l 
/ 
I 
l 
/ 
/ 
/ 
/ 
/ 
| 
/ 
I 
Sense Variation and Lexical Semantics Generative Operations 
Patrick Saint-Dizier 
IRIT-CNRS, 118, route de Narbonne 
31062 Toulouse Cedex France 
stdizier@irit, fr 
Abstract 
In this document, we outline some elements 
related to sense variation and to sense delim- 
itation within the perspective of the Gener- 
ative Lexicon. We then show that, in some 
cases, the Qualia structure can be combined 
with or replaced by a small number of rules, 
which seem to capture more adequately the 
relationships between the predicator and one 
of its arguments. 
1 Introduction 
Investigations within the generative perspective aim 
at modelling, by means of a small number of rules, 
principles and constraints, linguistic phenomena (ei- 
ther morphological, syntactic or semantic) at a high 
level of abstraction, level which seems to be appro- 
priate for research on multi-linguism and language 
learning. These works, among other things, attempt 
at modelling a certain form of 'creativity' in lan- 
guage: from a limited number of linguistic resources, 
a potentially infinite set of surface forms can be gen- 
erated. 
Among works within the generative perspective, 
let us concentrate on the Generative Lexicon (Puste- 
jovsky 91, 95), which has Settled in the past years 
one of the most innovative perspective in lexical 
semantics. This approach introduces an abstract 
model radically opposed to 'flat' sense enumeration 
lexicons. This approach, which is now well-known, 
is based (1) on the close cooperation of three lexi- 
ca\] semantic structures: the argument structure (in- 
cluding selectional restrictions), the aspectual struc- 
ture and the Qualia structure, (2) on a detailed type 
theory and a type coercion procedure and (3) on a 
refined theory of compositionality. The Generative 
Lexicon (GL) investigates the problem of the mul- 
tiplicity of usages of a sense of a lexeme and shows 
how these usages can be analyzed in terms of possi- 
ble type shirtings w.r.t, the type expected by a usage 
of that sense defined as the core usage. Type shift- 
ing is modelled by a specific inference mechanism: 
type coercion. The GL shows very clearly the inter- 
dependence between arguments and predicates. 
In our perspective, we are not only interested in 
deciding whether an expression is an acceptable ar- 
gument for a predicate and for what reasons as it 
is also the case in the GL, but we want to be able 
to 'reconstruct' or infer the meaning of the proposi- 
tion from its parts (the predicate and its arguments), 
and possibly also from the implicit semantics con- 
veyed by the syntactic form (Goldberg 94). We as- 
sume that the impossibility of building a semantic 
representation for a proposition entails that it is se- 
mantically ill-formed w.r.t, our grammar, lexicon 
and composition rules. The work presented here is 
not definitive, it is a feasability study which aims 
at opening new perspectives and new treatments for 
the GL, while remaining in the same spirit. We will 
in particular show that in some cases, the Qualia 
structure can be combined with, or replaced by, a 
set of relatively general rules, which, in addition, 
better capture the relations between the predicator 
and the predicated elements. 
In this paper, we contrast a rule-based approach 
(also used by other authors such as (Copestake and 
Briscoe 95), (Ostler and Atkins 92), (Numberg and 
Zaenen 79) with different perspectives) with the 
Qualia-based approach and comment on their re- 
spective advantages. We show how, in fact, they can 
cooperate. Another view is that presented in (Jack- 
endoff 97, chapter 2) with his principle of enriched 
composition, which is in fact quite close to our view, 
but restricted to a few coercion situations (aspec- 
tual, mass-count, picture, begin-enjoy). As will be 
seen, these systems are not incompatible, they cover 
different forms of knowledge and may be useful in 
different situations. The rules we present here are 
not lexical rules, as in (Copestake and Biscoe 95), 
but they are part of the semantic composition sys- 
Saint-Dizier 121 Sense Variation and Lexical Semantics 
Patrick Saint-Dizier (1998) Sense Variation and Lexicai Semantics Generative Operations. In D.M.W. Powers (ed.) 
NeMLaP3/CoNLL98: New Methods in Language Processing and Computational Natural Language Learning, ACL, pp 121-130. 
tern. Compared to the Qualia system (Pustejovsky 
91, 95), the different senses which can be derived 
from a more central sense are constructed in our sys- 
tem in context i.e. when dealing with arguments, not 
a priori as in the Qualia roles. Finally, in the last sec- 
tion of this document, we show that some problems 
advocated by (Copestake and Briscoe 95) concern- 
ing unification can be resolved by a constraint-based 
approach. 
Before going on into more details about our ap- 
proach, here are some general comments about lex- 
ical semantics generative systems. 
In a lexical semantics generative system, it is of 
much importance to have a clear analysis of the no- 
tions of word-sense and of sense delimitation. In- 
deed, depending on the strategy adopted (e.g. nar- 
row senses as in WordNet, or very large ones as in 
many AI works), the nature and the scope of gen- 
erative operations, and therefore the philosophy of 
language behind these operations, may entail very 
different sense delimitation strategies. We do not 
think that it is possible to define a real theory of 
sense delimitation, but it is certainly possible to de- 
fine a few principles or a strategy. Then, given a 
strategy for sense delimitation, we can adjust the 
complexity and the scope of generative operations. 
We assume that a verb potentially contains a num- 
ber of subsenses which are 'triggered' by semantic 
composition rules depending on the arguments found 
in a sentence. There are several possibilities to rep- 
resent a sense and its relations with subsenses, our 
approach is, as shall be seen below, to develop under- 
specified Lexical Conceptual Structure (LCS) (Jack- 
endoff 91) representations, that will also be used as 
a basis for developping generative mechanisms, It 
should be noted that, apart from fixed forms, sub- 
senses share quite a lot of features and are often 
distinguished by a few features. 
As shown in (Pustejovsky 95) and (Copestake and 
Briscoe 95), type coercion does not occur within the 
verb, but at the level of the proposition, allowing a 
sentence such as: 
Mary enjoys the film and eating ice-cream. 
This type of sentence is problematic for unification 
because the verb must bear two types for the first 
object, syntactically an NP and a proposition (S- 
bar) and semantically a physical object and an event. 
A constraint-based approach which allows domains 
to be assigned to variables (the variable representing 
the semantics/syntax of the object) would be a good 
solution (see section 6). 
In spite of evidence of its existence (explored in 
the Euronet project, the European WordNet), the 
Qualia structure turns out to be quite difficult to 
describe. It is well-designed for nouns, and looks 
more artificial for other lexical categories, but it 
is in fact essentially useful for predicated elements, 
mostly nouns. In fact, it seems that Qualia struc- 
tures are not really motivated and relevant for pred- 
icators. This can be noted in a concrete way when 
looking at verb Qualias which do not turn out to be 
very satisfactory and expressive, but rather look a 
little bit artificial. The descriptions made in roles are 
also often more pragmatic than semantic, and this 
may entail additional confusions in resolving type 
mismatches. This is in fact clearly stated in various 
analysis of the Aristotelician AITIAs, from which 
the Qualia roles emerged (M0ravcsik 75), where the 
author shows that there is a kind of continuum be- 
tween telic and agentive roles. Finally, the formal 
and constitutive roles are also somewhat redundant 
with the part-of and isa relations used to structure 
lexicons (or similar relations used in thesauri), it 
may not therefore be totally necessary to include 
them in the Qualia. 
A close observation of linguistic data and of sense 
variations shows that the most important role is the 
telic role. It is in fact the role to be considered as 
a default role for the coercion object -+ event. The 
formal role is far less frequently considered. We also 
think that this role should not only describe the com- 
ing into being of entities but also their destruction, 
which cannot enter into the telic role. 
We do not thing that generativity can only be 
based on an elaborated theory of types, and on the 
operation of type coercion. This operation is very 
powerful, simple and works satisfactorily for a num- 
ber of cases, but it turns out in some cases to be 
insufficient and also to largely overgenerate (in par- 
ticular it is quite d i~cult to use in natural language 
generation systems), unless a very rich and refined 
theory of type is defined, which may be contrary to 
the type philosophy. We think that the type system 
is well-designed to detect derived usages, characteri- 
zable by a type mismatch. The generality and gener- 
icity of the type coercion operation should however 
be kept because, via the reference to quite general 
types, it preserves a good degree of generality and 
systematicity to type shiftings. However, we feel 
that types are not sufficiently 'constrained' to ac- 
count for the constraints holding, for each verb, on 
the different sense/usage variations it may be sub- 
ject to. For example, an underspecified LCS form 
could be more appropriate because of its ability to 
partially represent meaning. The Qualia structure 
cannot encode all these constraints, in particular 
first because they are better designed for arguments 
than for predicators and second because some con- 
Saint-Dizier 122 Sense Variation and Lexical Semantics 
II 
II 
II 
II 
II 
!1 
II 
II 
II 
E 
| 
| 
I 
m 
m 
m 
m 
m 
I 
/ 
mm 
straints are relational between the predicator and 
the argument, which the roles cannot describe. 
In this document, we first present a standard 
classification of the different forms of relations be- 
tween a predicate and its arguments, then we outline 
some considerations on the difficult problem of sense 
charaterization and delimitation. We then show how 
it is possible to give up the principle of a Qualia 
structure and, in some cases, of type coercion in 
favor of (1) other sources of knowledge, essentially 
inherent properties of lexical items and (2) simple 
inference rules. This approach may be felt to some- 
what resemble meaning postulates, but it is orga- 
nized on a larger scale and treats different phenom- 
ena. 
2 Usage relations in the 
predicate-argument pair 
In this section, we consider the different types of 
usage relations between an argument and a predi- 
cate as introduced in the GL, and add some distinc- 
tions. By usage relation, we mean the analysis of 
the concordance/matching relation between a pred- 
icative term and one of its arguments in a language 
utterance. These relations are well-known and are 
discussed within our framework in (Mari 97) and 
(Marl and Saint-Dizier 97). 
The relational perspective adopted here is math- 
ematically richer than a functional one, it allows us 
to introduce ambiguity and non-determlnism. It also 
gives an equal role to the predicate and to the argu- 
ment (this is also the case in the GL). 
The typology of usage will then serve as a basis for 
an analysis of sense delimitation and for the devel- 
opement of generative operations. We consider here " 
the following forms of usage relations, which are, for 
most of them, very classical: 
• Selection: the type of the argument is sub- 
sumed by the type (possibly polymorphic or 
'dotted') expected by the predicate for that ar- 
gument. In general, the sense conveyed by the 
verb, which may be largely underspecified, is 
made more precise by the type of the argument 
by means of more or less complex compositional 
rules. For example, for the verb construire (con- 
struct), the expected object is, for example, of 
type 'physical-object v abstract-construction'. 
Selection is also called selective binding in the 
GL. Selection is clearly a wider phenomenon for 
general purpose (generic) verbs than for verbs 
with a more restricted meaning. Within our 
approach, instead of just using types, the dif- 
ferent subsenses of a verb are represented by 
Saint-Dizier 123 
means of underspecifled LCS forms. The pos- 
sible/potential instanciations are described by 
means of a system of constraints and LCS forms: 
Constraint : LCS form used to fill in under- 
specied positions. 
This is exemplified in sections 4 and 5. 
• Metonymy: in an argument, an entity is 
used to refer to another one, via lexical rela- 
tions, among which mainly the part-whole re- 
lation (and its different forms (Winston et al. 
87)), but also forms such as producer-product 
or institution-person relations. A number of 
metonymies have a high degree of systematic- 
ity in language (Lakoff 80), possibly over differ- 
ent languages. There are well-known examples 
of metonymies such as buses are on strike to- 
day (object for user or controller) or I bought a 
Ford (make for object). Most of these forms are 
treated by the GL via the constitutive or formal 
roles, but in a rather unconstrained way (except 
for the constraints implicit in the Qualia). 
We can somewhat structure metonymies by con- 
sidering 3 major forms: 
- metonymies based on the part-of relation, 
- metonymies based on various forms of ar- 
gument type shirtings such as the make for 
object illustrated above (I bought a Ford/a 
Sun). There is always a semantic link such 
as possession between the expected and the 
realized arguments. Another link is, e.g. 
physical object -+ monetary value. An in- 
teresting case is a sentence such as I bought 
French to express that I bought some prod- 
uct made in France or whose makeis known 
to be French. 
- metonymies introducing an emphasis on 
properties of the argument related to ac- 
tions and to its uses. From the argument's 
properties, argument type shiftings (also 
called alternations) can be defined such as 
the alternation physical object ~ event. 
The famous ambiguous example of the GL: 
'begin a book' is treated by this type of 
metonymy, using the general rule that any 
physical object was once created and can 
be used in a number of ways. 
Metonymies will not be developed in this paper. 
• Metaphor: this phenomenon is more com- 
plex and diverse than metonymy. It is however 
possible to isolate several types of metaphors 
which are relatively concrete, and which also 
Sense Variation and Lexical Semantics 
,i 
have a good degree of systematicity. These 
metaphors refer usually to partial homomor- 
phisms between ontologies of different concep- 
tual domains, with a focus on some aspects (or 
seines) carried prominently by the argument. 
General schemas have been defined, e.g. in 
(Lakoff 80), such as: Ideas are objects, ideas are 
food, creation is birth, activity is a substance, 
etc. which can, roughly speaking and within 
the GL perspective, allow a type to be changed 
into another. Metaphors are not, in general, 
treated by the GL, at least in a simple way. 
Rules for metaphors axe based on type shift- 
ings; their scope must undoubtly be restricted 
to some groups of nouns. The semantic inter- 
pretation of a metaphorical expression is in gen- 
eral not a trivial matter and much remains to 
be done in this area. 
. Co-composition deals with the emergence of 
unexpected new senses, often with an impor- 
tant sense variation, from the combination of an 
argument with a predicate. This phenomenon 
usually goes somewhat beyond compositional- 
ity and may be at the origin of semi-fixed 
or fixed forms. Compared to the definition 
given in (Pustejovsky 91), we have a more re- 
stricted view of co-composition. In our view, 
co-composition occurs when there is an impor- 
tant change in meaning, not simply when an ar- 
gument has influence over the verb's semantics. 
We do not think, in fact, that an argument may 
modify the meaning of a verb, it can just spe- 
cialize it, but this specialization remains com- 
positional. Situations of co-composition occur 
when the meaning of a proposition is not just a 
certain composition of the meaning of its parts, 
but something different. 
Examples abound of this phenomena, such as: 
produire des tdmoins, se payer un zdro, gagner 
sa vie. Co-composition is particularly difficult 
to analyze. Let us note two points. First, in 
a number of co-composition relations, a few, 
usually secondary, features are strongly high- 
lighted, producing the feeling of 'unexpected' 
use. Secondly, these forms are not formed at 
random and do not acquire a meaning also at 
random. Some of them may be the remnants of 
productive forms, perfectly compositional, sev- 
eral centuries ago. Fixed forms may then be 
considered as an impoverishment of previously 
compositional forms (see (Marl 97) for a few 
notes). 
Althought we haven't studied very far co- 
composition, we believe that a simple way to 
handle a number of them is to establish a link 
with one of its paraphrases, whenever possible, 
which involves a more regular form of compo- 
sition. For example, se payer un zdro will be 
associated with a verb (in English) such as to 
get and more precisely to get a zero. Similarly, 
se composer un visage = to compose one's face 
(section 5.1) will be associated with the expres- 
sion to hide one's feelings or opinions. 
These types of variations are claimed (e.g. in 
(Lakoff 80)) to have a quite good degree of system- 
aticity among verbs. It is important for our model 
to evaluate how stable they are within and over verb 
semantic classes in order to attain a certain general- 
ity and reusability. 
3 The problem of sense delimitation 
3.1 Preliminaries 
Sense delimitation is a very delicate, but crucial, 
problem. It has been studied at length by many 
authors and in many projects. 
Let us say very informally that, in our perspec- 
tive, we assume that a sense (more or less large and 
constrained) of a lexeme has a basic form and basic 
expressions called usage(s) (a surface form reflecting 
the basic sense). For example, let us assume that 
for the verb devour, its basic sense has a human or 
animal as subject and an edible object as its object. 
Then John devours meat is a basic usage. The basic 
sense expects some types of arguments but it may 
also originate derived usages (via the above genera- 
tive operations), which axe more or less constrained 
and limited. For the above example, we have John 
devours books, which is metaphorical. Since we can 
only observe linguistic realizations of these derived 
usages, sense delimitation is then defining sets of us- 
ages sharing the same semantic 'root', and identify- 
ing that root. Meaning is then assigned to linguistic 
expressions from the meaning of their constituents 
and by taking into account the semantic shifts intro- 
duced by the generative operations involved. 
This perspective may be felt to be somewhat naive 
and archaic: (1) examples abound where it is not 
possible to identify a root among usages, and (2) us- 
ages are not necessarily derived from a more or less 
unique root but may result from complex interac- 
tions among derived usages. Nevertheless, we can 
say that: 
• our perspective is workable for utterances from 
relatively standard texts. It is of course al- 
ways possible to find counter-examples, but 
Saint-Dizier 124 Sense Variation and Lexical Semantics 
!1 
I 
| 
I 
I 
I 
I 
I 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
!1 
II 
II 
II 
II 
II 
then their rate of occurence should be evalu- 
ated. 
We have at our disposal formal and practical 
tools to 'soften' the notion of Semantic root, e.g. 
polymorphic types, or complex systems for se- 
mantic representations. 
• Our approach is incremental, allowing the ad- 
junction, step by step, of new usages. 
Semi-fixed forms and fixed forms are treated 
apart, and we believe that a number of counter- 
examples to our approach could fall in this class. 
Sense delimitation is largely an open problem. It 
is indeed almost impossible to state precise and gen- 
eral principles that characterize the boundaries of 
different senses of a lexeme and what a sense exactly 
is. To make our approach workable, let us assume 
that senses exist independently of linguistic realiza- 
tions, whatever they are called (e.g. concepts), and 
that it is possible to formally represent them. Sense 
delimitation largely depends on the theoretical and 
practical perspectives and, to a large extent, on intu- 
ition. The diffÉculty is then to elaborate a coherent 
system of sense delimitation and of sense/usage vari- 
ation. Extreme solutions have been proposed, which 
are not totally satisfactory. For example WordNet 
(FeUbanm 93) tends to introduce one sense per us- 
age, where a usage is characterized by the semantic 
features of the arguments. For example, WordNet 
has 27 different senses for the verb give. Distinc- 
tions between senses are often very subtle and some- 
what hard to represent in a formal semantic repre- 
sentation. This approach is very useful in the sense 
that it provides a very detailed description of the us- 
ages of a large number of words in English, but we 
think it lacks generalizations about language which 
are often useful for NLP systems to work efficiently. 
On the other extreme, there are AI systems which 
tend to postulate a unique sense for a lexeme and 
very complex derivation procedures, involving com- 
plex logical systems, to produce different sub-senses. 
This may be explained by the fact that most M ap- 
proaches are not concerned with usages in language 
but just by meaning representation. 
Our perspective is in-between these extremes. We 
think that the different usages of a word should be 
organized around a small, relatively generic, num- 
ber of senses. From these senses, similarly to the 
GL, generative procedures should produce or recog- 
nize derived usages, with possible slight changes in 
meaning. Sense delimitation is addressed more for- 
really in (Saint-Dizier 98). 
A second problem is the definition of the nature 
and the form of constraints related to sense delimi- 
tation. For example, there are meaning components 
within a given sense which can be more or less con- 
strained or opened. Within a given sense, modali- 
ties, which are often more peripheral, such as: man- 
ner, means, goals, existence of an effect or telicity 
may be left unconstrained whereas elements such as 
the type of agent, the existence of a cause, or the 
type of the first object (e.g. physical object or not, as 
in vendre un objet = to sell versus vendre quelqu'un 
= to denounce) may be much more constrained. 
Then, given a set of usages assumed to be related 
to a unique sense, on what basis should the basic 
usage(s) and, therefore, sense be defined ? We will 
show that it is often the most usual usage which is 
the most primitive, probably the most concrete one, 
often one of the most widely used, and also possibly 
historically and ontogenetically the oldest (Marl 97). 
The semantic properties and representation of the 
basic usage must then be able to allow and possibly 
to explain the derived usages. 
3.2 Semantic representations of senses: a 
psycho-linguistics perspective 
These questions inevitably raise the problem of the 
semantic representation. Even if it is not compre- 
hensive, it turns out, from our experiments on dif- 
ferent semantic classes of verbs, that the LCS (Lex- 
ical Conceptual Structure (Jackendoff 90)) is a rela- 
tively adequate framework (possibly associated with 
a few attribute-value pairs for some properties) to 
represent the semantics of expressions subject to the 
sense variations we have identified, and to allow for 
the implementation of the generative operations ad- 
vocated in the previous section (see also B. Dorr's 
work on LCS forms for verb classes in English). The 
different, hierarchically organized, elements at stake 
in the LCS seem to correspond, in general, to the 
variation factors that we have identified and to be 
sufficiently fine-grained. 
It also turns out that, in the LCS, the primitive 
constituents and some basic formulae correspond to 
different steps of the categorization procedure when 
learning a language (Dubois, Marl and Saint-Dizier 
97). It will then be possible to modify or replace 
autonomous and often elementary parts of the LCS 
by others when representing sense variations, in a 
way similar to the human categorization and learn- 
ing procedures which modify the structure of con- 
cepts. Notice that operations of composition within 
the LCS are simply the embedding and the concate- 
nation of LCS formula within others. 
On a psycho-linguistic level, (Dubois et al. 97) 
Saint-Dizier 125 Sense Variation and Lexical Semantics 
has shown, for the possession verb family, elements 
of meaning which organize this family around pro- 
totypical kernels. For example, verbs such as ra- 
cheter and facturer aggregates very early in lan- 
guage developement stages to the kernel verb acheter 
(buy). Then, verbs such as indemniser, payer, rem- 
bourser, rdmundrer, dddommager come later. With 
each of these levels can be associated well-formed 
fragments of LCS, which can be combined with the 
kernel's representation and which make the mean- 
ing of these verbs more precise and distinct from 
the kernel (Dubois, Saint-Dizier 96). 
The generative operations we present operate on 
types and on LCS fragments, that they modify via 
substitution or concatenation to take into account 
sense variations. To allow for flexible substitutions, 
some LCS fragments will be represented apart, e.g. 
as default representations or as underspecified rep- 
resentations. Generativity, in a certain sense, is di- 
rectly related to the evolution of the language ac- 
tivity. This statement looks trivial, but it has some 
immediate consequences on the type of structural el- 
ement necessary to implement generative operations. 
4 Adjective Modification and 
Selection 
This section is devoted adjectival modification. The 
goal is to study the use and impact of the Qualia 
structure of the modified noun in the determina- 
tion of the semantic representation of the association 
Noun + Adjective. Similarly to verbs, we show that 
even for highly polysemic adjectives, there is a cen- 
tral sense, or a small set of closely related senses, for 
the adjective which is altered or specialized by .the 
modified noun. This is typically a selection (or se- 
lective binding) situation. The problems we address 
are the identification of what is exactly modified in 
the noun and how, by what means is it determined 
and what is the ensuing semantic representation. To 
illustrate this study, we have chosen one of the most 
polysemic French adjectives: bon (good). Other ad- 
jectives studied in the GL litterature are e.g. fast, 
sad, afraid and noisy. 
Bon in French has a major global meaning, namely 
'that works well', as in une bonne machine, un bon 
outil (a good tool, machine). It is also an adjec- 
tive of intensity, in quantity or quality, as in un bon 
repas (a good meal). These two meanings are very 
close as shall be seen below. Bon may also be com- 
bined with determiners expressing measures to indi- 
cate that the measure is slightly excedeed as in un 
bon litre (a good liter), and it is at the origin of a few 
fixed forms as un bon coup, une bonne gifle (a good 
slap), which are, in fact, synonymous of un mau- 
vais coup, une mauvaise gifle (a bad slap) even if 
bon and mauvais are opposites. This shows the non- 
compositional character of these fixed forms. Other 
fixed or semi-fixed forms are: bon suns, bon vivant. 
Let us now study in detail each sense. In general 
examples will be given in English to facilitate read- 
ing as long as they exactly correspond to the French 
ones. 
4.1 Bon = that works well 
This first sense applies to any noun of type tool, ma- 
chine or technique (in a general sense): a good car, 
a good screw-driver, a good computer, a good algo- 
rithm, a good cure, a good medecine, etc. It also 
applies to more abstract nouns denoting communi- 
cation acts or cognitive procedures, such as: a good 
advice, a good argument, a good talk, a good demon- 
stration. Exactly the same phenomenon occurs for 
its opposite mauvais (bad). 
Let us now consider the construction of the se- 
mantic representation. In (Pustejovsky 95), a role 
in the telic is considered, and ~an adjective such as 
fast, modifying a noun such as typist is represented 
as follows: 
he \[type'(e,z) ^ fast(e)\] 
where e denotes and event. This formula says that 
the event of typing is fast (fast may later be inter- 
preted in more depth). A similar representation is 
given for long, in a long record. This approach is fine 
to represent temporal notions in a global way, i.e. 
the event is said to be fast (i.e. short) or long. But 
this is not so straightforward for an adjective such 
as bon, and many other adjectives with no temporal 
dimension. Consider the representation of a good 
typist: 
)~e \[type'(e,z) A good(e)\] 
where it is not the typing event which is good but 
some more refined properties. Also, we do not want 
to consider a precise event, but to state that in gen- 
eral someone is a good typist. Such a statement 
implies at least a quantification over typing events 
of x. Finally, bon being polysemous, a single repre- 
sentation shouldn't be sufficient to accomodate all 
the senses. 
We can define an abstract representation for bon 
which includes a reference to the telic role of the 
Qualia structure of the noun. Let us assume that 
any noun which can be modified by bon has a telic 
role in which the main ftmction(s) of the object is 
described (e.g. execute programmes for a computer, 
run for a car): 
noun: N, Qualia: \[ Telic: T, ...\] 
where T denotes the set of properties associated with 
Saint-Dizier 126 Sense Variation and Lexical Semantics 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
m 
m 
m 
m 
m 
m 
m 
m 
B 
m 
B 
m 
m 
m 
m 
B 
the telic role of the noun N. Let us assume that T is 
a sequence of predicates of the form Fi(X, Y) where 
Y denotes the noun N. Let us assume that F~(X, Y) 
is the property modified by the adjective bon, identi- 
fied by means of semantic types. Then a LCS-based 
representation of bon is: 
\[stato BE+oho~,+id~.t(\[thi.g ¥ \], 
\[+p~op ABILITY - TO(F,(X, Y)) = good \])\] . 
This representation expresses the fact that the en- 
tity denoted by the noun works well via the evalu- 
ation function ABILITY-TO and the value 'good'. 
sFrom a certain point of view, this representation is 
close to the idea of (high) intensity, but specialized 
around the basic functionalities of an object, not its 
qualities, as emphasized in section 4.3. It is impor- 
tant to note here the role of the Qualia structure: to 
go beyond a general semantic representation and to 
introduce in a direct way a pragmatic or inter- 
pretative dimension via the instanciation of the 
predicate Fi (X, Y). 
The constant 'good' can be replaced by a more ac- 
curate representation, e.g. 'above average', but the 
problem of effectively measuring the activity remains 
open (in the pragmatics). Notice also that instead of 
quantifying over events, bon is described as a state, 
independently of events associated with the use of 
the object. The functionalities of the object remain 
good, even when it is not used effectively. 
If several properties are at stake, we may have a 
conjunction of functions: 
\[state BE+char,+ident(\[thin9 r \], 
\[+p,op A,=,,, ABILITY- TO(F~(X, Y)) = good \])\]. 
~From a compositional point of view, the combi- 
nation Adjective + noun is treated as follows, where 
R denotes the semantic representation of the adjec-. 
tive, T the contents of the telic role of the Qualia of 
the noun, and Y is the variable associated with the 
noun: 
f(Adj (R),Noun(Qualia(T)) = 
AX, Y BT(X,Y) 6 T, (N(Y) ^ R(~(X,Y))). 
4.2 Bon restricted to moral qualities 
A subsense of bon occurs in conjunction with persons 
or entities exhibiting some form of moral qualities, 
as in une bonne personne. The treatment is the same 
as in the above section, but the selection of the pred- 
icates r = Fi (X, Y) in the telic of the noun's qualia 
must be restricted to properties related to the moral 
behavior (charity, compassion, integrity), or, by ex- 
tension, to some psychological attitudes and cogni- 
tive capabilities (e.g. a good researcher). 
This case is interesting for the development of the 
Qualia structure in the sense that it motivates sub- 
divisions, or at least a typing of the predicates in the 
Qualia. This is advocated in (Pustejovsky 95), but 
not really made concrete, as far as we know. 
4.3 Bon as an intensifier 
Another main role of bon is to emphasize a quality 
of the object denoted by the noun. If the object is 
e.g. of type: 
• +edible, as in un bon plat/repas (a good 
dish/meal), then the meaning of bon is deli- 
cious, excellent, 
• +fine-art, as in un bon film/tableau/livre (a 
good film/ painting / book), then the mean- 
ing of bon is of good expressive and intellectual 
quality, 
• +smell, as in une bonne odeur (a good odor), 
then the meaning is pleasant. 
zFrom these examples, we can conclude that, 
when bon modifies a nouns, then there is a certain 
property associated with the telic of the noun that 
produces a certain pleasure. For example, watching 
a good film entails a certain pleasure, eating a good 
meal entails another form of pleasure. Let us con- 
sider again that the noun N modified by the adjec- 
tive bon is represented by the variable Y, the entity 
undergoing the pleasure is not explicit in the NP, it 
is represented by X in the scope of a A-abstraction: 
XX, \[ .... t CAUSE(\[ .... t F,(X,Y)\], 
\[state BE+char,+ident(\[thing X \], 
\[~a,e AT+ps¢(\[+p,op pleasure \])\])\])\]. 
We have here a second type of representation for 
bon, which is also quite far from those developed in 
the GL (see above). The term 'pleasure' is a di- 
mension that refers to an ontology describing e.g. 
mental attitudes and feelings. A primitive such as 
TOWARDS could also be used instead of AT, to 
indicate a path instead of a place. Similarly, we 
are developing for psychological verbs and verbs of 
feeling a more appropriate primitive system (Kamel, 
forthcoming), e.g. with a primitive such as FEEL, 
which could be appropriate also for the above rep- 
resentation. The notion of event at the higher level 
in the LCS can be subject to debates because there 
is also an idea of systematicity and regularity which 
must be captured. The semantic composition form 
is similar as in 4.1. 
4.4 Bon as a quantifier 
In expressions indicating measures where bon is 
combined with a kind of measure as in: 
un bon verve/litre/m~tre 
(a good glass, liter, meter), 
Saint-Dizier 127 Sense Variation and Lexical Semantics 
bon indicates that the measure is slightly above 
the unit considered. It therefore contributes to the 
quantification. 
Bon can also be a quantifier in: 
un bon salaire, une bonne somme (a good salary / 
amount) 
which indicates that the level of the salary is above 
the average. For these expressions, the semantic rep- 
resentation is that of quantifiers. 
4.5 Bon meaning exact or correct 
There are a few acceptions such as: 
un bon raisonnement, calcul, une bonne information 
(a good reasoning/computation/information), or 
un bon ticket / billet (a good ticket/bank note, op- 
posed to fake) 
which may mean 'correct', 'valid' or 'exact'. 
Another meaning of une bonne in/ormation, un 
bon catcul is a useful information and a well-planned 
affair. These forms are rather semi-fixed and prob- 
ably fall within the case of co-composition. 
5 Selection and Dimensions of 
Interpretation 
Selection is directly related to the problem of sense 
delimitation since it is more or less complex, depend- 
ing on the 'breadth' of a sense. The larger a sense 
is, the more complex is this phenomenon. It also 
well illustrates the use of LCS and how the meaning 
of a proposition is computed from the fundamental 
structure and semantics of the verb, its arguments 
and the taking into account of usage variations. 
We hypothesize that a verb sense receives a single 
LCS representation, possibly largely underspecified, 
and a list of instanciations constrained by the nature 
of the arguments and also possibly by pragmatic fac- 
tors. The global format is given informally for the 
Sell verbs in section 5.2. This 'polymorphic' repre- 
sentation is the representation of the verb. Usage 
variations entailed by metaphors or metonymies do 
not alter the meaning of the verb, but they operate 
on, and alter the meaning of, VPs or propositions. 
In a selection situation, the verb meaning becomes 
more specialized (a definition of subsumption in LCS 
is given in (Dubois and Saint-Dizier 96)). In the 
Generative Lexicon (Pustejovsky 91, 95), selection 
is treated by selective binding, which is an opera- 
tion entirely based on type concordance and type 
subsumption. No attempt seems to be made to con- 
struct a meaning representation, which is not the 
goal of the GL. 
Let us now present a few examples. Note that 
the verb classes considered here are those we have 
defined for French, they do not necessarily overlapp 
with those defined in (Levin 93), (Saint-Dizier 96b). 
5.1 The case of Construction verbs 
The construction verb class includes verbs like con- 
struire, batir, ddifier, rdaliser, composer, (build, 
construct, realize, compose), etc. Let us concentrate 
on the verb construire, which includes usages such 
as: 
construire une maison / un cercle / un projet /une 
relation. 
The sense variation goes from a central meaning with 
a concrete, physical object to an abstract object. 
The general representation of this verb is: 
A J, 1, \[event CAUSE(\[thin9 I \], 
\[¢~ent GO+ch~r,+ident(\[pc~th FROM+char,+ident( 
\[+prop EPIST - STATUS(d) = non - ezist \]), 
TO+eha~..+ident(\[+prop EPIST- STATUS(J) = ezlst \]), 
FRO M+cha,,+id~nt (part -- of(J)), 
VlA+ehar,+ident (definition - cor~$titutive(J))\])\]. 
which describes the coming into being of J. Two 
functions, related to lexical data, are used: part- 
of(J) which gets the parts of J, and definition- 
constitutive(J) which gets the definition of J (e.g. a 
circle is a set of points equidistant from a particular 
point: the center). If this definition is not available 
in the lexical entry corresponding to the lexeme J, 
then the fimction remains as it is, justs stating that 
J has a certain constitutive definition. 
Construire is probably the generic element of the 
class. If we consider the following sense of the verb 
composer, which is more specific, as in: composer 
une sonate (to compose a sonata), which is basically 
restricted to musical pieces (imposed by constraints 
proper to the verb), we get exactly the same phe- 
nomena and restrictions. Note that this verb has 
metaphorical extensions such as composer un menu 
/ une salade (to compose a menu, a salad) where 
the property outlined is that the menu or the salad 
is going to look like a piece of art. These extensions 
are treated exactly as above. The form se composer 
un visage (to compose one's face = to hide his opin- 
ions/feelings) is also metaphorically derived from the 
sense considered here, but is rather a semi-fixed form 
since it is quite remote from the original sense and 
weakly compositional. 
5.2 The Sell verbs 
The 'sell' verb class introduces a simple default rep- 
resentation. Let us consider the verb vendre (sell), 
generic element of the class. Its basic argument is a 
physical object (which has an intrinsic value). Be- 
sides this usage, we have slightly more metaphorical 
ones, such as: 
Saint-Dizier 128 Sense Variation and Lexical Semantics 
II 
II 
II 
II 
m 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
!1 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
II 
vendre des r~ves / des illusions (to sell dreams / il- 
lusions), 
if we assume that, in this latter case, a dream has 
no intrinsic value, it is its association with vendre 
which makes it emerged via the expectations on 
the argument. We also have expressions like ven- 
dre quelqu'un = to betray someone. These usages 
defines the possible sense variations of the verb sell. 
We can then say that these objects, in association 
with verbs of the 'sell' class (and a few other classes 
as well) get e.g. a fictive value, represented by the 
function: FICTIVE-VALUE(J), and there is also a 
type shifting on J. 
The basic representation of sell is the following: 
hi, J, K, \[event CAUSE(\[thing I \], 
\[.re., GO+po.s(\[thing J \], 
\[path FaOM+possf\[thin9 1 \]), TO+l~oss(\[thln 9 K \])\]), 
ao+poss(\[thin" P \], 
\[path FROM+po,.,(\[thin9 K \]), TO+po,s(\[thinu I \])\])\])\] 
ADEFAULT(P, VALUE- OF(J), J, 
TYPE(J) = physical - object, COERCED - TYPE(J) = no.e). 
where P is the anchoring point for the default, ac- 
tivated when J, the variable concerned, is of type 
physical-object. In this case, which is the standard 
one, J need not be coerced to any other type. The 
default representation represents the basic usage, for 
the other cases, this default is not used and other 
types of representations are anchored at P. 
The general form of a default is then: 
DEFAULT(anchor, representation, variable con- 
cerned, expected type for argument, coerced type 
if appropriate). 
When the type is not physical object, then a differ- 
ent value is anchored to the position P, as explained" 
above. The other possible values may equivalently 
(1) be specifi.ed in the representation of the verb, 
similarly to the default, but not with the status of a 
representation by default, and associated with con- 
straints of use, or (2) by means of a rule. If the first 
case is chosen (with constraints on the type of the 
object), then it has the following form: 
OTHER-Iq.EPT(P, FICTIVE-VALUE(J), J, TYPE(J)= 
abstract artefact, COERCED-TYPE(J) = physical-object). 
5.3 The Measure verb class 
The measure verb class includes verbs such as: 
4valuer, mesurer, apprdcier, explorer, etc. (evaluate, 
measure, appreciate, explore). They can be repre- 
sented by an LCS form, but this form needs to be 
paired with additional information. Of interest is, 
for example, the quality or certainty of the measure, 
which can be best represented by a non-branching 
proportional series (Cruse 86) where the scale orders 
verbs by increasing precision of the measure. 
The object argument J of these verbs may be very 
diverse. It is however related to a measurable prop- 
erty (directly or metaphorically measurable). In the 
following LCS, we introduce the conceptual category 
'epistemic', as defined in (Pinker 93). We first have 
the extraction of the property being measured, then 
state that it becomes known, and finally indicates 
that the value becomes known to the subject I. The 
sense variations introduced by the argument J is cap- 
tured by the function PROPERTY-OF(J) which ex- 
tracts a property in J (these functions are advocated 
in very recent works by Jackendoff (Jackendoff 97)). 
The representat!on is the following: 
hi, J, P, \[event CAUSE(\[thi~t9 1 \], \[ev,at ACT+epist (\[thin9 1 \], 
TOWARD+ehar.+ident (\[event 
GO+ekar.+ident(\[ P = PROPERTY(J)\], 
\[path FROM+ehar,+ident( 
\[p,'o;, EPIST - STATUS(VALUEOF(P)) = unkonwn\]), 
\[path TO+cha.,+id.,*t( 
\[peop EPIST - STATUS(VALUEOF(P)) = known\]), 
FOa+.pist(\[e,J.,tt GO+epi.t(\[thin. VALUEOF(P)\], 
\[path TO+epist(\[thin9 Z \])\]))\]. 
If the object is not directly measurable, then a 
'metaphorical' value can be associated with it. The 
primitive FOR indicates the goal. 
6 Computational Aspects: towards 
constraint-based systems 
LFrom a computational perspective, we think that 
programming languages based on unification and 
the generate and test strategy may be too weak to 
handle the problem of polysemy in general. We 
indeed need languages which support the idea of 
e.g. conveying sets (disjunctions) of potential solu- 
tions. Constraint-based approaches handle domains 
of potential assignements for variables; domains be- 
come restricted as soon as constraints are formula~;ed 
on them, via dedicated constraint resolution mecha- 
nisms. 
Constraint logic programming (e.g. (Benhamou 
and Colmerauer 93)) is one such paradigm where 
logical implication has been paired with other mech- 
anisms for handling various forms of constraints, in 
particular on finite domains. In that case, more or 
less complex algorithms have been developed and 
integrated into logic programmes. These algorithms 
basically handle classical operations on sets (e.g. in- 
tersection). 
For example, in sentences such as: 
John knows+wh,-wh \[that whales are mammals\]_wh 
and \[whether they have lungs or not\]+wh. 
Mary enjoys \[the film\]object and \[eating ice- 
Saint-Dizier 129 Sense Variation and Lexical Semantics 
cream\]event • the verbs know and enjoy must select two a priori 
incompatible types, syntactic or semantic. For the 
latter sentence, a rule of the following form, in DCG 
associated with constraints can be used: 
vp --> v(Sel-restr), sentence(Sem), 
{domain(X,Sel-restr), domain(Y,Sem), 
included-into(Y,X)}. 
sentence (Sem) --> 
sentence(Seml), \[and\], sentence(Sem2), 
~domain (X, Sere), domain (Y, Seml), 
domain (Z, Sem2), 
included-into (Y, X), included-into (Z, X) }. 
Between brackets are the constraints, which are 
not interpreted as Prolog terms, but as predicates 
related to set manipulations. The predicates doma±n 
and £ncluded-£nto express constraints, they are 
therefore not evaluated to true or false. 
The same approach can be used to handle type co- 
ercion in the GL, where a domain for a variable can 
be the (finite) set of all the types the entity denoted 
by the variable can be coerced to. The domain then 
expresses the generative expansion of the entity. No 
commitment to any particular derived type is made 
a priori. 
7 Conclusion 
In this paper, which is mainly a feasability study, we 
have proposed some generative principles, based on 
rules, which possibly use some aspects of the Qualia 
structure, in particular the contents of the telic role. 
This approach supports the idea of an enriched se- 
mantic composition system, as advocated earlier by 
Jackendoff, instead of lexical redundancy rules or 
type coercion. This approach includes the construc- 
tion of the meaning of a proposition, taking into ac- 
count sense variations. The analysis is carried out 
using notions of verb semantic class and Lexical Con- 
ceptual Structure, which both give a certain degree 
of genericity to our descriptions. 
It is clear that this work is essentially preliminary, 
it needs a lot of extensions and more fruitful com- 
parisons with the GL. 
Acknowledgements I would like to thank Alda 
Mari for working jointly with me on preliminary 
studies from which this work emerged. I also thank 
James Pustejovsky, Federica Busa, Mouna Kamel 
and Fran~oise Gayral for discussions which greatly 
helped improving this work. 

References 
Benhamou, F., Colmerauer, A.,(eds.), (1993), 
Con- 
straint Logic Programming, MIT Press. 
Bnsa, F., (1996), Compositionality and the Seman- 
tics of Nominals, PhD. Dissertation, Brandeis Univer- 
sity, MA. 
Copestake, A., Briscoe, T., (1995), Semi-Productive 
polysemy and sense extension, journal of semantics, vol. 
12-1. 
Cruse, A., (1986), Lezical Semantics, Cambridge Uni- 
versity Press. 
Dubois, D., Saint-Dizier, P., (1996) Construction et 
representation de classes sdmantiques de verbes: une 
eoopdration entre syntaxe et cognition, in proe. RFIA96, 
Rennet, France. 
Fellbaum, C., (1993), "English Verbs as Semantic 
Net", Journal o\] Lexicography. 
Goldberg, A., (1994), Constructions: A Construction 
Grammar Approach to Argument Structure, University 
of Chicago Press. 
Jackendoff, R., (1990), Semantic Structures, MIT 
Press. 
Jackendoff, R., (1997), The Architecture of the Lan- 
guage Faculty, MIT Press. 
Lakoff, G., Johnson, M. (1980), Metaphors we Live 
By, University of Chicago Press. 
Levin, B., (1993), English verb Classes and Alter- 
nations: A Preliminary Investigation, Chicago Univ. 
Press. 
Mari, A., (1997), Une analyse de la gdn~rativitd en 
s~mantique lexicale utilisant la structure lexicale con- 
ceptueUe, reserach report universitd de Lausanne and 
IRIT. 
Marl, A., Saint-Dizier, P., (1997), Gdndrativitd: au- 
delh d'une th~orie des types, in proc. RFIA, Grenoble, 
IMAG. 
Martin, R., (1979), revue de linguistique, vol. 17. 
Paris. 
Moravcsik, J.M., (1975) Aitia as a Generative Factor 
in Aristotle's Philosophy, Dialogue, 14. 
Nunberg, G.D., Zaenen, A., (1992), Systematic Poly- 
semy in Lexicoilogy and Lexicography, proc Euralex92, 
Tampere, Finland. 
Ostler, N., Atkin~, S., (1992), Predictable Meaning 
Shifts: some lexical properties of lexical implication 
rules, in J. Pnstejovsky and S. Bergler (eds.) Lexical Se- 
mantics and Knowledge Representation, SApringer ver- 
lag. 
Pinker, S., (1993), Lcarnability and Cognition, MIT 
Press. 
Pnstejovsky, J., (1991), The Generative Lexicon, 
Computational Linguistics, vol 17-4. 
Pustejovsky, J., (1995), The Generative Lexicon, MIT 
Press. 
Saint-Dizier, P., (1996), Verb semantic classes based 
on 'alternations' and on Word.Net-like semantic criteria: 
a powerful convergence, in proc. Predicative Forms in 
Natural language and in lexical knowledge bases, I1LIT, 
Toulouse, to appear in a volume published Kluwer Aca- 
demic (P. Saint-Dizier, ed.). 
Saint-Dizier, P., (1998), Generativity in Lexical Se- 
mantics: Rules versus Qualia ?, to appear in a volume 
under preparation, P. Bouillon and F. Busa (eds.), CUP. 
Winston, M.E., Chaffm, R., Hermann, D., (1987), A 
taxonomy of part-whole relations, Cognitive Science, vol. 
11. 
