Grammar Specs

Revision as of 12:31, 10 February 2010

UNL-NL grammars are sets of rules for translating UNL expressions into natural language (NL) sentences and NL sentences into UNL expressions. They are normally unidirectional, i.e., the enconversion grammar (NL-to-UNL) is different from the deconversion grammar (UNL-to-NL), even though they share the same basic syntax. In order to standardize the language resources in the UNL framework, the UNDL Foundation recommends the adoption of the following specifications for both UNL-to-NL and NL-to-UNL grammars. This formalism, however, is not supported by the UNL Centre's tools, and it is only required by those interested in using UNDL Foundation's tools.

Types of rules

In the UNL Grammar there are two basic types of rules:

Transformation rules

Used to generate natural language sentences out of UNL graphs and vice-versa.

Disambiguation rules

Used to improve the performance of transformation rules by constraining their applicability.

The Transformation Rules follow the very general formalism

α:=β;

where the left side α is a condition statement, and the right side β is an action to be performed over α.

The Disambiguation Rules, which were directly inspired by the UNL Centre's former co-occurrence dictionary and knowledge base, follows a slightly different formalism:

α=P;

where the left side α is a statement and the right side P is an integer from 0 to 255 that indicates the probability of occurrence of α.

We present both types of rules and their role in the UNL System. We introduce, first, the basic symbols that are used both by transformation and disambiguation rules; next, we present the transformation rules and their several subtypes; and finally we present the disambiguation rules.

Node

Transformation and disambiguation rules apply over nodes, either in a natural language sentence structure or in the UNL hypergraph structure. The node is the basic unit in the grammar and may be referred by:

• a string, to be expressed between "quotes";
• a dictionary entry, to be expressed between [square brackets];
• a dictionary sub-entry
• a feature (attribute, value, or attribute-value pair) or set of features of a dictionary entry;
• a pointer (index) to an existing node; and
• a relation between nodes (i.e., a hypernode).

Nodes can also be a combination of such values. Examples of nodes are

• "ing" (constant)
• [book] (entry)
• NOU (value)
• VA(01>02) (value)
• POS=NOU (attribute-value pair)
• %01 (pointer)
• %01%01%01 (pointer)
• %source (pointer)
• %01#01 (pointer to a dictionary sub-entry)
• agt(VER;NOU) (relation)
• PP(PB(PC(PRE;NP);PP);DP) (relation)
• [book],POS=NOU,%01 (combination of several possibilities)

Both transformation and disambiguation rules use the same set of basic symbols to address nodes:

Transformation rules

Natural language sentences and UNL graphs are supposed to convey the same amount of information in different structures: whereas the former arranges data as an ordered list of words, the latter organizes it as a hypergraph. In that sense, translating from natural language into UNL and from UNL into natural language is ultimately a matter of transforming lists into networks and vice-versa.

The UNDLF generation and analysis tools assume that such transformation should be carried out progressively, i.e., through a transitional data structure: the tree, which could be used as an interface between lists and networks. Accordingly, the UNL Grammar states seven different types of rules (LL, TT, NN, LT, TL, TN, NT), as indicated below:

• ANALYSIS (NL-UNL)
  • LL - List Processing (list-to-list)
  • LT - Surface-Structure Formation (list-to-tree)
  • TT - Syntactic Processing (tree-to-tree)
  • TN - Deep-Structure Formation (tree-to-network)
  • NN - Semantic Processing (network-to-network)

• GENERATION (UNL-NL)
  • NN - Semantic Processing (network-to-network)
  • NT - Deep-Structure Formation (network-to-tree)
  • TT - Syntactic Processing (tree-to-tree)
  • TL - Surface-Structure Formation (tree-to-list)
  • LL - List Processing (list-to-list)

The NL original sentence is supposed to be preprocessed, by the LL rules, in order to become an ordered list. Next, the resulting list structure is parsed with the LT rules, so as to unveil its surface syntactic structure, which is already a tree. The tree structure is further processed by the TT rules in order to expose its inner organization, the deep syntactic structure, which is supposed to be more suitable to the semantic interpretation. Then, this deep syntactic structure is projected into a semantic network by the TN rules. The resultant semantic network is then post-edited by the NN rules in order to comply with UNL standards and generate the UNL Graph.

The reverse process is carried out during natural language generation. The UNL graph is preprocessed by the NN rules in order to become a more easily tractable semantic network. The resulting network structure is converted, by the NT rules, into a syntactic structure, which is still distant from the surface structure, as it is directly derived from the semantic arrangement. This deep syntactic structure is subsequently transformed into a surface syntactic structure by the TT rules. The surface syntactic structure undergoes many other changes according to the TL rules, which generate a NL-like list structure. This list structure is finally realized as a natural language sentence by the LL rules.

As sentences are complex structures that may contain nested or embedded phrases, both the analysis and the generation processes may be interleaved rather than pipelined. This means that the natural flow described above is only "normal" and not "necessary". During natural language generation, a LL rule may apply prior to a TT rule, or a NN rule may be applied after a TL rule. Rules are recursive and must be applied in the order defined in the grammar as long as their conditions are true, regardless of the state.

Network-to-Network Rules

The network-to-network rules (NN) are used for processing networks, both in analysis and in generation. During analysis, these rules are used for post-editing the semantic network structure derived from the syntactic module in order to generate the UNL graph; in generation, they are used for pre-editing the UNL graph, transforming it into a semantic network that would be more appropriate for sentence generation.

The NN rules follow the general syntax below:

<NN RULE>::= <SEM>(<SEM>)*”:=”[{”-“,”+“}]<SEM>([{”-“,”+“}]<SEM>)*”;”
<SEM>::={agt,aoj,...,via}”(“<NODE>”;”<NODE>”)”
<NODE>:= see #Node

There are 4 different subtypes of NN rules:

Tree-to-Tree Rules

The tree-to-tree rules (TT) are used for processing trees, both in analysis and in generation. During analysis, these rules are used for revealing the deep structure out of the surface structure; in generation, they are used for transforming the deep into the surface syntactic structure. The TT rules follow the general syntax below:

<TT RULE>::= <SYN>(<SYN>)*”:=”[{”-“,”+“}]<SYN>([{”-“,”+“}]<SYN>)*”;”
<SYN>::={syntactic relation}”(“<NODE>”(“;”<NODE>”)*)”
<NODE>:= see #Node

Syntactic relations are n-ary: they can have as many arguments (nodes) as necessary.

There are 5 different subtypes of TT rules:

List-to-List Rules

The list-to-list (LL) rules are used for processing lists, both in analysis and in generation. In analysis, these rules are used for pre-editing the natural language sentence and preparing the input to the syntactic module; in generation, they are used for post-editing the output of the syntactic module and generating the natural language sentence.

The LL rules follow the syntax below:

<LL RULE>::= “(”<NODE>”)”(”(”<NODE>”)”)*”:=”[{”+”,”-“}]“(”<NODE>”)”(“(”<NODE>”)”)*”;”
<NODE>:= see #Node

There are 5 different subtypes of LL rules:

List-to-Tree Rules

The list-to-tree (LT) rules are used to parse the list structure into a (surface) tree structure. It is used only in analysis, and all LT rules follow the syntax below:

<LT RULE>::= “(”<NODE>”)”(“(”<NODE>”)”):=<SYN>(“<SYN>”)*”;”
<SYN>::={syntactic relation}”(“<NODE>”;”<NODE>”)”
<NODE>:= see #Node

There is a single type of LT rule:

Tree-to-List Rules

The tree-to-list (TL) rules are used to linearize the (surface) tree structure into a list structure. It is used only in generation, and all TL rules follow the syntax below:

<TL RULE>::= <SYN>(“<SYN>”)*:=“(”<NODE>”)”(“(”<NODE>”>”)”;”
<SYN>::={syntactic relation}”(“<NODE>”;”<NODE>”)”
<NODE>:= see #Node

There is a single type of TL rule:

Tree-to-Network Rules

The tree-to-network (TN) rules derive a semantic network out of a syntactic tree. It is used only in analysis, and all TN rules follow the syntax below:

<TN RULE>::= <SYN>”(“<SYN>)*”:=”<SEM>(<SEM>)*”;”
<SYN>::={syntactic relation}”(“<NODE>”;”<NODE>”)”
<SEM>::={agt,aoj,...,via}”(“<NODE>”;”<NODE>”)”
<NODE>:= see #Node

There is a single type of TN rule:

Network-to-Tree Rules

The network-to-tree (NT) rules reorganizes the network structure as a deep tree structure. It is used only in generation, and all the NT rules follow the syntax below:

<NT RULE>::= <SEM>”(“<SEM>)*”:=”<SYN>(<SYN>)*”;”
<SEM>::={agt,aoj,...,via}”(“<NODE>”;”<NODE>”)”
<SYN>::={syntactic relation}”(“<NODE>”;”<NODE>”)”
<NODE>:= see #Node

There is a single type of TN rule:

General Properties of Transformation Rules

PRIORITY

Rules should be applied serially, according to the order defined in the grammar. The first rule will be the first to be applied, the second will the second, and so on.

RECURSIVENESS

Rules should be applied recursively as long as their conditions are true.

COMPREHENSIVENESS

Grammars should be applied comprehensively as long as there is at least one applicable rule.

ACTION

The rules may add or delete values to the source and the target nodes, but only in the right side items:

agt(a;b):=agt(+c;);

agt(a;b):=agt(;-b);

CONSERVATION

Rules affect only the information clearly specified. No relation, node or feature is deleted unless explicitly informed.

For instance, in the examples below, the source node of the “agt” relation preserves, in all cases, the value “a”. The only change concerns the feature “c”, which is added to the source node of the “agt” in the first two cases; and the feature “b”, which is deleted from the target node in the third case.

agt(a;b):=agt(c;);

agt(a;b):=agt(+c;);

agt(a;b):=agt(;-b);

In any case, the ADD and DELETE rules (i.e., when the right side starts with “+”or “-“) preserve the items in the left side, except for the explicitly deleted ones:

INPUT: agt(a;b) obj(a;c) tim(a;d)

RULE: agt(a;b) ^mod(a;e):=+mod(a;e);

OUTPUT: agt(a;b) obj(a;c), tim(a;d), mod(a;e)

or

INPUT: agt(a;b) obj(a;c) tim(a;d)

RULE: agt(a;b):=-agt(a;b);

OUTPUT: obj(a;c) tim(a;d)

The REPLACE, MERGE and DIVIDE rules affect only their designated scopes. In that sense, NN may only replace, merge or divide semantic relations; TT may only replace, merge or divide syntactic relations; and LL may only replace, merge or divide list nodes. All other information is preserved, unless explicitly informed.

INPUT: agt(a;b) cob(a;c)

RULE: cob(;):=obj(;);

OUTPUT: agt(a;b) obj(a;c)

or

INPUT: agt(a;b) cob(a;c)

RULE: cob(a;):=obj(-a,+d;);

OUTPUT: agt(a;b) obj(d;c)

CONJUNCTION

Both the left and the right side of the rule may have as many items as necessary, as exemplified below:

SEM(A;B)SEM(C;D)SEM(E;F):=SEM(G;H)SEM(I;J)SEM(K;L);

DISJUNCTION

The left side of the rules may bring disjuncts, but not the right side.

{SEM(A;B),SEM(C;D)},^SEM(E;F):=+SEM(E;F);

SEM(A;B),{SEM(C;D),SEM(E;F)}:=-SEM(A;B);

agt(VER,{V01,V02};NOU,^SNG}:=;

EXTENDIBILITY

The left side of the rules may bring [Perl Compatible Regular Expressions] between "/", as indicated below:

/(agt|obj|aoj)/(A,%a;B,%b):=VS(%a;%b);

The rule above applies in case of agt(A;B), obj(A;B) and aoj(A;B)

/[a-z]{2,3}/(A,%a;B,%b):=VS(%a;%b);

The rule above applies in case of any sequence of two or three alphabetic characters in the position of relation of A and B

agt(/(VER|NOU)/,%a;%b):=VS(%a;%b);

The rule above applies in case of VER and NOU as features of the first node of the relation "agt"

agt(POS=/(VER|NOU)/,%a;%b):=VS(%a;%b);

The rule above applies in case of VER and NOU as values of the attribute POS of the first node of the relation "agt"

CONCISION

Rules should be as small as possible. In that sense, the source and the target nodes may be simple place-holders or indexes:

cob(;):=obj(;);

tim(%01;[[in]]),obj([[in]];%02):=tim(%01;%02);

tim(VER,%01;[[in]]),obj([[in]];NOU,%02):=tim(%01;%02);

tim(VER;in),obj(in;NOU):=tim(;%04);

In the DELETE rules, the right side may be omitted in case of deletion of the entire left side:

obj(PRE;):=;

READABILITY

There can be blank spaces between variables and symbols. Comments can be added after the “;”.

obj ( ; ) := ; this rule deletes every “obj” relation.

SCOPE GENERATION

Inside a relation, nodes can be relations (i.e., hypernodes).

SEM(A;SEM(C;D)):=SEM(A;C),SEM(C;D));

SEM(A;C),SEM(C;D):=SEM(A;SEM(C;D));

COMMUTATIVITY

Inside the same side of the rule, the order of the factors does not affect the end result, except for list-processing rules (LL, LT and TL).

SEM(A;B):=SEM(C;D)SEM(E;F); = SEM(A;B):= SEM(E;F)SEM(C;D);

SYN(A;B):=SYN(C;D)SYN(E;F); = SYN(A;B):= SYN(E;F)SYN(C;D);

But:

(A):=(B)(C); IS DIFFERENT FROM (A):=(C)(B);

SYN(A;B):=(C)(D); IS DIFFERENT FROM SYN(A;B):=(D)(C);

(C)(D):=SYN(A;B); IS DIFFERENT FROM (D)(C):=SYN(A;B);

Additionally, the order of the features inside a relation does not affect the end result, but the order of the nodes is non-commutative.

SEM( VER,TRA ; NOU,MCL ) = SEM( TRA,VER ; MCL,NOU )

But:

SEM( VER,TRA ; NOU,MCL) IS DIFFERENT FROM SEM( NOU,MCL ; VER,TRA )

DICTIONARY ATTRIBUTES

Dictionary attributes can be used as variables (see indexes).

SYN( A,^num ; B,num ):=SYN( A,num=%02; %02 );

DICTIONARY RULES

Dictionary rules are triggered by "!"<ATTRIBUTE>:

Dictionary

[foot] "foot" (NOU, NUM(PLR:=”oo”:”ee”)) <eng,0,0>;

[city] "city" (NOU, NUM(PLR:=”y”>”ies”)) <eng,0,0>;

Grammar

(@pl, NUM):=(!NUM,-@pl);

Output

foot>feet

city>cities

NLW SPLITTING

Sub-NLWs in complex entries are referred by # (see indexes).

Dictionary

[[bring] [back]] "bring back" (VER,MTW,VA(01>02), #01(HEAD,VER), #02(ADJT,PP)) <eng,0,0>;

Grammar

VC(VER,MTW,VA(01>02),%head;NOU,%obj):=VB(VC(%head#01;%obj);%head#02);

Indexes

Indexes (%) are used for indexing nodes, attributes and values inside and between the left and the right side of rules.

X(%a;)Y(%a;)

the first node of X is also the first node of Y

If omitted, indexes are assigned by default, according to the position

X(A;B)Y(C;D)Z(E;F)

%01 = A, %02 = B, %03 = C, %04 = D, %05 = E, %06 = F

The right side indexes are coindexed with the left side ones

X(%01;%02):=Y(%02;%01); (the first node of X is the second node of Y - index explicitly assigned)

X(;):=Y(;); (the first node of X is the first node of Y - indexes assigned by default)

Indexes can be replaced by user-customized labels made of any sequence of alphabetic characters and underscore

X(A,%a;B,%b)Y(C,%c;D,%d)Z(E,%e;F,%f)

%01 = A, %02 = B, %03 = C, %04 = D, %05 = E, %06 = F and

%a = A, %b = B, %c = C, %d = D, %e = E, %f = F

Numeric characters cannot be used as user-customized labels

X(A,%03;B,%05)

%01 = A, %02 = B (there is no %03 nor %05)

To avoid ambiguities, users are strongly recommended to replace default values by customized labels

• X(A,%a;B,%b)

instead of simply X(A;B) or X(A,%01;B,%02)

In case of sub-nodes, the parent node must be informed by the syntax <PARENT NODE><CHILD NODE>, where <PARENT NODE> may be, itself, a sub-node

X(Y(A;B);C)

%01 = Y(A;B), %02 = C, %01%01 = A, %01%02 = B

X(Y(Z(A;B);C);D)

%01 = Y(Z(A;B);C), %02 = D, %01%01 = Z(A;B), %01%02 = C, %01%01%01 = A, %01%01%02 = B

Indexation is not affected by repetition

X(A;B)Y(A;C)Z(A;D)

%01 = A, %02 = B, %03 = A, %04 = C, %05 = A, %06 = D (and %01 = %03 = %05)

Empty nodes are also indexed

X(;)

%01 = first node of X, %02 = second node of X

Indexes may be used both in the left and in the right side of rules

X(%a;%b):=Y(%b;%a); (the first node of the X relation becomes the second node of the Y relation)

X(%a;)Y(%a;):=Z(%a); (if the first node of the X relation is the first node of the Y relation then make it the single node of a Z relation)

Indexes may also be used to transfer attribute values expressed in the format ATTRIBUTE=VALUE

X(A,%a,ATT1=VAL1;B,%b):=X(%a;%b,ATT1=%a); (the value "VAL1" of "ATT1" of %a is copied to the node %b)

Special indexes (#) are used to make reference to the internal structure of the field <NLW> in the dictionary

(X)(Y):=(X,#02)(Y)(X,#01);

The rule above is used for complex dictionary entries such as:

[[A][B]] "uw" (X, #01(ATT=AAA), #02(ATT=BBB)) <flg,fre,pri>;

It means that, given (X)(Y), the output should be (B)(Y)(A).

Disambiguation Rules

Apart from the Transformation Rules, the UNL Grammar also comprises Disambiguation Rules, which are optional and may be used to:

• Prevent wrong lexical choices;
• Provoke best matches;
• Check the consistency of the graphs, trees and lists.

The formalism here presented is directly inspired by UNLC former co-occurrence dictionary and knowledge-base. The structure of the rule is as follows:

STATEMENT=P;

Where
STATEMENT is any network, tree or list relation; and
F, which can range from 0 (impossible) to 255 (possible), is the probability of occurrence of the STATEMENT

Examples:

agt(VER;VER)=0;

(PRES,@past)=0;

Disambiguation Rules are used, both in natural language analysis and generation, to induce (or prohibit) some rules or lexical choices. In what follows, we present an example of Disambiguation Rule and its role in the UNL System.

UNL DICTIONARY

[the] {} “” (POS=DEF) <EN,0,0>;

[book] {} “book(icl>document)” (POS=NOU) <EN,0,0>;

[book] {} “to book(icl>to reserve)” (POS=VER) <EN,0,0>;

INPUT

The book is on the table.

DISAMBIGUATION RULE

(DEF)(VER)=0;

FUNCTIONING

Before the Disambiguation Rule, there would be two candidate UWs for the NLW “book”. The Disambiguation Rule prevents the second alternative, which has the value “VER”, to apply.