A model for the interaction of lexical and non-lexical 
knowledge 
in the determination of word meaning 
Peter Gerstl 
IBM Germany, Scientific Center 
Institute for Knowledge Based Systems 
Schloflstr. 70 
7000 Stuttgart 1 
e-mail: gersti @ dsOiilog.bitnet 
Abstract 
The lexicon of a natural language understanding system that is not restricted to one 
single application but should be adaptable to a whole range of different tasks has to 
provide a flexible mechanism for the determination of word meaning. The reason for 
such a mechanism is the semantic variability of words, i.e. their potential to denote 
different things in different contexts. The goal of our project is a model that makes 
these phenomena explicit. We approach this goal by defining word meaning as a 
complex function resulting from the interaction of processes operating on knowledge 
elements. In the following we characterize the range of phenomena our model is 
intended to describe and give an outline of the way in which the interpretation 
process may determine the referential potential of words by the integration and 
evaluation of a variety of factors. 
1 Introduction 
A system with the capability of natural language understanding typically relies on knowl- 
edge about a restricted domain of application. For example, as a natural language com- 
ponent of an information system, it needs to be able to identify the relevant linguistic 
patterns. In case of an information system for flight scheduling words such as "plane", 
"departure", "late", ...will typically be more relevant than for example: "pahn", "this- 
tle", "pine" which might be appropriate for a different domain. In any event, there will be 
a whole range of words that are commonly used in conversation and, thus, are indepen- 
dent from the choice of a specific domain. It is therefore desirable to have a multi-level 
architecture which can be adapted to different domains without being forced to redesign 
the whole system. A text understanding system based on this kind of architecture would 
provide the kernel functionality that allows it to couple principles and mechanisms not 
immediately dependent on the domain a specific implementation of the system will be 
used for. The main problem of such a modular architecture is how and where to draw 
the boundary between domain-independent and domain-specific knowledge. There are at 
least two more reasons which motivate a domain-oriented design strategy: 
With regard to knowledge representation, the history in artificial intelligence research 
has lead from early enthusiastic plans of 'general problem solving capabilities' to more 
realistic applications of expert systems. One reason was the huge amount of data that 
would have to be represented together with a large set of regularities introducing a level 
of complexity which could not, be handeled in a realistic manner by the systems currently 
165 
available. Another problem is the inconsistency of data that would necessarily arise once 
a lot of different and sometimes conflicting information had to be integrated into a sin- 
gle knowledge base. Under this perspective task- and domain-orientation is a matter 
of rendering the knowledge base manageable and to allow reasoning processes to draw 
meaningful inferences on the basis of consistent data. 
The second argument in favour of a domain-oriented design strategy comes from the 
area of lexical semantics. It is a well known fact that the meaning of a word depends 
on a multitude of contextual influences. In a very broad notion of context the task and 
the domain of a text understanding system may be considered a part of the context that 
licences an effective restriction of the 'semantic scope' of a single word. This again is 
mainly an argument of tractability which in this case helps to minimize the amount of 
lexical information needed. In our example it is a natural design decision to assume that 
the lexicon of a language understanding system as part of an information system about 
flight schedules does not have to account for the 'plant'-reading of "plane". 
2 The variability of words 
The way in which a word might contribute to the determination of the meaning of linguistic 
expressions is indeterminate in different ways. W. Labov calls the semantic potential 
of words enabling them to constitute various links between linguistic expressions and 
elements of the domain (semantic) variability. It is this potential which is responsible 
for the already mentioned context dependence of word meaning. An important goal of 
our model is to classify types of variability according to a set of more or less specific 
properties. The questions guiding this classification are: Which kind of representation 
does the variability affect; by which means are the variants related and, how can the 
referential potential be restricted in order to single out the intended meaning? In the 
linguistic tradition, these variability phenomena fall into one of four classes which we will 
sketch out below. 
2.1 Morphological ambiguity 
This class represents cases of identical surface representations of words extracted from a 
discourse. Depending on the kind of representation in which the natural language input 
is encoded (i.e. orthographic, phonetic, ... form) cases of homography or homophony may 
belong to this class or not. Homonymy as a specific instance of morphological ambiguity 
results from the identity of lexical base forms. In a lexicon using orthographic representa- 
tions as is normally the case in dictionaries (singular nouns, infinitive forms of verbs, ... ) 
there are for example two homonymous entries for "firm" (the adjective and the noun 
variant). In addition, morphological ambiguity captures the more general case of an in- 
flected form that is identical to a base form or to another inflected form. An example is 
the occurence of "saw" which depending on the context can be understood as noun, as 
base form of the verb "saw" or as inflected form derived from the base form "see". 
It is a notorious problem in lexical semantics \[Kooij 71\] to justify the distinction be- 
tween coincidental identity of forms (morphologicM ambiguity) and semantic variants of 
a single lexical item (chapter 2.2). In our approach it depends on the purpose of the 
system and, thus, is a design decision comparable to modelling conventions for the do- 
main knowledge. Identical basic word forms which are in no relevant and transparent 
166 
way related to the representation of knowledge about the domain will be represented by 
different lexical items. The question whether two identical forms 'collapse' into a single 
entry is then directed by the choice of the domain and the task analogous to the way in 
which drawing a boundary between elements of the domain is motivated. Defining two 
word forms as being homonymous yields the consequence that once the occurence in a 
discourse has been morphologically identified and thus mapped onto the corresponding 
lexical item it is no more possible to skip to a different homonymous variant. Words which 
shall not expose this behaviour should not be modelled as homonymous hut as semantic 
variants of a single lexical item. 
2.2 Polysemy and polyfunctionality 
In the preceding chapter we outlined the situation where two lexical items realize the same 
word form. The potential of semantic variation encoded in a single lexical item is known 
as polysemy. It remains in effect once the appropriate lexical item has been identified 
by means of morphological processes. A special case of polysemy is what \[Weber 74\] 
calls polyfuaclionalily refering to the situation where two variants of a lexical item belong 
to different syntactic categories. Polyfunctionality occurs very frequently since many 
lexical items allow identical realizations which belong to different syntactic categories 
being related by means of conversion or other morphological processes which do not affect 
the word stem. Examples are the nominal and verbal reading of "point" and the variants 
of "clean" which are categorized as adjective, adverb and verb. The comparison with 
variants in a dictionary is not as effective as it was in chapter 2.1 because the entries in a 
dictionary tend to conflate phenomena we call polysemy and cases of variability outlined 
in chapter 2.3. 
2.3 Metonymy and change of semantic type 
In the previous chapter we mentioned that a polyfunctional expression can be the analyzed 
as the realization of different categories in different contexts. The difference in semantic 
potential that arises from this variability sometimes not only involves the transition to 
a different semantic type in the sense of Montague grammar but it may be paralleled 
by a more or less extensive shift in conceptual interpretation (cf. the one-place versus 
two-place predicate reading of "drink"). Apart from ambiguities which are reflected by 
morphosyntactic properties of a lexical item (eg. its argument structure) there are in- 
stances of semantic variation allowing to change the interpretation of a class of linguistic 
items in a systematic manner. The crucial question is if this class should be characterized 
by lexical information or on the basis of regularities found in the domain. Metonymy is 
a specific instance of this type of variability where the different readings are related by 
elements of a set of fundamental relationships such as 'part-whole', 'cause-effect', etc. 
\[Nunberg 78\] investigates more general mechanisms that licence the use of a word in place 
of another in cases where both of them are related by means of a context-specific relation. 
The phenomena range from eases of systematic correspondence such as in the 'newspaper- 
example '1 to more ideosyncratic ones as the famous 'ham-sandwhich-case '2. As Nunberg 
notes these are not cases of linguistic ambiguity because pointing to the sandwhich would 
1 The word "newspaper" eazl be used to refer either to the publisher or to the publication. 
2A waiter might apply the expression "the ham sandwhich is sitting at table 20" in order to identify 
a unique guest who ordered a ham sandwhich. 
167 
serve the same purpose as the utterance of the complex phrase. Nevertheless, it is not 
clear if the relations are considered as instances of lexical or encyclopaedic knowledge. 
An example similar to the 'newspaper-case' is the potential of words such as "school", 
"opera", ... to select one of the alternative meaning variants 'building', 'process', 'insti- 
tution', etc. The approach outlined in \[Bierwisch 82\] derives this potential from system- 
atic relationships between concepts representing entities of the domain. The conceptual 
knowledge about social instutions has to provide the background information that "the 
parliament is at the end of the street" is semantically well-formed though "?the govern- 
ment is at the end of the street" is not. Since the iexical specifications of "parliament" 
and "government" cannot account for the fact that it is naturally assumed that the former 
can be associated with a specific building whereas a similar assignment is not possible for 
the latter. A comparable argument may be found for the 'substance'-reading of words 
naming trees. The iil-formedness of "?this table is made of plane" in contrast to "this 
table is made of oak" results from the non-verifiability in the common-sense model of 
the domain. If an expert would affirm that is quite common to use the wood from palm 
trees for the construction of tables we would probably change our model of the domain 
and licence the acceptability of the first sentence. This example is different from the 
'school'/'newspaper'-cases since in addition to the conceptual shift it involves a modifica- 
tion of semantic properties of the underlying lexical items. This in turn is an argument 
in favour of a lexicalist position which would classify this type of variability as cases of 
polysemy. \[Pustejovsky 90\] shows that a lexicalist approach to event structure allows to 
systematically characterize a whole range of type-shifting phenomena together with their 
consequences with respect to well-formedness conditions. 
As a consequence of seeking the portability of domain specific knowledge we follow 
the lines of Bierwisch in distinguishing lexical and conceptual information. Yet we do 
not reject the lexicalist position since we consider semantic type-shifting effects as driven 
by regularities of the conceptual structure. In the following section, we generalize this 
position to a systematic distinction between linguistic and non.lingustic knowledge. This 
allows us to keep variants introduced by linguistic ambiguities systematically apart from 
those cases we classify as non-linguistic variations. 
Following \[Binnick 70\] we define polysemous variants of a word as those cases of vari- 
ability which are (at least in principle) distinguishable on the basis of linguistic properties 
of the corresonding lexical item. Polysemous variants of a lexical item thus differ in at 
least one morphosyntactie or semantic property. For non-linguistic variants introduced by 
means of metonymy or type-shifting linguistic properties of a lexical item do not help to 
identify the intended reading because the variants have exactly the same linguistic prop- 
erties. This situation calls for the disambiguation potential of contextual information in 
order to reduce the 'semantic scope '3. 
2.4 Contextual relativity 
Vagueness and indexicality also belong to the class of variability phenomena. They are 
usually associated with specific groups of linguistic expressions (graduable predicates in 
ease of vagueness and deietie expressions in case of indexieality). In contrast to the 
effects introduced so far, in these eases the class of potential referents cannot simply 
3We consider the 'semantic scope' of a word as the possible range of interpretation implied by the 
literal use of a word. The more general notion of 'referential potential' additionally accounts for cases of 
conceptual shift as in the 'ham-sandwich' example. 
168 
be characterized by an enumeration of alternatives. As \[Pinkal 80\] points out it is an 
inherent property of vague predicates to provide a 'grey area' where the decision whether 
the predicate is applicable or not depends on the discourse context. It is even impossible 
to precisely delimit the area of positive or negative applicability. Following the lines of 
\[Bosch 83\] we do not consider vagueness as an isolated semantic property of a specific class 
of words but as an instance of the more general notion of context-dependence 4. Since this 
is an aspect of the referential potential of words our model has to cover theses phenomena 
as well. 
Indexicality is the potential of deictic expressions to select their meaning by exploiting 
peculiarities of the discourse situation. It is similar to vagueness since the implied referen- 
tial indeterminacy cannot be resolved independently from the specific discourse context. 
Yet, even deictie expressions are not immune to other types of variability. For example, 
the pronoun 'T' may be used to refer to an entity which is somehow related to the speaker 
in a certain discourse situation. An example is the utterance of "I am over there" with 
the speaker pointing to a desk. The expression 'T' in this case can be used to refer to 
the place, where the speaker usually works. This is an example of a systematic shift in 
meaning motivated by a conceptual relation. It thus belongs to the phenomena described 
in the previous chapter. 
Another instance of context relativity occurs in cases of privative opposition. This 
relativity results from the lack of semantic information for a specific word which could 
be provided by the use of a different word. According to \[Zwicky/Sadock 75\] "dog" is 
ambiguous between the readings 'male dog' and 'female dog' because it can be forced to 
provide both readings in sentences like "that is a dog, but it isn't a dog". In contrast to 
"?that is a lion, but it isn't a lion" it seems that a meaningful interpretation can be found 
for the former (the one which forces the selection of different variants for both occurences 
of "dog") whereas the latter leads to a contradiction. The choice of a variant could be 
forced by the use of "bitch" instead of dog which is not possible for "lion" since there 
is not regular lexical specification for something like "lioness". This illustrates the fact 
that lack of semantic information for a lexical item can under certain circumstances yield 
the same effect as a disjunction of alternative readings. This may occur whenever the 
semantic 'gap' can be filled by one of a small set of possible alternatives. 
3 A classification of knowledge types 
In order to have a precise representational basis for our model of word meaning, this 
chapter is intended to introduce the basic notions used to classify the relevant phenom- 
ena. The distinction between linguistic and non-linguistic knowledge mentioned in the 
preceding section constitutes the methodical basis of our model. Up to a certain degree, 
this distinction allows an independent examination of properties characteristic for only 
one type of knowledge. By assuming this distinction we do not claim that linguistic and 
non-linguistic knowledge are in a way fundamentally different. We use this distinction as a 
methodological tool that makes it possible to isolate certain aspects of word meaning not 
directly involving the whole range of both types of knowledge. In the course of stepwise 
extending the complexity of interrelations between linguistic and non-linguistic knowledge 
we will have to carefully analyze the tenability of this distinction. 
4 In general the notion of context dependence applies to referential expressions such as definite nominal 
phrases. We will restrict our attention to cases of context-dependence which apply to single words. 
169 
We account for possible similarities between both types of knowledge by using the 
same formalism for the representation of linguistic and non-linguistic knowledge. It is 
a variant of order-sorted predicate logic \[Nebel/Smolka 89\] which combines properties 
of the KL-ONE family of knowledge representation languages with properties of feature 
based unification grammars such as HPSG \[Pollard/Sag 87\]. In order to concentrate on 
the description of our model we will not go into the details of our formalism here s. 
The central components of our model are the iezicon on the linguistic side and the 
ontology on the non-linguistic side. The lexicon and the ontology provide the 'basic 
building blocks' of linguistic and non-linguistic knowledge respectively. The elements of 
the lexicon are called categories; the elements of the ontology concepts. It is important to 
note that the lexicon in our model integrates specifications of syntactic categories (i.e N, 
V, ..., N', ..., AP .... ) with lexical items. 
The formal means for the description of categories and concepts are sorts which are 
related by means of attributes and rules. Attributes may be used to express characteristic 
properties or relationships motivating the choice of a specific distinction between sorts. 
Rules on the other hand are not considered as tools for the description of inherent and 
permanent properties but as representations of regularities which might arise under certain 
circumstances. Apart from that, the collection of attributive characterizations has to be 
consistent. That is not necessarily required for the system of rules as a whole. The 
fundamental organizational principle of subsumption relates categories and concepts are 
licencing the inheritance of attributes between sorts. The subsumption order does only 
apply between elements of one and the same type of knowledge. The notation used for 
the description of sorts is a feature-logic as in \[Shieber 86\] for categories and a simple 
relational notation for concepts. We represent the fact that A subsumes B as A C B. 
Rules of grammar and rules of inference are represented by using simple predicate logic 
notion. Sorts, attributes and rules will be called knowledge elements. All of them put 
together constitute the knowledge base of our system. 
According to our argument in favour of a design strategy specifically tailored to the 
domain and the task the system is intended for, the structure of the ontology must be 
covered by an appropriate theory about entities of the domain, their inherent properties 
and the diverse aspects in which they are related. We reiterate this methodological claim 
here since a similar argument can be applied to the organization of the lexicon. The 
typical task of a text understanding system is to facilitate the analysis and production of 
textual input. It depends on the capabilities required whether certain aspects of this task 
involve restrictions on the set of relevant linguistic phenomena 6. The choice of lexical 
items depends on the domain at issue since the lexical inventory should cover at least 
the range of non-linguistic phenomena represented in the non-linguistic component of the 
system. 
Categorial knowledge provided by the lexicon together with the rules off grammar con- 
stitutes the descriptional apparatus for the classification of expressions. On the one hand 
expressions serve as input for linguistic processing and on the other hand they represent 
sequential patterns of written or spoken language. This intermediate status makes them 
5 Most of our assumptions about the representation of linguistic and non-linguistic knowledge are based 
on experiences gained from work in the LILOG-project at IBM Stuttgart. A description of formalisms and 
methods applied in this project can be found in \[Geurts 90\]. 
6 One can for example reduce the computational complexity by limiting the relevant sentence-level 
constructions to simple propositional clauses if the system is not meant to deal with other types of 
modality. Even if this argument sounds quite trivial the determination of a set of requirements for the 
linguistic component are as important as they are for the representation of domain knowledge. 
170 
elements of discourse knowledge. Expressions belong to the type of knowledge which serves 
as a kind of record for the registration of linguistic interactions together with their spatio- 
temporal specifications. It directly corresponds to episodic knowledge on the non-linguistic 
side. Episodic knowledge has the same intermeditate status as discourse knowledge since 
on the one hand it serves as a record of 'statements' and other 'experiences' with respect 
to the domain and on the other hand it is used to characterize entities from the domain 
as individuals on the basis of conceptual knowledge. Conceptual knowledge combines the 
structural information conveyed by the ontology with the additional information expressed 
by rules of inference. Individuals which result from the processing of certain linguistic ex- 
pressions are called referents since they are open to further reference by linguistic means. 
The figure below shows the whole classification assumed as the basis of our model. 
/ 
g~eral 
/ 
Categorial / \ 
elements ruleJ / \ 
Lexical Grammatical 
Knowledge 
Linguistic \ 
indlvldual \ 
Discourse 
Non-linguistic / 
$~nera! 
/ 
Conceptual / \ 
elemeuts ruleJ / \ 
Ontological Inferential 
\ 
individu61 \ 
Episodic 
Expressions Referents 
Figure 1: The classification of knowledge types. 
4 Word meaning 
The task of our model is an approach to the various aspects of word meaning which 
are responsible for the variability effects described in section 2. In order to reduce the 
complexity of the linguistic domain we restrict the relevant linguistic expressions to those 
representing simple word forms which cannot be further decomposed by mophological 
processes other than inflection. As a consequence, the granularity for the representation 
of linguistic knowledge treats basic morphemes as minimal elements. Another consequence 
is the width of the temporal grid which specifies the minimal 'temporal distance' between 
elements of discourse knowledge. We introduce temporal indices, allowing to subdivide 
the linguistic input into a chain of word-level segments. Each index uniquely identifies 
a gap between two words and directly corresponds to a set of intermediate results in 
the course of processing the input. These results are what we call the context. Formally 
speaking, a context is a set of factors marked by a temporal index specific to a certain stage 
of processing at which the system is observed. Factors are functions between knowledge 
elements or between instances thereof. They are elements of specific contexts and therefore 
differ from attributes because their existence is strictly tied to a certain stage of processing 
associated with a temporal index. As a result of iterated forwarding a factor may remain 
applicable during a sequence of processing steps. Factors may be classified according 
171 
to their origin. Factors which are directly derived from knowledge elements are called 
primitive factors. Depending on the type of knowledge elements involved we distinguish 
two modes of origin. Primitive factors can be ... 
• selected as instances of attributes or 
• established by the application of rules. 
Complex factors are derived from primitive ones by one of the following operations: 
• the restriction of the domain and/or range of a primitve factor 
• the application of set-theoretical operations on primitive factors 
• the functional composition of primitive factors 
We present this classification because our analysis of word meaning crucially depends 
on the notion of contextual factors. It is the main goal of our project to reconstruct 
word meaning as the result of the interaction of processes with cope with an effective 
integration of various linguistic and non-linguistic factors primitive and complex in nature. 
Since we investigate word meaning under the aspect of the potential of words to refer to 
representations of entities of the domain, word meaning in our terminology is a complex 
factor which links elements from discourse knowledge (expressions representing words) 
to elements from episodical knowledge (referents). The temporary status of factors is 
responsible for the fact that for the identification of the referential meaning of a word 
the whole context has to be taken into account. The linguistic notion of 'word meaning' 
therefore derives from the analysis of subsets of factors that result from the intersection 
of contexts present in a sufficiently large group of different uses of the same word. 
4.1 The constituents of reference 
In order to characterize the interrelation between factors introducing variability effects 
(productive factors) and those limiting the 'semantic search space' (restrictive factors) we 
need to examine the way in which word meaning can be decomposed into a small number 
of factors 7. A segmentation of the interpretation process according to our classification of 
knowledge types leads to three components which by application of functional composition 
constitute word meaning. Since components are derived by functional decomposition of 
a complex factor (word meaning) they are factors as well. Components may be further 
analyzed as the results of set-theoretic operations on basic factors some of which limit and 
some of which extend the 'semantic scope' of a word form s. Factors extending the scope 
of interpretation are directly responsible for the variability effects described in section 2. 
Factors constraining the scope of interpretation are the topic of chapter 4.2. The following 
list introduces the three components of meaning together with examples of the relevant 
productive factors. An interesting criterion for the classification of productive factors is 
whether they are established by the application of rules or selected from attributes be- 
tween knowledge elements. 
7We do not assume contexts to be finite but our approach relies on the fact that a finite subset of 
the context sumces to describe word meaning precisely enough to demonstrate the requirements a system 
with reasonable disambiguation capabilities has to fulfil. 
8 In fact the same function may in one stage of the interpretatio process serve as productive factor and 
in another as a restrictive factor. Thus, the property of productivity or restrictivity cannot definitively 
assigned to specific factors. More precisely speaking, it is property of factors dependent on the current 
stage of processing represented by the temporal index. 
172 
(1) Categorization 
Categorization as the first component of the chain maps expressions representing words 
onto lexical items. Its relevant productive factor is established by the application of mor- 
phological rules in some cases involving morphological ambiguity. 
The following categorization of "saw" is selected from the lexicon because of the identity 
of phonological forms: 
PHON 
SYN 
carl4 
V 
SEM 
t sawt 
MAJOR V 
TENSE PRESENT 
SUBCAT ~ ... ~ v...v ~ ... 
TENSE PRESENT 
SUBCAT ,~ 
conc54 
A morphological rule establishes the categorization of "saw" as an inflected form de- 
rived from the lexical base form for "see": 
carl6 
PHON f3rdsng(' see') 
\[MAJOR V \] 
SYN TENSE PAST 
SUBCAT ~ ... ~ v...v ~ ... 
SEM c0nc75 
(2) Lexical meaning 
The semantic specification of a lexical item and the properties of the corresponding 
concept are related by means of the sEi value. Two basic factors which contribute to the 
relevant productive factor of lexical meaning originate from attributes in the knowledge 
base by means of selection. The linguistic constituent of lexical meaning may involve 
polysemy or polyfunctionality if it provides a range of semantic alternatives and the cor- 
responding morphosyntactic properties for a single lexical item. 
(2a) The linguistic constituent of lexical meaning 
The subcategorization entry of the lexical item selects the following three s polysemous 
readings for cat14: 
• . . 
SUBCAT SYN NP\[NOM\] \] SEM cone 8 ~" 
\[ SYN NP\[ACC\] \] \[ SYN NP\[NOM\] \] 
V < SEM conclo ' SEM cone s > 
V < (\[ SYN PP\[WITH\] 1)\[ SYN NP\[ACC\] \] \[ SYN NP\[NOM\] \] ~ 
SEM c0nc35 ~ SEM cOnCl3 ~ SEM cone 8 
The non-linguistic constituent of iexical meaning is responsible for variabilities origi- 
nating from systematic relationships between different concepts related to a single lexical 
9As a matter of illustration the subcategorization frame does not exhaust the range of alternative 
readings. It again depends on the task of the linguistic component wether the lexlcal item has to provide 
further polysemous variants such as the intransitive reading of "see". 
173 
item by means of the semantic specification. 
(2b) The non-linguistic constituent of lexical meaning 
conc75 SITUITIO| E ... 
time : conc5 
location : cone3 
conc23 PERCEPTIO| E SITUATIO| 
actor : concs 
theme : cone13 
instrument : conc35 
conc34 REALIZATIO| E SITUATIO| 
actor : concs 
proposition : conc19 
conc39 VISTI|G~ITUATIO| 
visitor 
visited 
(3) Indlvlduation 
spatio-temporal properties 
The filler of the actor role visually perceives 
the filler of the theme role 
by using the filler of the instrument role 
The filler of the actor role 
realizes the filler of the proposition role 
SITUATIO| 
conc2 The filler of the visitor role 
conc 4 . . . 
This last factor in the chain of meaning components becomes established by the ap- 
plication of rules of inference. It maps concepts onto referents. The productive factor of 
individuation is referentiality extending the range of possible referents a concept can be 
individuated to. Referentiality here serves as cover term for the phenomena described in 
chapter 2.4 together with cases of conceptual variation which qualify as 'ad-hoc-anaphora' 
because they succeed to identify a unique referent in a specific context but cannot be char- 
acterized as instances of general principles guiding a shift in conceptual interpretation l°. 
Additional parameters of the discourse situation (the time and location of the utterance 
as well as a proposition rl) allow to establish an individuation which maps conc34 onto a 
referent r2 with the following properties: 
actor(r2) = rspeaker ^ 
proposition(r2) : rl ^ 
location(r2) = r a ^ r3 C Sdiscourse ^ 
time(r2) = r4 A r4 < tdi,¢our,e 
The three components of word meaning can be considered intermediate steps of the 
interpretation process. They may be analyzed and described in isolation since their in- 
teraction results from the way in which the range of the preceding component fits to the 
domain of the following. The task of the interpretation process on this background is to 
find a 'path' leading from an expression to an individual which under consideration of all 
the available contextual factors qualifies as plausible candidate for the referential meaning 
of the expression. 
4.2 How the semantic scope can be restricted 
The crucial question now is how the diverse components interact in order to reduce the 
range of word meaning by the exclusion of implausible variants. Here we pick out three 
l°Nunberg's ham-sandwich is an example instance of this kind of context specific ad-hoc-anaphora. 
174 
example groups of factors which in a typical situation may support the reductive factors 
of meaning components and such help to reduce the referential potential of a word. 
(1) Word-specific factors 
The first group are factors which result from structural relationships expressed by 
morphosyntaxtic attributes and rules of grammar. As we mentioned in chapter 2.3 these 
factors only Mfect variabilities which are introduced by the linguistic part of our knowledge 
base. Factors of this group thus may help to resolve cases of morphological ambiguity, pol- 
ysemy or polyfunctionality but they have no effect on variants that result from metonymy, 
change of semantic type or other instances of contextual relativity. 
Consider the following part of discourse: 
"I tried to find a possibility to escape. Then I saw a hole in the fence." 
We'll give a sketch of an analysis of the meaning of "saw" in this example on the 
basis of the knowledge elements introduced in the previous section. The rules of gram- 
mar suppress the nominal reading of "saw" since the principles of X-syntax require the 
constituent "a hole in the wall" to be 'absorbed'. Morphological rules do not support 
the disambiguation process. On the contrary, their productive potential causes the in- 
troduction of the variant derived from the base form "see". The variant cat56 is ruled 
out because of incompatibilities between its subcategorization frame and the syntactic 
environment of "saw" in our example. 
ca¢56 
PHON 
SYN 
V 
SEM 
MAJOR 
TENSE 
SUBCAT 
V 
t sawt 
v 
PRESENT 
\[ SYN NP\[NOMI \] 
SEM concl8 ~> 
\[ SYN NP\[ACC\] I \[ SYN NP\[NOM\] \] 
SEM conc42 ' SEM cone18 
TENSE PRESENT 
SUBCAT ~ 
c0nc75 
The intransitive polysemous variant fails because of the same reasons as the nominal 
homonymous variant. The transitive reading of cat56 would force an optional preposi- 
tional argument to be headed by "into "11. Thus, the polysemous variant conc42 can be 
singled out purely on the basis of word-specific factors if the rules of grammar do not 
account for the adjunction of a locative PP with the head "in". In case the grammar 
licences the existence of a prepositional adjunct, our model of the domain would have to 
contribute the restrictive factor that the concept associated with "the fence" does not fit 
with conditions on 'sawing'-events. Since the reductive influence of word-specific factors 
ends with the selection of polysemous variants we cannot expect a further restriction of 
the 'semantic scope' without additionally considering other types of factors. 
11 In order to simplify the example this alternative does not occur in the feature structure of cat56. 
175 
(2) Selectional restrictions 
A different group of factors belongs to the semantic level 12 of our model. Factors in 
this group neither are clear instances of linguistic regularities nor of non-linguistic ones. 
They are partially linguistic and partially non-linguistic in nature and therefore considered 
as complex factors derived by the integration of elements from both types of knowledge. 
They may help to reduce variabilities affecting categorization or lexical meaning. Yet, like 
word-secific factors they do not constrain contextual relativity. 
Consider the following part of discourse: 
a piece of wood } "I saw in the bathroom." 
a cup of coffee 
The semantic specification for the internal argument of "saw" leads to a concept conc4~ 
representing a class of entities which qualify as fillers of the corresponding role in the 
conceptual representation of the 'sawing'-event. 
conc75 BAWl\]l(\] C SITUATIO\]I 
actor : concls 
obj ect : conc42 
instrument : conc29 
for example: a human beeing 
for example: a concrete object 
for example: a set of tools 
The compatibility between the semantic specification of the internal argument of the 
two polysemous readings of "saw" and the type specification of a role belonging to the cor- 
respondint SEM-value account for the existence of selectional restrictions. The conceptual 
representation of '% piece of wood" must be compatible with conc42 in order to establish 
the lexical meaning leading to the concept SaWI\]IG.EVE\]IT. In the case of "this hand-saw saws 
well" the external argument would because of requirements on fillers of conceptual roles 
have to be mapped on the instruemt role of conc75. The situation is more tricky if we 
compare the instances of the external argument of cat23 in the following example: 
(1) The policeman saw an accident. 
(2) *The ball saw an accident. 
(3) The automatic traffic control camera saw an accident. 
(4) ?The morning saw an accident. 
An interesting aspect of this phenomena is that selectional restrictions may be can- 
celled by contextual factors or by means of rhetoric devices. The example sentences show 
how difficult it might be to identify an obligatory set of selectional restrictions. Compar- 
ing (1) with (2) suggests being an instance of the concept PERS0\]I as a reasonable choice. 
Example (3) however shows that the critical property is something like 'having an optical 
sensoring mechanism capable of detecting objects'. Sentence (4) might imply a metaphor- 
ical interpretation in spite of its apparent semantic illformedness. This again yields an 
argument in favour of a domain-driven design strategy for the semantic level linking be- 
tween categories and concepts. 
12 The semantic level is the 'interface' between linguistic and non-linguistic knowledge represented by the 
SEM values in |exical items together with a set of rules which attune semantic specifications of argument 
structure to attributes of the corresponding conceptual definitions. 
176 
(3) The set of possible referents 
The last group of factors exemplified here are the only means available to reduce the 
semantic scope resulting from contextual relativity. As we saw in chapter 2.4 contextual 
relativity is a fundamental property all referential expressions have in common. Since the 
'semantic scope' introduced by variabilities of this type cannot be subdivided into a set 
of alternative readings neither lexical nor conceptual information does help to restrict the 
range of indiviuation. 
The only way out of this dilemma is to derive a set of possible referents from knowledge 
about the domain and from information occuring in the preceding part of discourse. The 
latter calls for an investigation of discourse properties on the basis of pragmatic devices 
such as the Gricean conversation principles. Bridging phenomena 13 as generalizations 
of anaphoric binding are promising candidates for an approach to the determination of 
possible referents. C. Sidner emphasizes: "anaphor interpretation can be studied as a 
computational process that uses the already existing specification of a noun phrase to 
find the specification of an anaphor" \[Sidner 83, p.269\]. The actual limits of a set of 
possible referents thus very much depend on the inferential capabilities of our system 
to reconstruct the conceptual relationships undelying text coherence. The notion of fo- 
cus presented in the work of Sidner certainly plays a crucial role in the reducion of the 
computational complexity a computation of all possible bindings would involve if realistic 
discourse situations were to be considered. 
5 Conclusion 
Up to now we merely picked a collection of phenomena with an impact on the referential 
potential of words ranging from lexical properties to 'genuine' discourse phenomena. We 
presented a moderately general framework designed in a way that each of these phenomena 
has a place to fit in. The example analysis of "saw" illustrated the interaction of elements 
of our model and thereby pointed to problematic aspects that yet have to be resoved. A 
crucial problem is to distinguish between linguistic aspects of lexical meaning which the 
lexicon has to account for from non-linguistic aspects which derive from relations in con- 
ceptual structure. The next step in the evaluation of our model requires the determination 
of criteria which help to keep linguistic and non-linguistic aspects of the semantic level 
apart. The ultimate goal of this distinction is to narrow down the flow of information that 
passes through the semantic 'interface' between linguistic and non-linguistic knowledge. If 
this strategy succeeds it should be possible to adapt the conceptual knowledge to different 
domains of application not affecting linguistic knowledge as the basis for capabilities of 
discourse understanding. 

References 
\[Bierwisch 82\] M. Bierwisch, "Semantische und konzeptuelle Repr/isentation 
lexikalischer Einheiten", R. RfiSi~ka / W. Motsch (eds.), Unter- 
suchungen zur Semantik, Studia Grammatica XXII, Berlin, 1982. 
R. I. Binnick, "Ambiguity and vagueness", Papers from the Sixth 
Regional Meeting of the Chicago Linguistic Society, Chicago, 1970, 
pp. 147-153. 
P. Bosch, "Vagueness is Context-Dependance: A solution to the 
sorites paradox", T. T. Bailmer / M. Pinkal (eds.), Approaching 
Vagueness, Elsevier Science Publishers B.v. (North-Holland), 1983, 
pp. 189-210. 
H. H. Clark and S. E. Haviland, "Comprehension and the given-new 
contract", R. O. Freedle (ed.), Discourse Production and Compre- 
hension, Norwood, New Jersey, 1977. 
Bart Geurts (ed.), "Natural Language Understanding in LILOG: An 
Intermediate Overview", IWBS Report 187, IBM Germany GmbH, 
Scientific Center, 1990. 
J. G. Kooij. "Ambiguity in Natural Language. An investigation of 
certain problems in its linguistic description", Amsterdam, London, 
1971. 
B. Nebel / G. Smolka, "Representation and Reasoning with Attribu- 
tive Descriptions", IWBS Report 81, IBM Germany GmbH, Scientific 
Center, 1989. 
G. D. Nunberg. "The Pragmatics of Reference", PhD Thesis, Repro- 
duced by the Indiana University Linguistics Club, 1978. 
M. Pinkal, "Semantische Vagheit: Phiinomen und Theorien: Teil I" 
Linguistische Berichte, 1980. 
C. Pollard / E. A. Sag, "Information-based Syntax and Semantics: 
Volume r', CSLI Lecture Notes Number 13 Stanford, 1987. 
J. Pustejovsky, "Semantic Function and Lexical Decomposition", 
Schmitz / Schuetz / Kunz (eds.), Linguistic Approaches to Artifi- 
cial Intelligence, Lang, 1990, pp. 243-303. 
S. M. Shieber. "An Introduction to Unification-BasedApproaches to 
Grammar", CLSI Lecture Notes Number 4, Stanford, 1986. 
C. L. Sidner. "Focusing in the comprehension of definite anaphora", 
M. Brady / Robert C. Berwick, (eds.), Computational Models of Dis- 
ourse, Cambridge, Mass. and London, 1983. 
H. J. Weber, "Mehrdeutige Wortformen im heutigen Deutsch: Stu- 
dien zu ihrer grammatischen Beschreibung und lexikographischen 
Erfassung", Tiibingen, 1974. 
A. M. Zwicky and J. M. Sadock, "Ambiguity tests and how to fail 
them", J. P. Kimball (ed.) ,Syntax and Semantics, pp. 1-36, San 
Diego, 1975. 
