The hyperonym problem revisited: 
Concep, tual and::lexical.:hierarchies.in:,Janguage,,generation::  
Manfred Stede 
Technical University of Berlin 
Dept. of Computer Science 
KIT Project Group 
10587 Berlin/Germany 
st ede~cs, tu-berlin, de 
Abstract 
When a lexical item is selected in the  
production process, it needs to be explained 
why none of its superordinates gets selected in- 
stead, since their applicability conditions are 
fulfilled all the same. This question has received 
much attention in cognitive modelling and not 
as much in other branches of NLG. This pa- 
per describes the various approaches taken, dis- 
cusses the reasons why they are so different, and 
argues that production models using symbolic 
representations should make a distinction be- 
tween conceptual and lexical hierarchies, which 
can be organized along fixed levels as studied in 
(some branches of) lexical semantics. 
1 Introduction 
Representations used in  processing 
owe much to the tradition of 'semantic net- 
works', which nowadays have been successfully 
formalized and organized especially around one 
particular kind of link between nodes: the ISA- 
link, which connects entities to subordinate en- 
tities. This link is, by definition, the root of 
the so-called 'hyperonym 1 problem': When a 
speaker utters a word, she presumably needs 
to retrieve a lemma from her mental lexicon, 
and the 'applicability conditions" of the lemma 
automatically render the lemma's hyperonyms 
also applicable, thus raising the question how 
the choice among a set of more or less specific 
words is made. 
In this paper, I briefly review approaches 
to the hyperonym problem in psycholinguis- 
tics, natural  generation, and lexical 
semantics. In doing that, I will refer to differ- 
ent branches of NLG according to their roots 
I Alternatively called 'hypernym' in many publica- 
tions: 'hyperonym" seems preferable, as the Greek root 
is 'hyper" (super) + 'onoma' (name). 
.... ~ ......................... • ....... : ........ ~ .: : ...... 
and main motivations. Generally acknowl- 
edged are the two poles of 'cognition-inspired' 
and 'engineering-inspired'  production: 
Cognition-inspired work (CI-NLG, for short) 
seeks to build models that replicate perfor- 
mance data and explain phenomena of human 
 production with the help of psycholog- 
ical experiments; engineering-inspired work (EI- 
NLG) seeks to build programs that provide lin- 
guistic output to some particular computer ap- 
plication. These goals are extremely different, 
and it seems that the gap between the respec- 
tive methodologies will persist for quite some 
time. In between the two, however, I would 
situate a third category, which may be called 
'linguistics-inspired'. For this branch, here ab- 
breviated as LI-NLG, the primary motivation 
is neither in modelling human performance nor 
in efficiently performing a technical application; 
rather, LI-NLG seeks production models that 
replicate 'competence data', i.e. that account for 
observed linguistic regularities, without con> 
miting to statements about the human produc- 
tion p~vcess. 
Arguing that progress hinges on a better un- 
derstanding of the structure of the mental vo- 
cabulary, which includes a clear picture of the 
nature of the ISA-link, I will sketch a framework 
of distinct (but related) conceptual and lexical 
hierarchies, which offers possibilities to account 
for at least some of the phenomena to be dis- 
cussed. 
2 The hyperonym problem 
Following tile psycholinguistics literature, the 
hyperonym problem is regarded as all aspect of 
lemrna retrieval. Roelofs \[1996, p. 308\] describes 
a 'lemma' as a representation of the meaning 
and the syntactic properties of a word, and the 
task of lemma retrieval as a crucial step in the 
93 
process of grammatical encoding, where build- situations of utterance. More concrete, given 
ing of a phrasal, clausal, or sentential structure a conceptual specification (in a wide sense, in- 
-requires the syntacti~information :thattemmas. : :ctuding,:eontextual. parameters=andcommun:iCa- - 
contain. 
Thus abstracting from .the other steps of lan- 
guage production (formulation, articulation) as 
well as from possible influences of context, the 
task is confined to retrieve a lemma that cor- 
responds to the Conceptual specification that 
is represented in some adequate way. For the 
psycholinguist, the~geneya~!_.prgb!em is that of 
convergence from an under-specified conceptual 
representation to one word that the speaker ut- 
ters. Levelt \[1989, p. 20I\] characterizes the hy- 
peronym problem: 
"There is one particularly nasty con- 
vergence problem that has not been 
solved by any theory of lexical access. 
l will call it the hyperonym problem 
\[...\]: When lemma A's meaning entails 
lemma B's meaning, B is a hyperonym 
of A. If A's conceptual conditions are 
met, then B's are necessarily also satis- 
fied. Hence, if A is the correct lemma, 
B will (also) be retrieved." 
The relation of hyperonymy is generally re- 
garded as transitive: If A is a hyperonym of 
B, and B is a hyperonym of C, then A is a hy- 
peronym of C. Following common practice, we 
call A a direct hyperonym of B, while it is only 
an indirect hyperonym of C. The same holds for 
the inverse relation, hyponymy. 
For CI-NLG, which is concerned with find- 
ing models that resolve the convergence prob- 
lem with the impressive speed displayed by hu- 
man speakers, the hyperonym problem is im- 
portant because it. serves to put implemented 
models of spreading activation to the test. For 
EI-NLG. on the other hand, it can usually be 
ignored, as most of today's practical applica- 
tions either do not require the production of a 
more general word (i.e.. there is a one-to-one 
mapping from concept to word) or can rely on 
fairly simple mechanisms that.,avoid ,lexical rep- 
etitions bv choosing from a fixed, pre-defined set 
of near-synonyms. For LI-NLG, the challenge 
of the hyperonynl problem is to explain how a 
sentence can be paraphrased by others that re- 
place a word by a hyperonym, and why speakers 
select from candidate hyperonyms in different 
rive goals), the task is to find the best candidate 
from a set of valid paraphrases, here especially 
on the grounds of replacing content words with 
hyperonyms. 
3 Psycholinguistic production 
models 
.... Lan gu age' prod n'ction ~m o dels~deve\[oped in- psy:--, 
cholinguistics are nowadays couched in neural 
network theory. Under debate are the computa- 
tional properties of the networks, i.e., the modes 
of activation spreading, tile existence of feed- 
back, of inhibitory links, etc. The main method- 
ological concern is to construct the models in 
such a way that they account for data gathered 
in human speech production experiments, of- 
ten involving production errors, which can shed 
light on the underlying mechanisms. 
A central point of content is the ques- 
tion whether the meaning of concepts and/or 
words is represented in a decomposed fashion 
or not. Here, the hyperonym problem is some- 
times used as evidence by proponents of non- 
decompositional models. Roelofs \[1996\], for in- 
stance, argues that if a number of nodes repre- 
senting semantic features are the basis for lex- 
ical access, in lemma retrieval it becomes ex- 
tremely difficult to control the activation spread 
in such a way that only the most specific lexical 
unit that combines these features gets selected. 
Roelofs concludes that a non-decompositional 
model is to be favoured: When lemma retrieval 
starts with activation of the 'lexical concept' 
FATHER, rather than with tile features MALE 
and PARENT, the output word will be father, 
without the danger of being outranked by a 
higher activation of parent (or person, or entity. 
presumably). 
This line is continued in a recent compre- 
hensive theory of speech production by Lev- 
elt. Roelofs, and Meyer \[1999\]. The focus of 
.this. theory_is more _on. the side. of.articulation, 
but their approach to (non-) decompos'itionan/:t 
hyperonyms follows the basic assumption just 
sketched. The model consists of three layers of 
nodes: A layer of concept nodes with labelled 
concept links, a layer of lemma nodes, and a 
layer of word form nodes that include morpho- 
94 
logical information. When a lexical concept is 
activated, the mechanism of activation spread- 
:ing ensures that ~the::~directly:..ecm:nected::lemma.... 
receives tile highest activation, and not a lemma 
associated with a hyperonym of the lexical con- 
cept (which is connected by an ISA-link). 
Working out the mechanics to ensure this 
behaviour is important for the implementa- 
tion, but from the particular viewpoint of word 
choice, approaches of this kind are not very ex- 
planatory. Levelt. et.al. :\[1999, ~..~,4\]i istate that 
"there is not the slightest evidence that speak- 
ers tend to produce hyperonyms of intended 
target words." But when lexical access starts 
with an appropriately activated lexical concept, 
the problem is effectively moved away, into the 
realm of conceptualization. The authors ac- 
knowledge the need for a component that es- 
tablishes a 'perspective' by selecting a specific 
set of words, but have not incorporated such a 
component into their model. Thus, why and 
how the lexical concept receives its activation, 
and where the intention of using a word arises 
from, is not covered by the theory. For these 
questions, we have to turn to work in natural 
 generation. 
4 Hyperonyms in NLG systems 
In contrast to psycholinguistics-inspired work, 
the vast majority of natural  genera- 
tion systems uses computations based on sym- 
bol manipulation, often connected with sym- 
bolic knowledge representation and reasoning 
techniques. In these systems, the hyperonym 
problem as one aspect of the general task of 
lexical choice arises only in systems that em- 
ploy a sufficiently rich model of the lexicon and 
tile concept-lexicon link. involving some sort of 
hierarchy information. As pointed out above, 
from an application-oriented perspective (i.e.. 
in EI-NLG) it is often sufficient to work with 
rather limited mechanisms that largely eschew 
the lexical choice task. 
The earliest and very influential device for 
performing lexical choice, Goldman's-\[.1.-975\] 
discrimination net hard-wires the sequence of 
choice points leading to a specific lexical item, 
which is in fact the general strategy taken in the 
majority of NLG systems: if you have a choice. 
then prefer the most specific term. 
The most substantial criticism on the prefer- 
the-specific heuristic has been voiced in the 
work of Reiter \[1991\]. One of his examples 
:is. ~a.. system., ~as~zerhlg~:the-N.uestio n .*Is; .Ter~y:a 
woman? Even if the system has the specific 
knowledge that Terry is a bachelor, the response 
No, Terry is a bachelor would not be appropri-.. 
ate here; the less specific No, Terry is a man 
is better since it does not prompt the hearer to 
draw ally conclusions as to tile particular rele- 
vance of Terry's marital status for the present 
Lc0:n~ersa, tion, Reiter?s-. main -pointis:to distin- 
guish the knowledge a generation system has at 
its disposal from the communicative goals fol- 
lowed in producing an utterance. The latter 
are explicitly represented in his system as a. list 
of attributes 'to communicate about an entity', 
which is a subset of the overall knowledge the 
system has of that entity. In the Terry-example, 
the goal is to inform the hearer that Terry 
has the attributes {Human, Age-status:adult, 
Sex:Male}. 
In the KL-ONE \[Brachman, Schmolze 1985\]) 
style knowledge representation used by Reiter, 
concepts can be marked as 'basic-level' in the 
sense of \[Rosch 1978\]. Thus, on the taxonomic 
path Tweety (instance-of) Robin - Bird - Ver- 
tebrate - Animal - Object, the concept Bird is 
a basic-level one, which leads to a preference 
for using the corresponding lexical item when 
referring to some kind of bird (i.e., some con- 
cept or instance subsumed by it). Simultane- 
ous to Rosch's work, Cruse \[1977\] (who in turn 
was building on earlier research by Roger Brown 
in tile 1960s) had pointed out that tile failure 
to use items of "inherently neutral specificity" 
(a notion that closely corresponds to the basic 
level) results in unwanted conversational impli- 
ca.tures I tile hearer will surmise the existence 
of some reason why the neutra.1 term could not 
be used in the specific situation of utterance. 
But using the basic level is not mandatory. 
of course. Given a suitable context where at- 
tention is directed to particular attributes of 
entitities, a speaker moves to a more specific 
or sometimes to a more ~ general :level. ~:Reiter's 
mechanism of to-communicate attributes tries 
to capture this: Covering these attributes with 
a suitable term can override the preference for 
the basic level. Other kinds of preferences are 
also accounted for, such as favouring shorter 
rather than longer words, which typically (but 
95 
not always) co-incides with the basic-level pref- 
erence. Reiter notes that humans also employ 
...... - some preferences.t:hat can~otbe explained ~wi,th 
the parameters investigated so far. He gives 
the example \[Reiter 1991, p. 248\] of a speaker 
pointing the hearer to a cow and a horse with 
the utterance Look at the animals / mammals / 
vertebrates, t None of the terms is basic-level or 
signigificantly shorter than the others, yet there 
is a clear order of-'normality' in the sequence of 
the three candidates. 
In my own work on lexical choice in the 
'Moose' generator \[Stede 1999\], I used - 
neutral conceptual hierarchies and the sub- 
sumption relation, inter alia to account for the 
fact that different s occasionally dis- 
play preferences for different levels of specificity. 
For example, in hi-lingual instructional text we 
find a regular correspondence between the gen- 
eral English to remove and numerous more spe- 
cific German lexemes ( abziehen, abnehmen, her- 
ausdrehen, ...); this might very well be a genre- 
specific tendency. Furthermore, Moose employs 
a model of lexical connotations that can over- 
ride the general preference for a more specific 
lexical item. For example, when referring to a 
POODLE in a derogatory manner, Moose can 
choose the appropriately connotated word mutt, 
which requires moving up the taxonomy to the 
DOG concept, where a range of near-synonyms 
(differing in their connotations) are attached. 
Another reason for considering hyperonyms in 
the lexical choice process is to avoid repeated 
usage of the same term when referring to some 
object multiple times. 
In the present Moose implementation, all 
more general words are inherited to the concept- 
to-be-lexicalized, and the preference mechanism 
selects one of them (in case of absence of any de- 
cisive factors, it chooses the most specific word). 
This mechanism is certainly not cognitively ad- 
equate (it was not intended to be) and also not 
particularly efficient: The range of candidates 
under consideration should be constrained be- 
forehand. 
-In conclusion, NLG systems, employ a mix- 
ture of constraints and preferences in their ap- 
proaches to hyperonymy. The factors used by 
various systems in the choice process are: 
o User's vocabulary and knowledge (e.g.. 
\[Mcl(eown et al. 199:\]\]) 
. Successul reference, i.e., discrimination 
from other candidate entities (e.g., \[Dale, 
Reiter1995\]) .:: :- ........ ~' 
• Basic-level and entry-level effects, conver- 
sational implicatures 
® Length of words 
® Stylistic features such as formality, posi- 
tive/negative attitude 
• Language, genre _, 
• Givenness of item, avoid repetition or "say- 
ing the very obvious" 
Not surprisingly, there is no generator yet that 
would incorporate all these factors within a 
single system. It is not clear which general 
lexical items should be inherited down to the 
concept-to-be-lexicMized and enter the prefer- 
ential choice mechanism; it is also not clear how 
exactly the various preferences would interact 
and which would take precedence in a particu- 
lar situation of utterance. 
5 Hyperonymy in lexical semantics 
Linguists studying lexical semantics are to a 
good extent concerned with sense relations be- 
tween words, and hyp(er)onymy is certainly one 
of the relations receiving the most attention. 
While the intuitive decision whether some en- 
tity is subordinate to some other entity is in 
most cases not difficult to make, spelling out 
the precise definition of hyponymy (and thus 
hyperonymy) and its consequences is anything 
but trivial. Lyons \[1977\], for example, proposes 
that fish and bird share the direct hyperonym 
creature- but not animal. That is, when I say 
There were plenty of fish in the creek, tile al- 
ternative sentence There. were plenty of animals 
in the creek would not be a felicitous utterance. 
even thougil it is "trutl>conditionally correct". 
And hence, there is a difference between fish 
ISA creature and fish ISA animal. 
An interesting distinction in this respect is 
offered by Cruse \[1986\], who separates hy- 
ponymy_ from the more constrained relation .of 
taxonym, y. A diagnosis for the latter is the ut- 
terance frame X is a kind of/type of Y. Exam- 
pies that "work" in this frame are: spaniel-dog, 
rose-flou, er, mango-fruit. Examples that seem 
not to work are: kitten-cat, queen-monarch, 
spinster-woman, u,aiter-man. Notice t hat bot h 
96 
groups are perfectly compatible with the ISA- 
test, though: No one would doubt that a waiter 
IS A man, a.q-ueen IS A'.monarch. 
Taxonomies, as Cruse proposes, typically 
have no more than five levels, and frequently 
have fewer. The levels are commonly labelled 
as 'unique beginner' - 'life form' - 'generic' - 
'specific' - 'varietal'. (The origin of these term 
in biology is obvious, but they can be trans- 
creature creature 
animal bird /N 
dog cat dog cat b~ & A 
collie spaniel robin blackbM slarling collie spaniel robin blackbird starling 
Figure 1: Variants of taxonomy, reproduced 
from \[Cruse 1986, p. 146\] 
.6 Synthesis: Toward a model of 
ferred to otherweatms, as-t3ruse notes.) Most ..... .- ::.~..:.coneepCu:at.van@-:lexical inhe~itance 
important is the generic level, which holds or- Due to the very different motivations, different 
dinary everyday names like cat, apple, church, 
cup. These items tend to be morphologically 
simple and are not metaphorically transferred 
from elsewhere. Most branches of hierarchies 
terminate at the generic level, and hence this 
is the level with the largest number of items. 
Items at specific and varietal levels are particu- 
larly likely to be morphologically complex, and 
compound words are frequent here. 
From the notion of explicitly defined levels, 
it follows that hierarchies do not need to have 
nodes at each level. Consider the examples in 
figure 1. Depending on what items people place 
on the generic level, they end up with one of the 
two variants; according to Cruse, most people 
subscribe to the second, which holds dog, cat, 
bird on the same, generic level. Another ex- 
ample are musical instruments: Most of them 
belong to a kind such as strings, woodwind, 
brass, percussion, but there is no obvious kind 
for bagpipes or concertina, which are thus di- 
rectly linked to musical instrument. 
Cruse elaborated the importance of the 
generic level in \[Cruse 1977\], where he states 
that for every line of noun taxonomy, there is 
one term that is 'inherently neutral' (cf. the no- 
tion of basic level mentioned above). There is 
a general rule that requires speakers to use this 
term in order to obtain an unmarked utterance 
in a given context:-:-.unless.this would- result 
in an 'abnormal communication', in which case 
the speaker should deviate from neutral level, 
but only to the minimum degree required to en- 
sure normality. Cruse then offers several condi- 
tions that would license such over- and under- 
specification, which we do not reproduce here. 
kinds of NLG have very different approaches to 
the hyperonym problem. EI-NLG can basically 
ignore or finess it. In CI-NLG, it is reduced 
to a merely technical question: getting the me- 
chanics of spreading activation right, so that 
lexical convergence enables the subsequent pro- 
cesses of syntactization and articulation (which 
the CI-NLG models place their emphasis on). 
A broader view is necessarily based on reason- 
ing with speaker's goals and contextual features, 
which for the time being is the realm of LI- 
NLG. Thus, before embarking on building more 
comprehensive con.nectionist models, the hyper- 
onym problem is best studied in the frameworks 
of LI-NLG -- but with the motivation of mod- 
elling human performance taken into account. 
Thus adopting the perspective outlined in 
section 4, we are interested in choosing words 
between more or less specific alternatives as well 
as between near-synonyms of the same speci- 
ficity. We thereby open the door to both 'ver- 
tical' and 'horizontal' lexical choice within a hi- 
erarchy, which raises a number of questions: 
* What is the granularity of conceptual, and 
that of lexical knowledge? 
• How are tile differences between near- 
synonyms represented? 2 
• Given an activated concept, which more 
general lexical items are considered in tile 
choice process; are there any restrictions on 
.-lexical inheritance-?- ......... 
o How is the eventual choice from the set of 
candidate lexical items being made? 
2This question is beyond the scope of this paper; the 
kind of approach I have in mind here is represented in 
\[DiMarco et al. 1993\], \[Hirst 1995\], \[Edmonds 1999\]. 
97 
collie -- (a silky-coated sheepdog with a long ruff and long narrow head developed in Scotland) 
=> shepherd dog, sheepdog, sheep dog -- (any .of various usually long-haird breeds,of do.g, ~ .... 
reared to herd-and guard sheep) 
=> working dog -- (any of several breeds of usually large powerful dogs bred to work as 
draft animals and guard and guide dogs) 
=> dog, domestic dog, .Canis familiaris -- (a member of the genus Canis"(probably... 
=> canine, canid -- (any of various fissiped mammals with nonretractile claws and 
typically long muzzles) 
=> carnivore -- (terrestrial or aquatic flesh-eating mammal; terrestrial carnivores 
have four or five clawed digits on each limb) 
=> placental, placental mammal, eutherian, eutherianmammal -- (mammals having a 
placenta; all mammals except monotremes and marsupials) 
.`=>~mamma1~-~a~amm~c~.~ded~er~eb~rte.having.~t~he`~in.~mur~.`~r~ess~¢.~Yered.~ 
=> vertebrate, craniate -- (animals having a bony or cartilaginous skeleton... 
=> chordate -- (any animal of the phylum Chordata having a notochord or 
spinal column) 
=> animal, animate being, beast, brute, creature, fauna -- (a living 
organism characterized by voluntary movement) 
=> life form, organism, being, living thing -- (any living entity) 
=> entity, something -- (anything having existence (living or nonliving)) 
Figure 2: Hyperonyms for collie from WordNet 
As we have seen, present models that admit 
hyperonyms into the choice process (in particu- 
lar those of Reiter \[1991\] and Stede \[1999\]) run 
into the problem of overgeneration: Too many 
candidates have to be compared for their prefer- 
ential features, and it is not clear that a decision 
can always be made. 
To illustrate the question of granularity and 
range of hyperonymic alternatives, contrast the 
path from collie to creature given by Cruse 
\[1986\] in figure 1 with the hyperonym chain 
for collie offered by WordNet \[Fellbaum 1998\], 
shown in figure 2. The WordNet chain includes 
many items that clearly do not show up in ev- 
eryday  use, and that a lexical choice 
process should prefer not to consider when pro- 
ducing an utterance about a collie. Chordate, 
for example, would in the vast majority of utter- 
ance situations not be an option. On tile other 
hand, all these terms are certainly 'correct', and 
a system should be able to respond affirmatively 
to the question Is a collie a chordate ? 
This divergence points to the need for a dis- 
tinction between conceptuaJ,and lexicalg:ranu- 
laritv and inheritance: The WordNet chain rep- 
resents rather a series of concepts than of words 
entering the lexical choice process, which ap- 
pears to be better represented by a Cruse-type 
chain with few designated levels (but needs to 
be augmented with near-synonyms for tile 'hor- 
ing, ... ........... ffntity 
creature .... " ........ li~q form 
animal, beast, ... ..... animal 
c>rdate 
ve~ebrate 
~ammal 
p~cental 
c~71ivore 
7~_1 ine dog J 
/ . wyking dog 
shepherd dog 
/ 
collie .......... collie 
Figure 3: Active-lexical and conceptual hierar- 
chy 
izontal' aspects of choice). 
The resulting situation is sketched in figure 3. 
On the right hand side, the nodes of the concep- 
tual chain also are linguistic units, but in lan- 
guage production they would be accessed only 
, if tile. '.to~com munical~e".attdbutes ex.plicitly, call 
for it, e.g., when comparing chordates to verte- 
brates. Otherwise, only items oll tile left hand 
side (tentatively called 'active-lexical') enter tile 
lexical choice process, which are characterized 
by their particular level in the vocabulary struc- 
ture, and further differentiated by stylistic and 
98 
other features. The generic, or basic, level is P. Downing. "Factors influencing lexical choice 
marked by a box. in narrative." In: W. Chafe (ed.): The pear 
When a hyperonym chain is thus not.merely -- .... stories:, cognitive,: c,ultural~ .and:li.nguistic~as-. 
an ordered list, but the signficance of the levels 
is recognized (assuming that Cruse's proposal of 
level structure indeed scales up to other areas of 
vocabulary), rules for deviating from the generic 
level can be stated that map contextual param- 
eters onto 'level movement instructions'. These 
rules would extend the lexicalisation framework 
of Reiter \[1991\], w.he~e tthe£fivsg:Gon~tion .~is..ad- 
hering t~ the hard constraints (the word must 
convey the essential attributes that are to be 
communicated), and the second is a preference 
for the basic level. Adding the instructions 
for level movement would "contextualize" this 
framework. 
The rules for moving between levels have to 
consider the specific function of the NP (refer, 
inform about category membership, etc.) and 
other factors as indicated in the previous sec- 
tions (and others mentioned by Cruse \[1977\]). 
Since the roles and interactions of these fac- 
tors are not well understood yet, at this point 
CI-NLG can make important contributions by 
designing experiments that shed more light on 
the parameters that prompt speakers to deviate 
from the basic level; one example here is the 
study on speaker's lexical choices in narrative 
by Downing \[1980\]. 

References 
R. Brachman, J. Schmolze. "An overview of the 
KL-ONE knowledge representation system." 
In: Cognitive Science 9 (2), 1985. 
D. Cruse. "The pragmaties of lexical speci- 
ficity." In: Journal of Linguistics 13, pp. 153- 
164, 1977. 
D. Cruse. Lexieal semantics. Cambridge, \[_l\[(: 
Cambridge University Press, 1986. 
R. Dale, E. Reiter. "Computational Interpreta- 
tions of the Gricean Maxims in the Genera- 
tion of Referring Expressions." In: Cognitive 
Science 19:233-263, 1995. 
C. DiMarco, G. Hirst, M. Stede. "The semantic 
and stylistic differentiation of synonyms and 
near-synonyms." In: Working notes of the 
AAAI Spring Symposium on Building Lexi- 
cons for Machine Translation. Stanford Uni- 
versity, March 1993. 
pects of narrative production. Norwood/N J: 
Ablex, 1980 
P. Edmonds. "Semantic representations of near- 
synonyms for automatic lexical choice." PhD 
thesis, Department of Computer Science, 
University of Toronto, September 1999. 
C. Fellbaum. WordNet -- An Electronic Lexical 
:Database~C~mb~idge /MA : MI T . l~,ress, 199.8. 
N.M. Goldman. "Conceptual generation." In: 
R.C. Schank (ed.): Conceptual informa- 
tion processing. Amsterdam: North-Holland, 
1975. 
G. Hirst. "Near-synonymy and the structure of 
lexical knowledge." In: Working notes of the 
AAAI Spring Symposium on Representation 
and Acquisition of Lexica\] Knowledge. Stan- 
ford University, 1995. 
J. Lyons. Semantics. Volume I. Cambridge/UK: 
Cambridge University Press, 1977. 
K. McKeown, J. Robin, M. Tanenblatt. "Tai- 
loring lexical choice to the user's vocabulary 
in multimedia explanation generation." In: 
Proceedings of the 31st Annual Meeting of 
the Association for Computational Linguis- 
tics (ACL). Columbus, OH, 1993. 
W. Levelt. Speaking: From Intention to Articu- 
lation. Cambridge/MA: MIT Press, 1989. 
W. Levelt, A. Roelofs, A. Meyer. "A theory 
of lexical access in speech production." In: 
Behavioral and Brain Sciences 22, pp. 1-75, 
1999. 
E. Reiter. "A new model of lexical choice for 
nouns." In: Computational Intelligence 7, 
240-251, 1991. 
A. Roelofs. "Computational Models of Lemlna 
Retrieval." In: T. Dijkstra, K. de Smedt 
(eds.): Computational Psycholir~gui.~tic.~. 
London: Taylor & Francis. 1996. 
E. Rosch. "Principles of categorization." In: E. 
Rosch, B. Lloyd (eds.): Cognition and cate- 
gorization. Hilldale, N J: Lawrence Erlbaum, 
1978. 
• M: Stede. Lexicai semantics and tcrmwledge rep- 
resentation in multilingual text generations.. 
Dordrecht/Boston: Kluwer, 1999. 
