Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
305 lines (226 sloc) 15.9 KB

Knowledge Base

Knowledge base files in LoMRF are text files having the suffix .mln (e.g., file_name.mln). The contents of a knowledge base file in LoMRF can be composed of the following:

  1. [Optional] Domain types with their possible values.

  2. Predicate definitions, representing the structure of each predicate.

  3. [Optional] Function definitions, representing the structure of a function.

  4. [Optional] Formulas represented in First-order logic, expressing a template for producing Markov Networks.

  5. [Optional] Definite clauses, a special case of formulas.

Domain Types

Domain types with their possible constant symbols. Each domain type has some unique name and represents a particular collection of constant symbols. All constant symbols in LoMRF are finite and discrete.

Consider, for example, the domain of time ranging from 1 to 12. This domain is defined in LoMRF with the following statement:

time = {1,...,12}

For domains containing integer symbols, like the example domain of time, we can define a sequence of integer numbers using the ... symbol.

Similarly, a domain that represents a collection of person names is defined as:

person = {Achilles, Agamemnon, Menelaus, Helen, Odysseus}

The possible values of domain types can automatically discovered from the specified knowledge base and evidence file(s). Specifically, for each domain type, LoMRF computes the union of unique constant symbols from the knowledge base and the evidence file(s). For example, we may not explicitly specify all person names, but can be automatically discovered from the evidence files and/or the contents of the given knowledge base.

Predicate Definitions

Predicate definitions, express the structure of each predicate. For example, the predicate that is named with the symbol Brothers defines that two persons are brothers. Therefore, the definition of predicate Brothers has two arguments, where each one takes constants from the domain of person:

Brothers(person, person)

Similarly, the predicate Happens associates the domain of action with time is defined as:

Happens(action, time)

Therefore, all predicates have a name (e.g., Brothers, Happens, etc) and a specific number of arguments. Each argument has some domain type (e.g., person, action, time, etc).

Function Definitions

Function definitions, express the structure of a function. All functions in LoMRF are finite with known domain types. Each function definition has a name, zero or more domain types as arguments and a resulting domain type.

For example, the function that is named as monthOf with a single argument that takes constants from the domain of time and returns constants from the domain of month (i.e., the month number):

month monthOf(time)

Similarly a function next that takes a single argument from the domain of time and returns the next one is defined as follows:

time next(time)

Build-in functions and predicates

Build-in functions and predicates are supported internally in the LoMRF and do not require any schema definition in the knowledge base file.

Functions

Function Description
x++ Increase the integer number x by one
x-- Degrease the integer number x by one
x + y Sum x with y
x - y Subtract y from x
x * y Multiply x with y
x / y Divide y from x
x % y The remainder of x divided by y
concat(x, y) Concatenate x with y

Predicates

Predicate Description
a = b Term a is equal with term b
a < b Term a is less than term b
a <= b Term a is less than or equal with b
a > b Term a is greater than term b
a >= b Term a is greater than or equal with b
substr(a, b) Term a is substring of b

First-Order Logic Formulas

Terms:

Terms are intuitively represent objects and can be any of the following:

  • Constants, starting with upper-case letter or numeric symbols, e.g., Achilles, Agamemnon, 1, 2, etc.
  • Variables, starting with lower-case letter symbols, expressing any constant of a specific domain type, e.g., x, y, z, t, id, name, etc.
  • Functions, starting with lower-case letter symbols, may contain zero or more terms as arguments, e.g., motherOf(X), fatherOf(Agamemnon), etc.

Predicates:

Similar to any first-order logic representation, a predicate expresses a relation among its terms that can be either True or False. In LoMRF, predicates start with an upper-case letter. For example, the statement that Agamemnon and Menelaus are brothers, is represented by the predicate Brothers(Agamemnon, Menelaus). Predicates may have constants, variables and function symbols in their arguments. For example, Brothers(Agamemnon, x) is composed of the constant Agamemnon and the variable x.

Please note that when a predicate does not contain any variable is called ground. For example, the predicates Brothers(Agamemnon, Menelaus) and Brothers(Atlas, fatherOf(Patroclus)) are ground, while the predicates Brothers(x, y), Brothers(Agamemnon, y) or Brothers(Atlas, fatherOf(y)) are not.

Formulas:

Formulas are represented in First-order logic, expressing a template for producing Markov Networks.

  • Each formula imposes a constraint over some predicates.
  • Each formula can be associated with some weight value, that is a positive or negative real number. The higher the value of weight, the stronger the constraint represented by the formula. In contrast to classical logic, all worlds (i.e., Herbrand Interpretations) are possible with a certain probability. The main idea behind this is that the probability of a world increases as the number of formulas it violates decreases.

Formulas may contain one or more predicates, connected to each other with logical connectives and quantified symbols. The logical connectives and quantifiers in LoMRF are outlined in the following table:

Symbol Description
^ Logical conjunction (And)
v Logical disjunction (Or)
! Logical negation (Not)
=> Logical implication
<=> Logical equivalence (if and only if)
Forall Universal quantification
Exist Existential quantification

By default all variables implicitly assumed to be universally quantified unless otherwise indicated.

A knowledge base may contain both hard and soft-constrained formulas. Hard-constrained formulas are associated with an infinite weight value and capture the knowledge which is assumed to be certain. Therefore, an acceptable world must at least satisfy the hard constraints. Soft constraints, on the other hand, capture imperfect knowledge in the domain, allowing the existence of worlds in which this knowledge is violated.

  • Hard-constrained formulas, do not have weights (the weight is assumed to be infinite) and capture the knowledge which is assumed to be certain.
  • Soft-constrained formulas are always associated with weights and capture imperfect knowledge, allowing the existence of worlds in which this knowledge is violated.

Definite Clauses

A special case of formulas are the Definite clauses, which can be used to define declarations of rules. The definite clauses, are processed by the LoMRF and automatically translated to equivalent formulas. Their syntax is simpler from the syntax of formulas and are ideal for defining domain-specific/common-sense knowledge.

A definite clause can be either hard-constrained or soft-constrained and is composed of two parts:

  1. The head part is a single positive literal.
  2. The body part is single literal or a conjunction of two or more literals. The literals can be either positive or negative (i.e., negated predicate). In contrast to first-order formulas, disjunctions, quantifiers, implications and equivalences are not allowed.

Example of soft-constrained definite clause in LoMRF, with a single positive literal in the body:

3.2 HeadPredicate(x, y) :- BodyPredicate(x, y)

Example of hard-constrained definite clause in LoMRF, in which the body is composed by one positive (i.e., BodyPredicate1(x)) and one negative (i.e., !BodyPredicate2(y)) literal:

HeadPredicate(x, y) :- BodyPredicate1(x) ^ !BodyPredicate2(y).

It is possible to have more than one definite clauses for the same head predicate, and thus define alternative definitions (i.e., disjunctions) for the same head statement.

2 Head(f(x), t) :- FooPredicate(x, t) ^ !BarPredicate(z(t))
-1.68 Head(f(x), t) :- AnotherPredicate(x, t)

LoMRF uses definite clauses to create formulas that express if-and-only-if conditions to head predicates. In that way LoMRF implicitly introduces closed-world assumption to head predicates. To better illustrate this, assume that we have the following definite clauses:

Head(f(x), t) :- FooPredicate(x, t) ^ !BarPredicate(z(t)).
-1.68 Head(f(x), t) :- AnotherPredicate(x, t)

LoMRF will translate the given clauses into the following formulas:

//
// 1. Convert definite clauses to: (weight) body => head
//
FooPredicate(x, t) ^ !BarPredicate(z(t)) => Head(f(x), t).
-1.68 AnotherPredicate(x, t) => Head(f(x), t)

//
// 2. Introduce closed-world assumption via predicate completion
//    for the head predicate Head(f(x), t), as hard-constraint
//
Head(f(x), t) => (FooPredicate(x, t) ^ !BarPredicate(z(t))) v AnotherPredicate(x, t).

As we can see from the resulting translation, LoMRF produces 2 + 1 formulas for the definitions of Head(f(x), t). The first two formulas are straightforward translations of the definite clauses to (weight) body => head formulas. Specifically, if the definite clause is hard-constrained, the corresponding translation will also remain hard-constrained. Otherwise, the resulting translation will keep the same weight value. The last resulting hard-constrained formula introduces the opposite direction, that is head => disjunction of all body parts. That formula is always hard-constraint and states that in order to have the head satisfied, at least one of the body parts must also be satisfied. Any other possibility doesn't affect the state of the head predicate. With that translation LoMRF implicitly introduces closed-world assumption to head predicates, using a technique that is called predicate completion (see McCarthy, 1980; Lifschitz, 1994).

LoMRF tries to simplify the resulting knowledge base, by specializing as much as possible the predicate completion for each distinct head predicate. Specifically, the original predicate completion will result to a more general form, producing a single formula with an equivalence:

Head(a, t) <=> (a = f(x) ^ FooPredicate(x, t) ^ !BarPredicate(z(t))) v (a = f(x) ^ AnotherPredicate(x, t))

The problem with that translation is that we cannot keep the weight values from the original definite clauses and that increases the number variables (i.e., the additional variable a).

In contrast to the original predicate completion approach, LoMRF expresses the predicate completion in a decomposed form, that is one formula per definite clause and one single formula with the disjunctions of body parts. The implementation is based on the transformations presented in Skarlatidis et. al. (2011, 2015).

Technically, the definite clauses are translated into logically stronger first-order formulas. In particular, LoMRF performs predicate completion for each unique head predicate in the knowledge base. In cases where some definitions of head predicates are missing, the corresponding missing definition are implicitly considered as False. For further details see Skarlatidis et. al. (2011, 2015), as well as the Advanced Cases section.

Advanced Cases

Variables that appear only in the body part

In situations where there are variables in the body part that do not appear in the head, then existential quantification is introduced. For example, in the following clause the variable z appears only in the body part:

1.86 Head(x,y) :- Foo(x, z) ^ Bar(z, y)

In such case, the translated formulas will be the following:

1.86 Exist z Foo(x, z) ^ Bar(z, y) => Head(x, y)

Head(x,y) => Exist z Foo(x, z) ^ Bar(z, y).

Please note that during grounding existentially quantified formulas are replaced by the disjunction of their groundings (see Domingos and Lowd, 2009). This may lead to a large number of disjunctions and a combinatorial explosion of the number of clauses, producing unmanageably large Markov networks. This type of clauses should be avoided, if it is possible.

Partial definitions

In some cases we may not have definitions for each unique head predicate. Consider, for example, the following knowledge base:

time = {1,...,12}
values = {Foo, Bar}

Head(values, time)
P(time)
Q(time)

Head(Foo, t) :- P(t) ^ Q(t).

In the given knowledge base, the domain type values is composed of two constants, that is Foo and Bar. Furthermore, the knowledge base contains a single definite clause for the head predicate Head(Foo, t). LoMRF will compute the decomposed predicate completion for the predicates having Head/2 as signature. However, the definition for the head predicate Head(Bar, t) is missing and thus LoMRF will implicitly define it as false - i.e., Head(Bar, t) <=> False. As a result, the final form of the knowledge base is given below:

BodyPredicate1(t) ^ BodyPredicate2(t) => Head(Foo, t).

Head(Foo, t) => BodyPredicate1(t) ^ BodyPredicate2(t).

!Head(Bar, t).

As we can see from the resulting knowledge base Head(Bar, t) <=> False is equivalently expressed as single negated unit clause !Head(Bar, t)..

Limitations of the decomposed form

In some cases we may have a knowledge base with definite clauses of a particular head predicate, but the heads does not contain the same level of variables. For example, we may have the following two definite clauses:

Head(Foo, t) :- P(t) ^ Q(t).
Head(x, t) :- R(x) ^ Q(t).

The first term of the head predicate of the first clause is a constant (Foo), while the in the second one the corresponding term is a variable (x). In that case the second clause contains a higher level of variables in the head predicate (Head(x, t)).

In such situations, LoMRF computes the predicate completion with respect to the head predicate with the higher level of variables. Therefore, the first clause is handled as:

Head(x, t) :- x = Foo ^ P(t) ^ Q(t).

and the resulting knowledge base is the following:

x = Foo ^ P(t) ^ Q(t) => Head(x, t).
R(x) ^ Q(t) => Head(x, t).

Head(x, t) => (x = Foo ^ P(t) ^ Q(t)) v (R(x) ^ Q(t)).
You can’t perform that action at this time.