ECAI.

84: Advances in Artificixl [nfslligence

T. O'Shea (Editor)
Elsevier Science Publishers B.V. (North-Holland) 147
@ ECCAI, t984

A SET-ORIENTED SEMANTIC NETI'IORK FORMALISM FOR THE

REPRESENTATION OF SENTENCE MEANING

Jean-Louis Binot
Computer Sc'ience Department
Institut Montefiore, B2B - Sart Tilman
University of Liège
Liège - BELGIUM

present SF ("Sentènce Formalism"), a nev,, highly-struc-

ÿ'le
tured set oriented semantic network formalism for the rep-
resentation of sentence meaning. SF is the output formalism
of a domain-independent parser of French sentences called
SABA. Its main feature is the new concept of $definition
structure", which not only captures the necessary asser-
tional/descriptional distinction, but also allows an easy
representation of references and plays a significant role
in the interpretation of quantifier structures,
1. INTRODUCTION

There has been recently, as noted in [5], a surge of interest for portable NL in-
terfaces.One key problem in the design-of such an interface is the choice or de-
sign of a domain-independent output formalism, that is : a formalism able to re-
present sentence meaning independently of any specific application domain.
[,le present here a new semantic network formalism called SF ("Sentence Formalism,')
for the representation of sentence meaning. SF will be output by a modified ver-
sion of a domain-independent parser of French sentences called SABA [11. SF is ba-
sically set-oriented. Its key feature is the new concept of "definitfoi structure",
which not only captures the necessary distinction between the descriptional and
assertional parts of a sentence; but offers an easy way to represent pronouns and
noun-phrases referents and, lastly, plays a significant role in the interpretation
of quantifier structures.
Sf doesn't compare with semantic network formalisms primarily aimed,at knowledge
representation, such as those of Hendrix [4] or Schubert et al. [10]' because such
formalismsare basically domain-dependent. They presuppose indeed, in order to re-
present a sentence, that the nodes representing the objects talked about by the
sentence have already been identified in the knowledge base of the application
domai n.

SF has mayoe more in cormon with sonre logical formalisms specifically aimed at sent-
e-nqe meaning !epresentation, such as thoie of Nash-llebber & Reiter [8] or ]loore
ü1, the later being used in the TE/ll'l NL processing system [31. There is however a
basic difference of approach. These formalisms start with aliedicate calculus no-
tation, which is gradually extended r',ith types, sets, additional quantifiers and
other features in order to achieve a sufficient expressive power. SF, on the con-
trary, is fundamentally set-oriented i at its core are the set equality and set
inclusion relations expressed by our definition structures. l{e believe this leadd
to a more unified account of noun-phrase structures, including reference and quafl-
tification aspects.
Although our work is primarily concerned with French, all the examples presented
hereafter have been, for the sake of readability, translated into English.
148

?. FUNDAMENTALS

I n order to set the stage for the fol I ow'i ng di scussi on , wê shal I brief 1y descri be
the complete SF representat'ion of example. Let us consider (1). Its complete
SF representation is shown in (2) în
( 1 ) Les ehiens pour.suiuent LLn chat nol-y,
The dogs are following a black cat
(?)
tr
i-
- )tÆ --)rnt
x I AGENT r- 9BJECT t:'Y
>1 *___ _)ûç
I
__;riÂr,
ÿ coloR
<*>( - - -- *
,ü 0

BlÊ.0

Eq
Such a representation includes three types of structures :
a) there is first the "relational structure" of the sentence, shown in (3). This is
a traditional "case-like" structure : the reiation is represented by a node, to
which arguments are tied by role links (note that names of roies usèd in this pa-
per are oniy illustrative; in practice, a specific set of roles can be defined for
each application domain).
(3)
AGENT OBJET

x follow v
FACT DEC FACT

:
Nodes in SF are of three types RELATIoNS (drawn as'0"),oBJECTS itr*") and FUNC-
TIONS (not discussed here). Each node has a M0DE. The mode "follows" (3) is
of in
DECLARATM (DEC) : it corresponds to an asserted instance of a relation. The mode
of the x and y nodes is : these nodes refer to specific entities.
FACTUAL (FACT)
0ther possible modes are (both seld-explaining), and DE-
INTERR0GATIVE, IMPERATIVE
PENDENT (which characterizes relations in subordinate clauses and vaiues of attri-
butes, like "black" in (2)). Our use of modes is inspired from Simmons modalities
[1], although we don't use the same modes as he.
Besides role links, SF has also "attributive links" (like the C0L0R link in (2)),
which are used to refer to specific properties of objects.
b) (3) expresses only the assertional content of the sentence. The next step is to
provide each factual node with a "definition structure" (4), which expresses the
descriptional properties allowing to identify the specific entities represented by
that node ("the dogs" for x, "a black cat" for y). Definition struètures are dis-
cussed in section 3.
c) Lastly, relation and argument nodes are organized in a "binder structure" - the
nodes drawn as '[R]" in (2) - which specifies the scope of the involved quanti-
fiers. Binder structures are discussed in section 4.
 t49

(4)

X AGENT OBJ ECT v

) 0<
follow

DEP

3. DEFINITION STRUCTURES

EACH FACTUAL NODE IN SF REPRESENTS A SET OF ENTITI ES ( tfre case of an uni que ndi -
'i

vi dua 1 bei ng handl ed as a s'i ng I eton ) . Thi s set 'i s defi ned by descri pti ona'l 'inform-
ation provi ded by the sentence (ad5ect'ives, compl ements , Ï^e I ati ve cl auses ) . Tl'»is
i nformation wi I I be expl i ci t1y represented i n a "contextual defi n'iti on structure"
(5).
(5 ) (contextual def i ni t'ion structure) ! I =
(factual node) + (s'ize)
ü (definition l'ink)
(set def i n'i ti on structure)
The idea behind such a structure is to relate two sets : the "defined set" D -
represented by the factual node -, which is the set of entities the speaker.v',ants
to talk about, and the "definition set" E - represented by the "set definition
structure" - which is the set that can be identified with the contextual infonna-
tion provided by the sentence. The different items of (5) are discussed below :
a) the "size" of the defined set is expressed by two properties ass'igned to the
factual node :
0PERATOR : ),=,)=,/=,(,or (=
NUMBER :apurenumber;
b) "definition 'links" are of two types :
- therrtotal definition n ("Ttr), which expresses set equality: D is described
as equal to E,
- the "partial definition" ('Pu), which expresses set inclusion : D is described
as included in E.

The choice between these types depends on the definite or indefinite nature of the
represented term. To exemplify points a) and b), (6)a shows how "a black cat" is
represented by a factual node with size "=1", related to the set of all black cats
by a partial definition.
(6) a : b

(the set cats> (the set of al I cats>

150

c) "Set definition structuresrrare used to represent definition sets, like "the

set of all black cats". They are of three types : generic, by name and'by referen-
ce.

9srçric-ÉeIirl!islt
A generic definition structure defines a set E by describing the properties of its
prototypical element (represented by a generic node, drawn as "(*)") and by indic-
ating a larger set E1 in which E is included. Thus the set of all black cats is de-
fined in terms of the set of all cats as in (6)b.
Note that (6)b introduces a new set, which must in turn be defined. Generic struc-
tures induce a recursive definition process, which must bottom-out sonrewhere I Con-
sequently we need a set definition primitive. It will be the "conceptual class,,,
defined as the class ôf all entities conforming to the -supposed)y known - defin.i-
!!on of the corresponding concept. In practice, a set of concepts should be spec.i-
fied for each application domain; we shall simply assume here that a concept exists
for every meaningful word. A conceptual c1ass, like "the set of all cats,,, will be
represented by a concept node, drawn as in (6)c. (6)a, b and c put together yield
the entire definition structure of "a black cat" as used in (4).
Generic definitions allow also easily to represent relative clauses. Thus (8)
shows the SF representation of Woods notorious example ( Û2],p.61) :
(7) The dog that bit the man had rabies
(B)
BENEFICIARY OBJ ECT

It can easily be seen that this representation avoi"ds the "shared-subpart-fa1lacy,,

problem pointed out by Woods [tZ], (e) is the representation of (7) büt can in nô
way be understood as the representation of (9) :
(9) The dog that had rabies bit the man.
QeIlri!isr:-!y-rcue
Definition by name occurs when a proper noun is used in an NL utterance, as in
(10) :
(10) lnon conduit La uoitw,e de PauL
drives Paul's car
Jean
The definition set of the entity "Jean" is taken as the set of all entities whose
name is JEAN. This set will be represented by a "named set node", drawn by putting
the name inside rectangular brackets, as in (11) :
9etlri!is!e-qy reference
We have a1 ready seen howdefinition structures can represent adjectival and nomi-
nal complements ((2) and (11)) and relative clauses (B). A sl ight extens'ion of the
 l5l

(tt1

Ol^JNED-BY
--0

same concept allows SF to express also pronoun references. This extension consists
in admitting that a factual node can be used as <set definition structure> for an-
other factual node. The defined set D corresponding to the first will then be ta-.
ken as the definition set E of the second. Thus, (12) will be represented as in
( 13) :
(12) .racques tz.atta.t.LLe dans Le jardin. IL tond La pelouse.
Jacques works in the garden. He mows the lawn.
(13)
OBJ ECT
0Ç-
mow

oê -sJryAIIqN
work

The same extension applies to the representation of definite noun-phrases refe-

rents. Thus "woman" and "animal" in (14) refer.to two entities previously mention-
ed. This is easily represented as shown in (15).
(14) Uarthe trat:aiLlait dans Le jazdin. Ltn Loup sottit du bois. Ltaniral attaqua
La fermne.
Marthe worked in the garden- A wolf came out of the wood. The animal attacked
the woman.

(15)

*_srJUATroN_;, oÇ jLEU -* tr -4
-l
work

[wnnrH r]
152

Although we didn't investigate these aspects further, we believe examples such as

(13) or (15) show clearly that definition structures could prove a powerful tool
for research on the representation of definite and indefinite descriptions and on
noun phrase resolution.

4- LOGICAL ADEQUACY AND BINDER STRUCTURES

A representation formalism is logically adequate if it has at least the same ex-

pressive poÿÿer as formal logic. This property is important since formal logic cones
with a clearly defined semantics onto which a sound theory of inferences can be
built. I'le have taken care to ensure the logical adequacy of SF, as we show below.
BreÉ1se!c:-end-!eru:
A basic formula of formal logic is obtained by applying a predicate to some terms
(constants, variables, functions or even sets in a set-oriented formalism). More
complex formula can then be obtained through the use of connectors and/or quanti-
fi ers.
Predicates correspond in SF to non-factual relations, which can be expressed ex-
plicitly as relation nodes or implicitly as attributive links (see section 2).
Terms are expressed in SF by factual nodes. A factual node has four properties
which allow to capture entirely its Iogical status, including the nature of the
quantifier possibly applied to him. These properties are the NUMBER,0PERATOR,
VIEW (defined later) and the type of definition link starting from the node. Their
values are attributed to factual nodes on the basis of the determiners leading the
noun-phrases they represent; the table below indicates the values chosen for some
cormon French determiners used in declarative sentences. Some important aspects
of factual nodes arê discussed in the next subsections.

FRENCH ENGL I SH OPERATOR NUMBER DEFINITI0N VIEt^l

Les the(p1 ural ) >1 T UNKNOUüN

ehaque every, each )=0 T DISTRIBUTIVE
tous al I )=0 T UNKN0I^JN
Le the(singular) -1 T UNKN0I^JN
des, eertains some(p1ura] ) >1 P UNKNOWN
un a, one -1 P UNKNOI^JN
au moins n at least n )=n P UNKN0IJN
nreæaetement n n,exactly n =n P UNKNOI^,N
Lesn ' the n =fl T UNKNOI^IN

Constant nodes
A constant node is a node with size "=1" and with a total definition. l,le only con-
sider here nodes of 0BJECT type. Such a node can be interpreted as a logical cons-
tant, provided that its definition structure allows indeed the identification of
an unique object of the application domain ! Thus a proper noun like Jean can be
translated as the logical constant JEAN if only one object of name JEAN is known
about. The verification of this condition should be made by the application pro-
cess interpreting the SF formalism.

Ieri.e!Ier-e4-!vpel
A non-constant factual node can be interpreted as logical variable. A network such
as (3) is then viewed as the expression of an open formula :
( 16) F0LLOl'l(x,y)
If we add the definition structures, as in (4), we still have an open formula, but
one in which each variable is constrained to take only values of a certain type :
 153

(tl'1 FoLLoW(x,y) IIHERE xctset(e) lmember(e,dogs))

AND yc{Set(e) lmember(e,cats).color(e,black))}
There are of course other ways than (17) to express types in 1ogic, one of the
most interesting being the 'lambda notation presented in [B].
uantifiers and binders
A non-constant node is usually bound by some kind of quantifier. Two things must
then be indicated : the nature of the quantification and its scope.
A.FUNDAMENTAL PROPERTY OF SF IS THAT THE NATURE OF THE QUANTIFIER IS ALUIAYS INDI-
CATED BY THE PR0PERTIES 0F THE FACTUAL N0DE ITSELF. To simplify, Iet us discuss
for now only the classical universal and existential quantifiers. The key property
to choose between them is the type of definition link starting from the node : a
total definition corresponds to an universal quantificati.on, a partial definition
to an existential one. This can easily be justified by refeming to the sets D and
E defined in section 3 : for a totally defined node, a relat.ion asserted of the
elements of D will be true of all elements of E, while for a partially defined
node, such a relation will be true only of some elements of E.
The scope of the quantifier, on the other side, will be indicated explicitiy by
a "binder structure" (figure (18)). A binder can be described as a built-in func-
tion with two arguments "r" and rrtrr ' rrrrr is tied to a non-factual relation node
expressing an open formula and "t" is tied to a factual node representing a term
which is still free in that formula. The result of the binding is a new formula
in which the term is bound by the appropriate quantifier. This new'formula is re-
presented by the binder node itself (drawn as "[R]"); a new binder may be in turn
be applied to it. Thus any combination of quantifiers can be represented by suc-
cessive bindings. A constant node, of course; doesn't need to be bound.
(18)

t,rle are at least able to understand the complete representation of (1) asshown
now
1n (2). to the x node in (2) expresses formula (19); the sec-
The binder applied
ond binder, applied to that formula and the y node, yields (20).
(19) (Vx lxc{Set(e) lmember(e,dogs)}) F0LLOl,l(x,y)
(20) Gy lyc Set(e) lmember(y,cats). color(y,black)]
(vx lxctSet(e) lmember(e,doss)]) F0LLOt'l(x,y)
Our binders are inspired from Shapiro [S], who used special nodes to represent-
quantifiers. The main difference howeveris that in our case the nature of the
quantifier is not specified by the binder node, but is viewed as an intrinsic prop
erty of its argument.
I!e-qi:!riDu!ivelssllec!1ve-9is!irs!ier
Until now, we only considered factual nodes in a "distributive" vüay : the proper-
ties asserted of the set represented by such a node were held as separately true
of each member of this set. Following Martin
rrcollective" interpretation, under which a set[6],
we want to introduce also a
should be considered as a whole en-
ti ty.
In SF, the distributive/collective distinction is considered as an individual prop-
erty of each factual node, the meaning of which is defined as follows :
- if a collective factual node is argument of a relation, that means that all mem-
bers of the set represented by the factual node are simultaneously and collecti-
vely participating in an unique action, event or state described by the relation;
- if a distributive factual node is argument of a relation, that means that each
member of the set represented by the factual node participates individually in a
separate action, event or state of the kind described by the relation.
154

0f course, for a constant node, there is no difference between the two interpret-
ations.
The distinction will b indicated by the view property of the factual node. Graphi-
cally, a collective node will be drawn as "(/)'. Collective nodes can be bound ,as
shown the following example:
(21) The policemen are tracking bandits.
Suppose that we are interpreting "the policemen" iteratively and "bandits" collec-
tively. The sentence has two readings. The first is that each pol iceman js track-
ing his own set of bandits; the second is that they are all tracking.by separate
means the same set of bandits. If we accept to use a logical formalism with set va-
riables, these interpretations can be respectively expressed as (22) and (23);
(22) (l set S lS C {band'its})(vx lxc{policemen})TRACK(x,S)
(23) (vx lxc{policemen})(3 set S ls C {band.i ts})TRACK(x,S)
In these formulas, " fset S lS C {...})" should be read as : "there exist a set S
such as S is included in ...". C1early, the distinction between the two is a matter
of quantifier scope. The corresponding SF representations are shown in (24)a and U
respecti ve1 y.
(24)a :

t--
t
-
I
I It lr I

I AGENT l-lqIryI ) b<-oitcr-l )=1

track

Iuuer jsel-sge!!1I!er:
As it is well known, NI allows more quantification schemes than the classica] uni-
versal and existential ones. I'le shall discuss briefly the case of numerical quanti-
fiers. Let us consider a famous example due to Woods (fl21,0.72).
(ZS) fhree lookouts saw two boats
Woods himself saw three possible readings of (25) :
- "each of the three lookouts saw two boats (for an unknown total of boats between
two and six)"
- "each of the two boats was seen by three lookouts'r"
- "one group of three1ookouts that jointly participated in seeing a group of two
boats " .
The difference between the two first readings is only a matter of quantifier scope.
They are represented in SF in (26)a and b.
(zo)a :

t t
tRl [n]'l( -1 t-- )tRl--r - )tRl
^l1 ^
I I ,{\ I
I lp It
AGENT lr OBJ ECT
I

I
*- AGENT OBJECT I

) 0(
I I

-*
I
*- 3 -- )o(
a
SSAW SAW
P
 155

From the third reading, it is obvious that Woods makes also some kind of collect-
ive/distributive distinction. Now, the interesting point is that if we combine ex-
haustively the two binder structures of (26) with the collective/distributive view
for each iactua] node, we get not 3, but I different vÿays to represent (25). Two
of these ways being logically equivalent, we obtain SEVEN different readings for
the sentence! These read'ings are enumerated in (27). To express them logically, we
have introduced two new notations : a numerical existential "flnx" ("there exist
n x") and a collective existential "Jlset n" ("there exist a set of cardinality
n").
(?7):
a. (3 set #3 S1 lS1 c {lookout})(312x lxc{boats}) SEE(Sl,x)
"A set of three lookouts have seen, in two separate events, two boats"
,b. (312x lxc{boats})(3set#3 S1 IS1 C {lookouts}) SEE(S1 ,x)
"The same two boats have been seen separately, each by a set of three lookouts"
c. (313x lxc{lookouts})(3set#2 S1 lS1 C {boats}) SEE(x,S1)
I'Three lookouts have each seen a set of two boats"
d. (fset#2 Sl ls1 C {boatsi)(fl3x lxc{lookouts}) SEE(x,sl)
"A same set of two boats has been seen separately by three lookouts"
e. (ll3v lyc{lookout})(312x lxc{boats}) SEE(y,x)
"Thrèe lookouts have each seen separately two boats"
f. (312x lxc{boats})(313y Iyc{lookouts}) SEE(v,x)
'iTwo bôats have each been seen separately by three lookouts"
9. [A(3 set43 51 lsl C {lookouts})(fset#2 sz lsz C {boats}) SEE(SI,S2)
set of three lookouts jointly saw a set of two boats"
h. (fset#2 sz lsz C tboats])(fsetff3 51 lsllc {loorouts}) SEE(SI,S2)
equivalent to g.
l,loods interpretat'ions correspond respectively to readings e,f and g. This example
shows the considerable flexibility that can be obtained in SF.

5. CONCLUSIONS
SF is a highly-structured, powerful, domain-independent and 1ogica11y adequate se-
mantic network formalism for the representation of sentence meaning. These charac-
teristics are provided by the two key concepts, discussed in this paper, of "def-
inition structures'r and of "binder structures". Many other features of SF could not
be discussed here, notably the representation of negations, of connectors and of
comparatives, the use of "selectors" to represent ordinals as well as superlatives,
and the concept ofrrsimultaneous binding". A more detailed discussion of 5F can be
found in [2].
SF has also the additional and important property of being an "incremental" repre-
sentation system: as can be seen in figures (3),(4) and (2), each of the three
steps of the representation of a sentence is carried out by adding a new structure
and without modifying the already constructed part of the network. This property
facilitates the design of a NL parser for producing SF structures.
SF is to be output by a domain-independent parser of French sentences called SABA,
implemented in MACLISP on a DEC2050. SABA is basically a case-frame instantiation
system with a pure bottom-up parsing strategy and an unusual right-to-left parsing
order ['l]. A first version of SABA, including a simplified'version of SF, has al-
ready been tested on a medical application. An improved version supporting the en-
tire SF formalism is currently underimplementation. Our long-term goal is to de-
velop a portable NL interface for expert systems,
ACKN0t,ILEDGMENTS

Thanks are especially due to three people : Ph. Lemaire, who helped to design a
first version of the formalism presented here (this version! albeit imperfect and
156

not'logically adequate, contained already the outline of the concept of "defini-

tion structure"), P.-J. Gailly, who provided many helpful critics ànd suggestions,
and D. Ribbens, for useful comments on an ear'lier draft of this paper.

REFERENCES

[t] Oinot, J-L., Graitson, M.,Lemaire, P. and Ribbens, D., httonatic processing
of ttritten Erench Langnge, p?oc. of C0LINGBO (Tokyo, 1980), 9-14.
[Z] ainot, J-L, S.F : a Sentence FoymaLism, internal report, Dept. of Comp. Sc.,
Lièse Univ. (1984).
[3] Grosz, 8.J., IEAl,l : a transpoqtable natutaL-Langtage intez,face s7stem, proc.
of the Conference on Applied Natural Language Processing (Santa-Monica, '1983),
39-45.
[4] Hendrix, G.G., Eneoding knoaledge in partitioned netuoz,ke, in Findler, N.V.
(ed.), Associative networks : representation and use of knowledge by computers
(Academic Press, 1979).
[5] Hendrix, G.G., NatutaL language Interface, Summary of a Panel of the Workshop
on Applied Computational Linguistics in Perspective, JACL 8,2 (1982).
[6] Martin, N.A., RoLes, co-descriptots and the fortnal representation of quanti-
fied English erpz,essions, MIT/LCS tech. memo 139 (1979).
ÿ] Moore, R.C., Problems in Logical fonn, proc. of tgth meeting of ACL (1981),
1 17 -124 .

[8] Nash-Webber, B.L. and Reiter, R., AnapVnra and. LogicaL forrn : on formalmean-
ing representation fot, natutaL T,angtage, proc. of IJCAITT(1977),121-131.
[S] Strapiro, S.C., A net sttactute for semantic inforration storage, deduetion
and retriettal, proc. of IJCAITl (1971),51?-523.
[10] SctruUert, L.K., Goebel, R. and Cercone, J., The structute and. otganisation of
a senwntic net for conprehension cnd inference, in Findler, N.V. (ed.),
Associative networks : representation and use of knowledge by computers
(Academic Press, 1979).
[1t] Sirrnons, R.F., Semantic networks : tTpiy conrputation anduee for undetstal.d-
ing e_ngLish.sentencee, in Schank, R.C. and-Colbÿ, K.M., (eds.), Computer
models of thought and language (Freeman, 1973).
[tZ] Wooas, W.A., What's in a Link : foundations for semantie netuotke, in Bobrow,
D.G. and Collins, A.C. (eds.), Representation and understanding (Academic
Press,1975)