atomic unreified term. Hence, non-atomic unreified terms whose functor is an instance of #$ReifiableFunction will be replaced by reified versions of those terms. The motivation for this is to support efficient indexing of non-atomic terms. In the standard Cyc Browser, you can tell the difference between a naut and a nart by the clickable left parenthesis; narts have clickable left parens, and nauts don't. They look very similar in the Cyc Browser, but at the HL they are totally different things.
EL variables are replaced by HL variables in order to support an effient unification algorithm in the inference engine.
EL sentences are replaced by HL assertions if indicated by arg-type constraints (e.g. #$CycLAssertion). This results in a meta-assertion. For more information on meta-assertions, see #$NoteAboutAssertionsAsArgConstraints and#$NoteAboutMetaAssertionsAcrossMts. Also see #$MetaAssertionsForPolyCanonicalizingAssertions for an explanation of how polycanonicalization interacts with meta-assertions.
Adds #$termOfUnit and #$evaluate literals when necessary
The canonicalizer may add some additional literals to the HL sentence, to facilitate the operation of the inference engine. #$termOfUnit relates a nart to its corresponding naut. It is an inference-internal predicate and you should never need to mess with it.
The canonicalizer will also add #$evaluate literals to facilitate the inference engine's handling of #$EvaluatableFunctions.
Simplification is performed at various points during canonicalization. This process includes such logical transformations as
(and P) -> P (or P) -> P (and False ...) -> False (or True ...) -> True (and True ...) -> (and ...) (or False ...) -> (or ...) (not (not P)) -> P
and many others. Simplification is performed so as to map as many logically equivalent EL expressions as possible to the same HL expression.
6.2.3.i. Summary of how the Canonicalizer affects the OE process
Often, the operation of the Canonicalizer is invisible to the ontologists, except for the occasional notification of ill-formedness. But sometimes, the Canonicalizer will cause assertions in the KB to look considerably different from the EL form you entered. Being aware of how the Canonicalizer works will help you to recognize these canonicalized assertions as your own, and to understand how and why they were transformed.