Rule Types and Syntax
PyClause uses conventions for type names and rule formatting from this publication .
Rule Syntax
To load rules with the c_clause.Loader they must be represented in a structured syntax. An illustrative example of a B-rule string is as follows:
has_part(X,Y) <= member_meronym(X,A), hypernym(A,B), derivationally_related_form(Y,B)
The head and the body of a rule are distinguished by the “ <= “ delimiter, with spaces required on both sides.
Head
The head contains a single atom in the format rel(X,Y), where rel is a relation string in the knowledge graph. The head atom may have X on the left and Y on the right. No other variables appear in the head. Depending on the rule type, X or Y —but not both— might be replaced with an entity string (a constant) or the Y is set to X.
Body
Atoms within the body are divided using a comma “, “ with a required space after it. The two terms within an atom are separated by a “,” with no spaces included. Body atoms can also include entitiy strings (constants). Possible variables and entities in an atom can be inferred by the following heuristic:
Given the ordering,
variables = ["A", "B", "C", "D", "E", "F", "G", "H", "I", ...],
the following holds for the body atoms:
The atom at position i=0 contains variables[0] and X or Y, depending on the rule type.
Intermediate atoms, where 0<i<length-1 (length being the number of body atoms), include variables[i-1] and variables[i].
The final atom, at position i=length-1, involves variables[i-1] and an entity string or Y.
The ordering within a particular body atom is free and changing the positions will result in a distinct rule. Moreover, the head relation is allowed to be included in the body atoms:
rel1(X,Y) <= rel2(X,A), rel1(A,Y)
rel1(X,Y) <= rel2(A,X), rel1(Y,A)
rel1(X,ent_1) <= rel2(X,ent_2)
rel1(X,ent_1) <= rel2(ent_2,X)
rel1(X,Y) <= rel2(Y,X)
rel1(X,Y) <= rel2(X,Y)
rel1(X,Y) <= rel1(Y,X)
Note
Entity string format. Entities (constants) within rules must exist in the knowledge graph of the loaded data. They should consist of at least two characters.
Supported Rule Types
We show below the supported rule types. The examples follow the correct syntax, e.g., they could be parsed by the c_clause.Loader.
Let
rel1, rel2, rel3, rel4be relation strings.Let
cc, ddbe entity strings.Let
X A B C D E F G H .. Ybe variable strings.
B-Rules
The head together with the body form a cycle when directions within the atoms are ignored. Any length supported. Examples:
rel1(X,Y) <= rel1(Y,X)
rel1(X,Y) <= rel1(X,A), rel1(A,Y)
rel1(X,Y) <= rel2(X,A), rel3(A,Y)
rel1(X,Y) <= rel2(A,X), rel3(A,Y)
rel1(X,Y) <= rel2(X,A), rel3(A,B), rel4(B,Y)
U_c Rules
One entity constant in the head and one in the body (either for X or Y). Note that Y must be contained in the first atom if X is substituted. It is allowed that cc and dd are the same entity. Any length supported. Examples:
rel1(X,cc) <= rel2(X,dd)
rel1(cc,Y) <= rel2(Y,cc)
rel1(cc,Y) <= rel2(cc,Y)
rel1(X,cc) <= rel2(A,X), rel3(A,dd)
rel1(cc,Y) <= rel2(Y,A), rel3(dd,A)
rel1(cc,Y) <= rel1(A,Y), rel2(A,dd)
rel1(cc,Y) <= rel2(A,Y), rel3(A,B), rel4(B,dd)
U_d Rules
One entity constant in the head no entity constant in the body. Note that the variables in the body follow the generic heuristic above for intermediate atoms without having a Y or entitiy in the last atom. Any length supported. Examples:
rel1(X,cc) <= rel2(X,A)
rel1(cc,Y) <= rel2(Y,A)
rel1(cc,Y) <= rel2(A,Y)
rel1(X,cc) <= rel2(A,X), rel3(A,B)
rel1(cc,Y) <= rel2(Y,A), rel3(B,A)
rel1(cc,Y) <= rel2(A,Y), rel3(A,B)
rel1(c,Y) <= rel2(A,Y), rel3(A,B), rel4(B,C)
Zero (Z)-Rules
This rule type is directed and has no length. The first rule only makes prediction in tail direction given a head=X and vice versa for the second rule. It can not be used for triple scoring.
Examples:
rel1(X,cc) <=
rel1(cc,Y) <=
U_xxc Rules
This rule type only has one body atom.
Examples:
rel1(X,X) <= rel2(X,dd)
rel1(X,X) <= rel2(dd,X)
U_xxd Rules
This rule type only has one body atom.
Examples:
rel1(X,X) <= rel2(X,A)
rel1(X,X) <= rel2(A,X)