INTERPRETATION OF NOUN PHRASES IN INTENSIONAL CONTEXTS 
Leonardo LESMO and Paolo TERENZIANI 
Dipartimento di Infor,natica 
Universita' di Torino 
C.so Svizzera 185- 10149 TORINO- ITALY 
ABSTRACT 
This paper presents a network formalism for representing the 
meaning of noun phrases occurring in the context of intensional verbs 
such as seek and want. The basic assumption is that an intermediate 
representation is useful to carry out the interpretation process perspi- 
cuously. The proposed representation keeps apart de-re (transparent) 
and de-dicto readings, stating, by means of suitable arcs and nodes, 
that in the first case there is a real counterpart of the described entity, 
whereas in the second case no such counterpart exists. Moreover, tbe 
concept of relevance of the description is emphasized and the rela- 
tionships between intensional contexts and the dicotomy existing 
between value-free and value-loaded readings of definite descriptions 
is discussed. Finally, t!ae impact of the intermediate representation on 
the discourse history is considered, in order to explain how such a 
representation accounts for the contextual data and updates them 
according to the informational conte,tts of the sentence being 
analyzed. 
1. INTRODUCTION 
Formally, an intension is a mapping that associates with eve,y 
possible world (actually, with indices that identify possible worlds) a 
denotation, where a denotation is a set of entities (individuals, rela- 
tions, functions). If we take a term, say "animal", it denotes, in the 
present world, a set of individuals; but the meaning of the term can- 
not be characterized by that set. For instance, the terms "human" and 
the term "featherless biped" denote the same set, but their meaning is 
different, since the fact that there are ~ featherless bipeds except 
human beings, in this world, e.'m be regarded as a mere accident. It is 
not difficult to conceive of a different world where another species 
shares the same property with humans. 
The characterization of possible worlds has been subject to 
deep philosophical debates, which were mainly concerned with the 
relationships existing between "conceivable worlds" and "possible 
worlds", as well as with the connections existing between alternative 
worlds, but, for tim present purpose, we only need to note that: 
- The concept of "possible world" is rather easy to render formally, 
expanding the usual concept of interpretation in logics by introduc- 
ing "indices" on worlds. Thus, for instance, the mapping which 
assigns to each predicate symbol a relation (of the same arity) on 
the universe, must be viewed as a set of mappings (one for each 
world), since the denotation of the predicate can vat 3, fiom world 
to world. Equivalently, the intension of a n-ary predicate can be 
delined as a function from world indices to n-sty relations. 
- The concept of "possible world" is rather common in the everyday 
practice. For instance, in saying 
(1) John seeks a unicorn 
no relation between John and a real world entity referred to by % 
unicorn" can be assumed, since no such entity exists. Nonetheless 
(2) John seeks a centaur 
is different fi'om (1). Intuitively, the expression "centaur" has a 
meaning different fiom "unicorn", because in some (fictitious) 
world they are used to refer to different entities. The same cannot be 
said for expressions such '.,s "centaur" and % being hall' human and 
half horse", since the second can be used to define the first ~. 
On the other hand, the fact that possible worlds are common to logics 
and intuition does not mean that there is an exact match in the way 
the concept is used. 
Let's consider, for instance, the best known treatment of 
natural language in logics, i.e, the work of Monlague's. In order to 
maintain the strict compositionality of his approach, Montaguc llad to 
make the interpretation of linguistic entities considerably complex a. 
For instance, the interpretation of a NP is given as a set of properties, 
where a property is a function from indices to set of individuals. For 
instance the property "redness" is that function that, for each index, 
says what is ted at that index. Of course, this is intuitively accept- 
able, but why should "John" be interpreted as a set of properties? It is 
true that with any person we know we associate a bundle of proper.. 
ties that that person has (i.e properties such that the individual 
denoted by "John" happens to be in their denotation evaluated at this 
index), but it seems also true that when I say "John" I refer to an 
individual and not to his properties ~. Even more complex is tim 
treatment of NP's in object position; for instance the phrase "a 
unicorn" in (1) is interpreted as a property of properties, in order to 
accont for the fact that, loosely speaking, sets have to be evaluated 
"here and now" and so the set of properties that t, unicorn has is the 
same as the set of properties that a celltaur has. 
We cannot go in further details here (more detailed informa- 
tion can be found in \[Dowry et al. 81l), but a couple of questions 
naturally arises: Is it possible to maintain the basic principles of a 
possible worlds analysis without introducing the complexities of 
Montague's inmnsional logic? Is it possible to introduce an inter- 
nlediate level in the semantic interpretation process, which preserves 
the idea of compositionality? 
Recently, two interesting proposals concerning the semantics 
of NL have been put forward. It seems that they can give some 
suggestions about how to answer the questions above. We will briefly 
outline them here, but we must first notice that we do not aim at 
presenting the philosophical position of those authors, since the dis- 
cussion that follows is strongly biased to the views we hold about the 
1 This can simply be due to my scarce knowledge about mythology. If there 
is a more precise characterization of centaurs, such that it is possible to conceive of a 
world where a being half human and half horse is not in the denotation of "centaur", 
then the same situation as "featherless biped" arises. 
2 We cannot discuss here our disagreement with a conception of syntax that 
assigns to a sentence as (1) two different syntactic stnmtures in order to cope with 
the so-called spec~c and non-specific readings (the one where a particular unicorn is 
wanted vs. the one where any unicorn could serve John's purpose). We can only 
note that, in our view, syntax is mainly concerned with surface phenomena. Note 
also that in \[Cooper 83\] a Montague-based proposal is presented, that overcomes this 
difficulty. 
3 Formally, the two alternatives are equivalent if we assurae that no two indi- 
viduals share the same "total" set of properties. For this reason, the solution proposed 
by Montague is perfect, if we assume, as he did, that no intermediate representation 
of the meaning of linguistic entities is ever built. In fact, the formulae of intensional 
loeics are only tools to renresent constraints on oossible worMs, so tha~ their com- 
plexities was not a matter at hand. As we will see, our position is different. 
37B 
topic at hatuL 
The first proposal is "Situation semantics" by Barwise and 
Perry 1183\]. They view the meaning of a sentence as a relation 
between the discourse situation (who are the speaker, the addressee, 
etc.), the connections of the speaker (i.e. its relationships with the 
environmenl), and a described situation (the event, or "type of event", 
which is being described). What is flmdamental, in our view, is the 
rejection o1' truth values as "senses" of sentences. This move accounts 
for the non-.equivalence of attitude reperts involving logically 
eqtdvalent embedded sentences. For instance, 
(3) John knows that two plus two is fonr 
is not the same as 
(4) .iohn knows that the square of 37 is 1369 
although bt,th facts are necessary lruths (with respects to the 
difference Imtween meaning and intension, see also \[Lewis 72\]). 
Also important is the exploitatioo of what is called ,'esource 
situation in the interpretation of constituent of sentences. In the 
present concern~ we. can fix our attention on NP's. Let ItS consider, 
for example, the meaning of defufite noun phrases: 
"We can also think of. d,c \[the n\] 4 as giving a partial function 
fi'um situations e to individuals a~,: 
d,c \[tbe u\] re) is defined and = ae 
iff 
d,c \[tile u\] a,,e " IBarwise & Perry 8L pp.149-150\] 
where the latter means that in the context where tile expression is 
evaluated tb~ "rehttionar' meaning of u umst single out exactly one 
individual. If we have available a particular resource situation, then 
we get an individual (a,,), but if such at context is not awdlable, then 
the interpretation of the NP is just the non-evaluated function fi'om 
situations to individuals. In the lirst case, we get the referential (or 
value-loaded) use, since any predicate the sentence expresses is about 
the referred individual (a~,); in the second case we get the attributive 
(or value-free) use; what is the predicate about in this case? It will be 
about the eventual (possible, future, fictional) fillet" of rite description. 
The simplesl way to obtain such a resuh is to incorporate the descrip- 
tion of the individual into tile inlerpretalion of the sentence, so that, 
whenever th.~ actual fillet" will be avaihtble, the predicate will be true 
of it. This approach does not seem to us in contrast with a possible 
worlds analysis: in each such a world there will be all individual 
denoted by the definite CXl)ression (possibly different individuals at 
different indices); thus, the property of satisfying rite description will 
imply the property expressed by the predicate. In any case, the 
iuterpretatior, of sentences on the basis of some specific contextual 
situation suggests that the context is made explicit, in particular the 
connections of the speaker; of course, in an "nnderstanding" system 
(be it a human or a computer) such counectious are, at best, 
presumed, that is the hearer must "imagine" the connections on tire 
basis of a "shared" environment or whatever she/he knows about tile 
speaker. More important, in case the hearer is a machine, no real 
sharing of environment can be assumed, so that it must use an "inter- 
nal" representation of the outer context; In the present work, this 
representation is available in terms of general world knowledge 
(represented in a KL..ONE like format), plus the history of the user- 
machine dialogue carried on so far. 
Another piece of research that must be taken into flccoant con- 
corns mental apaces \[Fauconnier 8511. This framework provides us 
with a tool for representing the relationships between the individuals 
mentioned in a sentence and their particular view of reality. We 
believe that this is one of tire most common nses of the pro- 
theoretical notion of possible world. Of course, tile knowledge, the 
beliefs, the opinions, the desires of different persons are not the same, 
and when we want to understand natural language we must try to 
model these differences. Although mental spaces are not claimed to 
be such a modelling technique s, nonetheless they seem useful in 
modelling the process that leads to the construction of rite correct 
interpretation 6. 
4 "d" stmtds for the discourse situation, c for the "connections", " \[the 1¢\] " for 
file meaning of 'the 'g', "e" for the described situation. 
This paper does not present an implementation of the mental 
spaces approach, but uses one of the basic ideas of that work: some 
linguistic contexts "set up" local spaces that constrain the meaning of 
the entities occurring in them. So, a mental space represents a local 
view of linguistic entities, that may be used to isolate the (referred) 
beliefs, desires, etc., of the mentioned individuals. 
It should be clear from the above discussion that we are not 
adopting in full any of the positions of the reviewed works. It seems, 
however, that each of them can have a strong impact on the overall 
framework. Thus, an adaptation of mental spaces can help in model- 
ling the process of building the meaning of sentences; a "situational" 
analysis can help in unde,'standing how such a meaning must merge 
wifll the environment; a possible world approach can provide the 
basic insights about the tinal result of the interpretation. This paper 
is mainly concerned whh the first of these three aspects. 
2. OVERVIEW (IF THE PtlENOMENA TO MODEL 
This paper is concerned with the interpretation of noun 
phrases and in particular with NP's occurring in intensional (or 
opaque) contexts and to the dichotomy existing between value-loaded 
and value-flee interl)retations. The final goal is to show bow the for- 
malism of lri-eoncepts (introduced in \[Di Euge,tio & I.esmo 87\]), 
extended in a suitable wily, accounts for tile phenomena under study. 
Particularly rclovaut iil the present context seem tile works by 
Fawcett and Hirst \[19861 and Kronfeld \[1986\]. The theory oulined in 
\[Fawcett & Ilirst 1986\] implies the existence of seven different read- 
ings of the sentence 
(5) Nadia wanls a tlog like Ross's 
Those readings depend on three different parameters: 
I) Transparent (de re) vs. opaque (de dicto) 
II) Source of description (the agent, i.e. Nadia, vs. the speaker) 
Ill) Time of application (present vs. future vs. repeated) 
We believe that only three readings are directly related to the infor- 
mational conleut of the sentence, since neither the source of the 
description, n(~r the time of applicatiou scent to be among the con- 
corns of the sentence itself. For instance, leawcett and Hirst distin- 
guish betweeu 
a) Transparent reading, agent's description 
(Nadia says "I want that dog, which is just like Ross's". The 
speaker reports this as (5) perhaps without being familiar wilh 
Ross's dog) 
b) Transparent reading, speaker's description 
(Nadia says "I want that dog", and tbe speaker, who knows 
Ross's dog, reports this as in (5)). 
The problem is that (5) does not seem to be concerned with what 
Nadir or the spe.'tker know about Ross's dog; it only assumes that "a 
dog like Ross's" is a description wtlid for the bearer of (5) so that it 
can enable him to huild an internal representation of what Nadir 
wants. On the other hand (5) ceaainly involves a description of 
Nadia's desires; taking this into account, a perspicuous representation 
should be able to characterize those desires more precisely: Does 
Nadia wants a dog of a kind that casually happens Io be the same 
kind of Ross's dog? Or the dog she wants must be of tlnlt kind just 
because it is the same kind of Ross's dog? In other words how 
relevant is the description? 
The relevance of descriptions seems to be one of the central 
concerns of the work described in \[Kronfeld 86\]. In that paper, a 
computational model of reference is provided, and definite descrip- 
tions outside intensional contexts are considered. Some examples are: 
s "The conslruction of spaces should not be over-interpreted. They are not 
representations of realit!¢ or of partial 'possible worlds'. Ill particular, a belief space 
('Max believes ') is not a belief or state of belief; at most, it is a way of talk- ing about belief~7\[~conuier 85, p.1521 
6 "The space elements are not referred to by expressions of language; they are 
set up, identified, etc., and may then be used for purposes of reference (or possibly 
purported reference), tIowever, this distinction has an obvious and yet nontrivial 
consequence: theories of refarence based on the use of language cannot bypass men- 
hal spaces; they will have to forsake Ihe idea of a direct link between linguistic strec- 
tares and referents aod take into account the important intennediate process of space 
construction". \[Fauconnier 85, p.158\] 
379 
(6) Smith's murderer is insane 
(7) The winner of the race will get $10000 
Let's consider ex. (6): traditionally, "Smith's murderer" can be used 
referentially in ease the speaker "knows" who is the murderer; other- 
wise, it must be used attributively (i.e. when "whoever he is" is 
appropriate) 7. Note that, even in case the speaker does not have any 
clue about the identity of the culprit, still the expression refers to a 
single, well defined individual. In fact, one of the accomplices of the 
murderer, after hearing a policeman say (6) on a TV interview, could 
report that as "Hey, Bob, he told that you are insane". Analogously, 
the fact that no winner of the race exists yet does not mean that the 
reference in (7) is not definite or precise, since the "role" of winner is 
uniquely defined even when the race has not been run. Again, if Fred 
wins, it is possible to report (7) telling him "You will get $10000". 
Notice, however, that in both examples the reports are correct but not 
precise, since they fail to convey the "reason why" the thing reported 
is true, So it seems to us that: in "attributive" readings there is a com- 
ponent of meaning concerning some kind of dependency (Causal- 
ity?) 8. It is worth noting that this proposal is completely compatible 
with Barwise and Perry's. In \[Barwise & Perry 83\], in case of attri- 
butive reading, the description must be evaltmted in all the situations 
which contain an individual who matches such a description. That 
means that the assertion made about tiffs description must be relevant 
in all these situations. This implicitly states that there is a sort of 
dependency connection between the description and the assertion, and 
that this connection is valid in all possible situations which contain an 
individual satisfying the description. What we aim to show is that 
the distinction between attributive and referential readings provides us 
with an important criteriuna that also affects intensional readings. 
We believe that the two "opaque" readings we envisioned ear- 
lier are distinguished by means of the same "relevance" critefium 
used above. In classical exalnples as 
(8) Mary wants a dog 
no relevance criterium can be applied because of the simplicity of the 
description. In fitct, if Mary said "I want a nice and faithfol animal", 
either we assume that, for Mmy, the intension of "nice and f;tilbful 
animal" coincide with the intension of "dog" (i.e. in all possible 
worlds the two expressions denote the same set of individuals) or the 
report is not accurate. On the contrary, we could assume that "a dog 
like Ross's" is right only in this world (in any world she wants the 
same kind of dog, which is the same as the kind of dog that Ross 
owns in this world) or in every world (whichever is the kind of dog 
Ross owns, in any world, Nadilt wants to have one of the same kind). 
Let's consider one more example to make the point clearer (see \[Fau- 
connier 85, pp.45-46\]): 
- George says: "The winne,' will go to Hong Kong" 
- The speaker assumes that "the winner" and "the best dancer" are 
equivalent descriptions (or "equal roles in the speaker's mental 
space" in Fauconnier's te,'ms) 
- The speaker says "George thinks the best dancer will go to Hong 
Kong" 
We claim that the speaker's report is incorrect (although sincere), 
contrarily to what Fauconnier assumes, because in an "attitude" con- 
text (as "thinks" in the present example and "wants" in the previous 
case) what must be correctly modelled is the attitude of the agent. 
The speaker's description always has a subsidiary role: the one of 
describing correctly the mental state of the agent. 
3. LAYOUT OF THE SYSTEM 
TNs short section provides some information about the overall 
organization of the system which the semantic interpreter is a part of. 
The GULL (General Understander of Likely Languages) system 
works on the Italian language and is based on production nfles. The 
syntactic structures are represented as dependency trees, where all 
nodes, and not only the leaves are associated with surface words. For 
7 What "knowing" means in the present context is not very clear. If the speak- 
er determined the identity of tlSe murderer by finding his driving license near the 
corpse and this is the only thing he knows about him, then it can safely be stated 
that be "knows" who is the murderer, but still the only possible reading is the attri- butive one. 
s "Conversationally relevant use" in Kronfeld's terms. 
380 
each subtree-, the root represents the "head" of the constituent and its 
immediate subtrees are the "modifiers". The syntactic production rules 
say how the tree must be extended when a word of a given syntactic 
category is analyzed (for more details see \[Lesmo & Torasso 83, 85a, 
85b\]). 
As soon as the syntactic rules extend the tree, the control is 
passed over to the semantic rules: they update the semantic represen- 
tation according to the meaning of the new piece of tree (in other 
words, syntax and semantics operate synchronously and the 
knowledge sources are separate). Backtracking is not used: whenever 
the analysis is blocked, the current situation is inspected by another 
set of rules ("restructuring rules") that modify the dependency tree in 
order to provide a new interpretation; then, the analysis is restarted on 
the new tree. Semantically ambiguous words that belong to the same 
syntactic category ate disanthiguated on the basis of the local context 
where they appear: no possibility of undoing the choice on the basis 
of a larger context is currently provided. The antecedents of the 
semantic rules pick up any newly entered word (the "current node of 
the tree") and match it, together with its surrounding context (i.e. its 
mother node and its dependants), against the domain knowledge, 
represented in terms of a KL-ONE-Iike semantic net, which fulfills 
the role of the terminological component in an hybrid knowledge 
representation system. The access to the net binds stone variables 
appearing in the antecedent of the rules to nodes of the net (con- 
cepts), which are then used in the consequent to build up the 
representation that is going to be described. Of course, the import of 
non-content words (as articles and quantifiers) is different from the 
one of content words. In a previous paper \[Di Eugenio & Lesmo 87\] 
we described how the introduction of non-atomic nodes called tri- 
concepts enabled us to deal deterministically with definite and 
indefinite determiners. Now, we will see how the problems about 
intensional contexts described before can be solved. In parallel, we 
worked on quantifiers: some results have been presented in \[Lesmo et 
al. 88\]. 
Tim semantic representation built by the interpreter consists in 
two planes (CP: Content Plane; RP: Reference Plane). A third plane 
is used in the drawings to refer to file stable dmnain knowledge (the 
terminological component mentioned above, called SP - Semantic 
Plane). The CP includes a set of nodes and arcs that represent the 
main assertion and the description of the entities involved in that 
assertion (i.e. the restrictions imposed on them by prepositional 
phrases, relative clauses, etc.). The RP contains a set of non-atomic 
structures called RASs (Reference Ambiguity Spaces) composed of a 
Concept node, an Extension node, and a Class node (three aspects of 
a "concepf' that are needed to establish the possible referent of a 
linguistic expression). These structures will not be used often in the 
present paper, since they are mote concerned with the representation 
of articles. A fourth node can be present in a RAS: it represents a 
subset of the extension and is used when a specific reading is chosen 
for definite noun phrases. On the contrary, since indefinite deter- 
miners are not used in class assertions, the Class node does not 
appear in RAS's associated with indefinite descriptions. Each RAS is 
linked to the defining concept of the SP via a SEMREF (SEMantic 
REFerence) arc. Also in the CP there may appear complex structures 
(DAS: Distributivity Ambiguity Spaces): They contain two nodes (Set 
and Individuals) and are used to distinguish between distributive and 
collective readings. The arc which connects a node in the CP to its 
associated RAS in the RP is labelled SAME for definite NPs (it sim- 
ply states that the given node of the CP coincides with the reference 
node in the RP), I-OF (Instance-OF) for indefinite ones. 
In order to simplify the figures, we will use the conventions 
shown in fig.1. Note that the initial representation (as shown in fig.l) 
will be augmented during the processing of the sentence if the NP has 
specifications (e.g. adjectives, relative clauses etc.). These 
specifications are not directly linked to the CP node that refers to the 
entity but to another node connected to the RAS by a DEF-AS (or 
NAME, if the specification is a proper noun) arc 9. 
9 The introduction of a pair of nodes for a unique referent is justified by the 
need of keeping apart the entity that partecipates in the main predication and the way 
it is defined. 
4. REPRESENTATION OF INTENSIONAI, CON'FEXTS 
The main extension to the work reported in \[Di Eugenio & 
Lesnro II'/l consists in the introduction of CONTEXT SPACES (CS), 
which enable us to treat the intensional contexts along the lines pro.. 
posed in \[Fauconnier 85\]. 
A Context Space is a subplane of the CP, which contains a set 
of nodes and of arc connecting them. CS are created when particular 
conditimm occur ill tile input sentence, in our case the presence of an 
intensional verb. In this case the CS is created and connected to the 
node which represents the verb in the CP by a SP-BUtLD (SPace- 
BUILDer) arc. The fact that the arc leaves a node representing an 
attitndin;d verb specifies (itnplicitly) the type of the CS lO and its gen- 
erator. The nodes appearing in a CS refer to entities that do not 
necessaril7 exist in the enter context (the CP itself is assumed to 
correspond to the otttennost context, i.e. it contains terms that have to 
be interptl;ted according to the speaker's view of reality, as expressed 
in the SP). A node of a certain CS may refer to a local entity ill that 
CS or to a counterpart in that CS of an entity of tile outer context. 
The COUNP (COUNmrPart) arc specifies the seeping of nodes in the 
semantic representation: an mv COUNP that connects a node NI in a 
CS (say CS1) m another node (say N2) external to CS1 states that N2 
is a countt.'tlxtrt of N1 (or, better, of the entity which NI refers to) in 
the outer t;pace; on the contrary, a COUNP arc staldng from N1 and 
entering N1 itself specilies that NI has no counterparts th outer con- 
texts, i.e. it refers to a local entity of space CS1. The absence of a 
COUNP ace leaving from NI represents art ambiguous case in which 
tile scope of N1 has not been determined yet. Such ambiguous 
representation enable us, as we will later see in the paper, to develop 
a determitfistic attalysis of sentences also in case of de-re or de-dicto 
ambiguity, 
Let's show how these struclnres are nsed by means of some of 
tile previotts examples. During the interpretalion of sentence (5) the 
analysis of the fl'agntent "Nadia wttttls" leads to the creation of a new 
CS (CS1 in fig.2). CS1 represents tbe "world of Nadia's desires as 
mentioned iu the sentence" and is linked to node N2 (instance of "to 
want") tlmmgh a SP-BUII.,D arc. Node N3 represents what Nadia 
wanls as an instantiation of the "dog" concept and is introduced in 
CS1 when "a dog" is analyzed. Up to this point, lhe representation 
must be considered as atnbiguous, sitlce it does not specify which is 
fire status of N3 wilh respect Io the speaker's world. In case no 
further processing is possible, then tile sentence remains ambiguous 
(fig.2 shows the basic ambiguous representation of sentence (5)); in 
other cases, perhaps depending on the context where tile sentence is 
uttered, it could be possible to decide whether tile intended reading is 
10 Ill Ibis pallor we will only deal with "attialdinar' CS's. Actually many oth- 
Er types of C~;'S may be used to cope with other kinds of linguistic pitettorttclta, For 
example a "temporal" CS may be used to deal with tile antbiguity of "In 1970, tile 
president met my sister", where "the preside~lt" could refer either te tile current president or to file person who was president in 1970. 
"de-m" or "de-dicto" (also preference criteria could be adopted; for 
instance, it seems that, outside any specific context, the preferred 
reading of (5) is the intensional one); when such a disambiguation is 
possible, then the representation is expanded by including a suitable 
COUNP arc: it will leave fi'om node N3 and will enter node the same 
node N3 iu case the reading is intensional, otherwise it will enter a 
newly-created node outside CSI (tile latter alternative is reported in 
fig.3). It is clear that it is the presence of node N8 that relates tlae 
representation to tire existence (in the speaker's world) of a specific 
entity wanted by Nadia, contrarily to what happens in the intensional 
reading, where no node (referring to a dog) existing outside CS1 is 
linked to N3. 
As stated above, two different "intensionar' readings can be 
obtained for sentence (5) depending on the relevance of tile descrip- 
tion. The representation we obtained till now corresponds to the 
purely referential l'eading of (5) (actually, this use of "l~eferential" is 
slighhy inappropriate, but the previous discussion should enable tile 
reader to intmpret it correctly). If we want to represent the "attrihit- 
rive" reading we tnust add something to the Stl'tJcture; what we need 
is a specification of the "dependency" of the statement made through 
the utterance on the description used. To this aim we include the 
WIllS (WH-ever IS) arc, as shown in fig.4. At this point, the import 
of this arc shotfld be clear. 
The proposed formalism enables the system to operate deter- 
ministically, i.e. to build the liual representation without destroying 
strttctures built previously (this was one of the goals of the introduc- 
tion of tri-concepts, as explained in \[Di Eugenio & Lesmo 87\]). In 
fact the system builds a basic ambiguous representation (as shown in 
lig.2 for sentence (5)) which is possibly expanded with the introduc- 
tion of new arcs when sotne disantbiguation is possible. Of course, 
the insertion of suitable arcs does not modify the set of nodes initially 
stored in the CS's, which remain stable (perhaps expanded) 
throughout the analysis. As regards the disantbiguation criteria, it 
does not seem that there are general roles for choosing one interpreta- 
tion or tile o/her. Of course, if the verb does not open an intensional 
context, then no problem arises (the CS is not created) Jr; On the con- 
traly, if we consider a verb as "seek", it sectns that only the object 
NP can provide the clues to select one particular reading; in "Mary is 
seeking a street", the preferred reading certainly is non-intensional; in 
"Mar5' is seeking a fl'iend" still seems to be preferable to assume the 
existence of a ptu'ticular fliend; on the contraly, in "Mary is seeking a 
ftmding" there seem 1o be few doubts thai the funding does not exist 
yet. We are currently trying to build heuristic roles, based on nominal 
hierarchies appearing in the terminological component of an hybrid 
knowledge representatiott formalisnt, able to suggest a particular read- 
ing. Moreover, subsequent sentences may provide further disantbi- 
guation criteria. Consider the example in \[Fauconnier 851, peg. 57: 
(9) Ursula wants to marry a ntillionaire. 
We nnty obtain different readings depending on the subsequent sen- 
tences: 
(t0) lie is really a eon-nlan 
leads to a de-re (referential) interpretation, while 
(11) But she won't thld one 
indicates a de-ditto reading tbr sentence (9). 
Concep~ T @P'~ '~ RAS ?ode 
Exte\[lslon ~" 
Node 
Possible ~ 
Sub.Extension 
DAS 
~~~'~'~, Individuals 
C~'~.,'4.. * Atomic Node Node 
...... (llotl-nnlbiguous) 
C# 
Fig,1 - Basic structures used in the representation 
A linal point concerns the use of the proposed representation 
in the subseqttent phase of irtterpretation. Generally, the search for tile 
correct referent in the discourse history is led by the presence of a 
RAS in the RP and by its incoming (I-OF or SAME) and outcoming 
(a DEF-AS arc which connects it with the representation of the 
referent's description) arcs. An indefinite specific reference (for 
exatnple "a dog" in "a dog is eating meat") causes the creation of a 
new entity, since the assumption is that indefiuites are used when the 
referent is not identitiable by the reader. On the contrary definite 
specilic reference (for example "the dog" in "the dog is eating meat") 
leads to the search.for the entity referred to into the discourse his- 
tory l:~ 
n Actually, the verb of the sentence is exatnined in all cases, since tile pres- 
ence of some auxiliaries folx;es the creation of an intensional context; consider, for 
iastaoce: "You tllusl reserve a seat", "I will build a lable". Of course, these verbs, 
when used in the past tense do not have any intensional characterization. 
l:~ Obviously, many other types of references are possible• For a complete 
analysis of referential problem connected with tile semantic of determiners see \[Di 
Eagenio & l~esmo 87\], 
381 
Let us see now the impact of the newly proposed extensions 
onto this environment. When a RAS is referred to by a node Nk 
through an 1-OF arc (indefinite specific reference) the possible pres- 
ence of Nk in a CS (say CS1) may be detected. If this is the case, 
node Nk must be inspected: if it has a COUNP arc, pointing to a 
node of the external context, then the usual creation process is exe- 
cuted; otherwise Nk refers to an entity which is local to CS 1 and has 
no counterparts in the external worid. So we need an interpretation 
which blocks any infet~ence on the entity associated to the node in the 
external world. This result can be obtained by copying the whole CS 
in the Iinal network representing the discourse history, and so biding 
its internal node to the external context. 
Let us specify now the interpretation of the WHIS arc. As 
stated above, when a RAS is refetTed to by a node Nk through a 
SAME arc (definite specific reference), an entity satisfying the 
description (represented by a Nj node of the CP connected to the 
RAS through a DEF+AS arc) is searched for in the discourse history 
network. In fact, in this case, we only have to cope with "the func- 
tional relevancy of the description" \[Kronfeld 86\] (i.e. its ability to 
convey infolxnation leading to the identification of the correct 
referent). But what does it happen if we also have to account for 
conversational relevancy, i.e., what is the interpretation if a WHIS arc 
connecting the description node Nj to the node Nk exists? In this 
case the usual searching for the referent is delayed. Both node Nk 
and its description (represented by the WHIS ate and by the subnet of 
CP whose head is Nj) are entered in the discourse history network. 
In this way the description itself becomes part of the final representa- 
tion of the sentence (as stated in \[Barwise & Perry 83\]). Moreover 
the presence of an arc WHIS in the final network states that the node 
entered must not be considered as the referent of the sentence hut 
rather as a "role" (see \[Fauconnier 85\]) which may assume different 
values in different worlds (i.e. in different CS's). In fact, different 
referents for the descriptiort may be detected by the usual searching 
process in the CS's and this leads to different inferences (different 
assignments of referents to the "role"). 
NAME 
8~ 
7 
~P 
r 
ROSS -'qEJ 
Fig.2 -Basic ambiguous representation of the sentence "Nadia wants a dog like Ross's". 
SP-BUII_D ~X ~N~ 
j ~ .... L,KE j 
(.b 
Fig.3 -Representation in the CP of the 
transparent reading of the sentence 
"Nadir wants a dog like Ross's" 
~lP N~ T 
8p- 
._ k .... ----,,,£m 
WHIS N4 
@ 
Fig.4 -Representation in the CP of 
the opaque attributive reading of the 
sentence "Nadia wants a dog like Ross's. 
382 
A final remark is needed: for the sake of simplicity we stated 
that the relevancy of tile description must be detected during the 
analysis of the input sentence, independently from the subsequent 
interpretation process. Actually the strategy of the system is quite 
more flexible, because also the interpretation process, merging tile 
semantic representation of the sentence and the discourse history net- 
work, supplies further information about the relevancy of a given 
description. For example, when the sentence does not include 
sufficient information to choose between referential and attributive 
reading), the absence in the discourse history of a referent litting the 
description (detected during tile final process of interpretation) forces 
an attributive reading. 
5. CONCLUSIONS 
We have described a semantic network formalism that is able 
to cope wifll the problems arising when NP's occur as objects of 
verbs inducing an intensional context. The mp,esentation preserves in 
an easily a,-cessible format all information needed to distinguish 
among the lU)ssible readings, and the construction of tile net can be 
carried on ({eterministically (see the discussion of paragraph 3 and 
note that di:~ambiguating among different readings only consists in 
adding arcs to existing subnets), llowever, some problems still have 
to be solved concerning the final representation obtained by the pro- 
cess of mew, ing the meaning of the sentence under analysis with the 
discourse hi~tory. Actually, the "assertional" cmnponent of hybrid 
lortnalisms (see, for-instance, \[Brachman et al. 83\]) does rtot lend 
itself to handling intensional objects created duting.~he analysis. As 
stated above, we have structured the assel,ional component providing 
facilities for copying the subnetworks containing the description of 
intensional entities into it. On tire other hand, some work still has to 
be made in {,rdcr to provide such a component with a clear semantics. 
We believe that basic insights can be found in the SNePS system 
\[Maida & Shapiro 821, \[Shapiro & Rapaport 86\], where the proposed 
organization meet,s more clearly the requirements set forth by the 
existence o1' intensional olrjects. 
The version of the interpreter described in this paper is imple- 
mented in FRANZ LISP and runs on a SUN workstation under the 
UNIX operating system. The maintenance of tim discourse history 
and the merging process is currently under implementation. 
ACKNOWLEDGMENTS 
This work has partially been fimded by tile Ministem della 
l'nbblica Istmzione of Italy, within the project "Software Architec- 
tures for lntdligent Systems". We must thank Kai von Luck and his 
colleagues of the Technische Universitaet Berlin, for having put at 
our disposal BACK, a hybrid Knowledge representation system 
whose terminological component has been used to implement the 
Semantic Plane. The second author (Paolo Terenziani) acknowledges 
Regione Piemonte, Progetto Finalizzato Sanita', that provided the 
funds for his Research Assistantship. 

REFERENCES 
\[Barwise & Peny 83\] l.Barwise, J.Perry: Situations and Attitudes. 
MIT Press (1983). 

\[Brachman et al. 83\] R.J.Brachman, R.E.Fikes, H.J.Levesque: KRYP- 
TON: Integrating Terminology and Assertion. Proc. AAAI-83, 
Washington, D.C. (1983) 31-35. 

\[Cooper 83\] R.Cooper: Quantification and Syntactic Theory. 
D.Reidel Publ. Co. (1983). 

\[Di Eugenio & Lesmo 87\] B.Di Engenio, L.Lesmo: Representation 
and Interpretation of Determiners in Natural Language. Proc. 
lOth IJCAI, Milano (1987) 648-654. 

\[Dowty et al. 81\] D.R.Dowty, R.E.WalI, S.Peters: Introduction to 
Montague Semantics. D.Reidel Publ. Co. (1981). 

\[Fauconnier85\]G.Faueonnier:MentalSpaees. MITPress (1985). 

\[Fawcett & Hirst 86\] B.Fawcett, G.Hirst: The Detection and 
Representation of Ambiguities in kntension and l)escription. 
Proc. 24th Annual Meeting of the AcL; New York (1986) 192-199. 

\[Kro,ffeld 86\] A.Kronfeld: Donnellan's Distinction and a Colnpnta- 
tional Model of Reference. Proc. 24th Annual Meeting of the 
ACL, New York (1986) 186-191. 

\[Lesmo & Torasso 83\] L.Lesmo, P.Torasso: A Flexible Natural 
Language Parser Based on a Two-level Representation of Syn- 
tax:. Proe. 1st Meeting of European Chapter of ACL, Pisa 
(1983) 114-121. 

\[Lesmo & Torasso 85a\] L.Lesmo, P,Torasso: Amdysis of Conjunc- 
tions in a Rule-based Parser. Proe. 23rd Annual Meeting of 
ACL, Chicago (1985) 180-187. 

\[Lesmo & Torasso 85b\] L.Lesmo, P.Torasso: Weighted Interaction of 
Syntax and Semantics in Natural Language Analysis. Proe. 9th 
IJCAI, Los Angeles (1985) 772-778. 

\[Lesmo et al. 88\] L.Lesmo, M.Berfi, P.Terenziani: A Netwm'k For- 
realism for Representing Natural Language Quantifiers. 1'roe. 
ECAI 88, Muenchen (August 1988). 

\[Lewis 72\] D.Lewis: General Semantics. In D.Davidsnn, GHarman: 
Setnantics of Natural Language. D.Reidel Publ. Co. (1972) 
169-218. 

\[Maida & Shapiro 82\] A.S.Maida, S.C.Shapiro: \[ntensional Concepts 
in Propositional Semantic Networks. Cognitive Science 6 (1982) 
291-330. 

\[Shapiro & Rapaport 86\] S.C.Shapiro, W.J.Rapaport: SNePS Con- 
sidered as a Fully Intensional Propositional Semantic Network. 
Proc. AAAI-86, Philadelphia (1986) 278-283. 
