A Fast Algorithm 
for the Generation of Referring Expressions 
Abstract 
We simplify previous work in the development of 
algorithms for the generation of referring expre~ 
sions while at the same time taking account of psy- 
cholinguistic findings and transcript data. The result 
is a straightforward algorithm that is computation- 
ally tractable, sensitive to the preferences of human 
users, and reasonably domain-independent. We pro- 
vide a specification of the resources a host system 
must provide in order to make use of the algorithm, 
and describe an implementation used in the IDAS sys- 
tem. 
Introduction 
In previous work \[Da189,DH91,Rei90a,Rei90b\] we 
have proposed algorithms for determining the con- 
tent of referring expressions. Scrutiny of the psy- 
cholinguistics literature and transcripts of human di- 
alogues shows that in a number of respects the be- 
haviour of these algorithms does not correspond to 
what people do. In particular, as compared to these 
algorithms, human speakers pay far less attention to 
reducing the length of a referring expression, and far 
more attention to making sure they use attributes 
and values that human hearers can easily process; in 
the terms introduced in \[Da188,Da189\], hearers are 
more concerned with the principle of sensitivity than 
with the principle of efficiency. We have designed a 
new referring expression generation algorithm that 
is based on the~ observations, and believe that the 
new algorithm is more practical for real-world natu- 
ral language generation systems than the algorithms 
we have previously proposed. In particular, the al- 
gorithm is: 
• fast: its run-time is linear in the number of distrac- 
tors, and independent of the number of possible 
modifiers; 
• sensitive to human preferences: it attempts to 
use easily perceivable attributes and basic-level 
\[Ros78\] attribute values; and 
• Supported by SERC grant GR/F/36750. E-mail ad- dress is E.Reiter@ed. ac.uk. 
tAiso of the Centre for Cognitive Science at the Univer- 
sity of Edinburgh. E-mail address is R. DaleQed. ac .uk. 
Ehud Reiter*and Robert Dalef 
Department of Artificial Intelligence 
University of Edinburgh 
Edinburgh EH1 1tlN 
Scotland 
• domain-independent: the core algorithm should 
work in any domain, once an appropriate knowl- 
edge base and user model has been set up. 
A version of the algorithm has been implemented 
within the IDAS natural-language generation system 
\[RML92\], and it is performing satisfactorily. 
The algorithm presented in this paper only gener- 
ates definite noun phrases that identify an object 
that is in the current focus of attention. Algorithms 
and models that can be used to generate pronominal 
and one-anaphoric referring expressions have been 
presented elsewhere, e.g., \[Sid81,GJW83,Da189\]. We 
have recently begun to look at the problem of gen- 
erating referring expressions for objects that are not 
in the current focus of attention; this is discussed in 
the section on Future Work. 
Background 
Distinguishing Descriptions 
The term 'referring expression' has been used by dif- 
ferent people to mean different things. In this paper, 
we define a referring expression in intentional terms: 
a noun phrase is considered to be a referring expres- 
sion if and only if its only communicative purpose is 
to identify an object to the hearer, in Kronfeld's ter- 
minology \[Kro86\], we only use the modal aspect of 
Donefian's distinction between attributive and ref- 
erential descriptions \[Don66\]; we consider a noun 
phrase to be referential if it is intended to identify 
the object it describes to the hearer, and attributive 
if it is intended to communicate information about 
that object to the hearer. This usage is similar to 
that adopted by Reiter \[Rei90b\] and Dale and Had- 
dock \[DH91\], but differs from the terminology used 
by Appelt \[App85\], who allowed 'referring expres- 
sions' to satisfy any communicative goal that could 
be stated in the underlying logical framework. 
We here follow Dale and Haddock \[DH91\] in assum- 
ing that a referring expression satisfies the referential 
communicative goal if it is a distinguishing de- 
acription, i.e., if it is an accurate description of the 
entity being referred to, but not of any other object 
in the current context set. We define the context 
set to be the set of entities that the hearer is cur- 
rently assumed to be attending to; this is similar 
ACRES DE COLING-92, NAtCrES, 23-28 AO't~q" 1992 2 3 2 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
to the notion of a discour~ focus space \[GS86\]. We 
also define the contrast set to be all elements of the 
context set except the intended referent. The role of 
tile conlponcnts of a referring expression can then 
be regarded as 'ruling out' members of the contrast 
set. For example, if the speaker wished to identify a 
small black dog in a situation wlmre tile contrast set 
consisted of a large white dog and a small black cat, 
she might choose the adjective black in order to rule 
out the white dog and the heart noun dog in order 
to rule out the eat; this results in the referring ex- 
pression the black dog, which matches the intended 
referent but no other object in the current context. 
The small dog would also be a succ~csful referring 
expre~-~ion in this context, under the distinguishing 
description model. 
Unnecessary Modifiers 
A referring expression must communicate enough in- 
fornlation to be able to uniquely identify the in- 
tended referent in the current discourse context 
(i.e., it must adhere to the principle of adequacy 
\[Da188,Da189\]). But.this is not the only constraint 
a good referring expression must obey; it is clear 
that many referring expressions that meet this con- 
straint are inappropriate because they couvey incor- 
rect and unwanted conversational implieatures 
\[Gri75,Rei90a\] to a human hearer. 
One source of such false implicatures can he the pre~ 
ence of redundant or otherwise unnecessary modifiers 
in a referring expression. For example, consider two 
possible referring expressions that a speaker might 
use to request tilat a hearer sit by a talfle: 
(l) a. Sit by the table. 
h. Sit by the brown wooden table. 
If the context was such that only one table was vis- 
ible, and this table was brown raid made of wood, 
utterances (In) and (lb) would both be distinguish- 
ing descriptions that unranbiguously identified the 
intended referent to the hearer; a hearer who heard 
either utterance would know where he was supposed 
to sit. However, a hearer who heard utterance (lb) 
in such a context might make the additional infer- 
enee that it was important to the disc~mrse that the 
tM)le was brown and made of wood; for, tile hearer 
might reason, why else would the speaker include in- 
formation about the table's colour and material that 
was not necessary for the reference task? This infer- 
enc~ is an example of a conversational implicature 
caused by a violation of Grice's maxim of Quantity 
\[Gri75\]. 
Inappropriate Modifiers 
Unwanted conversational implicatures can alse be 
caused by the use of overly specific or otherwise un- 
expected modifiers. One example is ms follows: 
(2) IL l,ook at the doq. 
b. Look at the pil bull. 
In a context where there is only one dog present, 
tile hearer would nommlly expect utterance (2a) to 
be used, since dog is a basic-level class \[Ros78\] for 
most native speakers of English. Hence the use of ut- 
terance (2b) might implicate to the hearer that the 
speaker thought it was relevant that the animal was a 
pit bull and not some other kind of dog \[Cru77\], per- 
haps because the speaker wished to warn the hearer 
that the animal might be dangerous; if the speaker 
had no such intention, site should avoid using utter- 
ance (2b), despite the fact that it fulfills the referen- 
tial communicative goal. 
Previous Work 
In previous work \[Dalgg,D}191,Rei90a,Rei90b\] we 
have noted that the presence of extra information 
in a referring expression can lead the hearer to make 
false implicaturcs, and therefore concluded that a 
referring-expression generation system should taake 
a strong attempt to ensure that generated refer- 
ring expressions do not include unnecessary infor- 
mation, either as superfluous NP modifiers or as 
overly-specific head nouns or attribute values. Dale 
\[DaI88,DaI89,DH91\] }ins suggested doing this by re- 
quiring the generation system to produce mtnimal 
distinguishing descriptions, i.e., distinguishing de- 
scriptions that include as few attributes of the in- 
tended referent as possible. Keiter \[Keig0a, Rei90b\] 
ha.s pointed out that this task is in fact NP-Hard, 
and has proposed instead that referring expressions 
should obey three rules: 
No Unnecessary Components: all components 
of a referring expremion must be necessary to ful- 
fill the referential goal. For example, the small 
black dog is not acceptable if the black dog is a dis- 
tinguishing description, since this means small is 
an unnecessary component. 
Local Brevity: it should not be po~ible to pro- 
duce a shorter referring expression by replacing a 
set of existing modifiers by a single new modifier. 
For exanlple, the sleeping female dog should not 
he treed if the small dog is a distinguishing descrip- 
tion, since the two modifiers sleeping and female 
can bc replaced by the single modifier small. 
Lexical Preference: this is an extension of the 
ba.sicAcvel preference proposed by Cruse \[Cru77\]; 
more details are given in \[Reigl\]. 
A referring expression that mt.~ts l\[eiter's con- 
straints cart be foumt in polynomial time if the lexical 
preference relation meets certain conditions \[Rei90a\]; 
such a referring expression can not, imwever, always 
be found in linear time. 
Psychological and Transcript Data 
Psychological Evidence 
Subsequent to performing the above research, we 
have looked in some detail at the psychological lit- 
erature on human generation of referring expres- 
sions. This research (e.g., \[FO75,Whi76,Son85, 
A(X'ES DE COLING-92, NANrES, 23-28 AOl'rl' 1992 2 3 3 I'SOC. hi: COLING-92, NANTES, AUG. 23-28, 1992 
Pec891; \[Lev89, pages 129-134\] is a useful summary 
of much of this work) clearly shows that in many 
cases human speakem do include unnecessary modi- 
tiers in referring expressions; this presumably implies 
that in many cases human hearers do not make impli- 
catures from the presence of unnecessary modifiers. 
For example, if human subjects are shown a picture 
of a white bird, a black cup, and a white cup, and 
are asked to identify the white bird, they frequently 
say the white bird, even though just the bird would 
have been sufficient in this ease. 
A partial explanation for this use of redundancy may 
be that human speakers generate referring expres- 
sions incrementally \[Pee89\]. An incremental gener- 
ation algorithm cannot always detect unnecessary 
modifiers; in the above example, for instance, one 
could imagine the algorithm choosing the adjective 
white to rule out the black cup, and then the noun 
bird in order to rule out the white cup, without then 
erasing white because the black cup is also ruled out 
by bird. 
Another explanation of redundancy might involve 
the speaker's desire to make it easier for the hearer 
to identify the object; the speaker might believe, for 
example, that it is easier for the hearer to identify 
a white bird than a bird, since colour may be more 
immediately perceptible than shape. 1. 
Both of the above explanations primarily justify ad- 
jectives that have some discriminatory power even if 
they are redundant in this particular context. In the 
above example, for instance, white possesses some 
discriminatory power since it rules out the black cup, 
even though it does happen to be redundant in the 
expression the white bird. It would be harder for 
either of the above factors to explain the use of a 
modifier with no discriminatory power, e.g., the use 
of white if all objects in the contrast set were white. 
2qaere is some psychological research (e.g., \[FO75\]) 
that suggests that human speakers do not use mod- 
ifiers that have no discriminatory power, but this 
research is probably not conclusive. 
Thc argument can be made that psychological real- 
ism is not the most important constraint for gener- 
ation algorithms; the goal of such algorithms should 
be to produce referring expressions that human hear- 
ers will understand, rather than referring expressions 
that human speakers would utter. The fact that hu- 
man speakers include redundant modifiers in refer- 
ring expressions does not mean that NL generation 
systems are also required to include such modifiers; 
there is nothing in principle wrong with building gen- 
eration systems that perform more optimizatious of 
their output than human speakers. On the other 
hand, if such beyond-human-speaker optimizations 
1Another possible explanation is that speakers may in 
some cases use precompiled 'reference scripts' instead of 
computing a referring expression from scratch; such refer- 
enoe scripts specify a set of attributes that are included as 
a group in a referring expression, even if some members 
of the group have no discriminatory power in the current 
context 
are computationally expensive and require complex 
algorithms, they may not be worth performing; they 
are clearly unnecessary in some sense, after all, since 
human speakers do not perform them. 
Transcript Analysis 
In addition to the l~ychological literature review, we 
have also examined a transcript of a dialogue be- 
tween two humans performing an assembly task. ~ 
We were particularly interested in questions of mod- 
ifier choice; if a discriminating description can be 
formed by adding any one of several modifiers to a 
head noun, which modifier should be used? In par- 
ticular, 
1. Which attribute should be used? E.g., is it better 
to generate the small dog, the black dog, or the 
female dog, if these are discriminating descriptions 
but jnst the dog is not? 
2. Is it preferable to add a modifier or to use a more 
specific head noun? E.g., is it better to say the 
small dog or the chihuahua? 
3. Should relative or absolute adjectives be used? 
E.g., is it better to say the small dog or the one 
foot high dog? 
In our analysis, we observed several phenomena 
which we believe may generalise to other situations 
involving spoken, face-tooface language: 
1. Human speakers prefer to use adjectives that com- 
municate size, shape, or colour in referring expres¢ 
sions. In tile above examples, for instance, a hu- 
man speaker would probably prefer the black dog 
and the small dog over the female dog. 
2. Human hearers sometimes have trouble determin- 
ing if an object belongs to a specialized class. In 
the above example, for instance, the chihuahua 
should only be used if the speaker is certain 
the hearer is capable of distinguishing chihuahuas 
from other types of dogs. If there is any doubt 
about the heater's ability to do this, adding an 
explicit modifier (e.g., the small dog) is a better 
strategy than using a specialized head noun. 
3. Human speakers seem to prefer to use relative ad- 
jectives, and human hearers seem to have less trou- 
ble understanding them. However, human-written 
instructional texts sometimes use absolute adjec- 
tives instead of relative ones; this may be a con- 
sequence of the fact that writers cannot predict 
the context their text will be read in, and hence 
how readers will interpret relative adjectives. In 
the above example, therefore, a speaker would be 
expected to use the small dog, but a writer might 
use the one foot high dog. 
ZThe transcript was made by Phil Agre and John 
Batali, from a videotape taken by Candy Sidser. We are 
very grateful to them for allowing us to use it. 
ACTF~ DE COLING-92, NANqT~, 23-28 Aotrr 1992 2 3 4 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
The Algorithm 
Based on the above considerations, we have created 
a new algorithm for generating referring expressions. 
This algorithm is simpler and faster than the algo- 
rithms proposed in \[Dai89,Rei90a\] because it per- 
forms much less length-oriented optimization of its 
outputi we now believe that the level of optimiza- 
tion suggested in \[Da189,ReigOa\] was unnecessary 
and psycholinguistically implausible. The algorithm 
has been implemented as part of a larger natural- 
language generation system, and we are pleased with 
its performance to date. 
Assumptions about the Knowledge Base 
Our algorithm is intended to be reasonably domain- 
independent. We. do, however, make some assump- 
tions about the structure of the host system's un- 
derlying knowledge base, and require that certain 
interface functions be provided. 
in particular, we assume that: 
• Every entity is characterised in terms of a col- 
lection of attributes and their values. An 
attrii)ute-value pair is what is sometimes thought 
of as a property; an example is (colour, red). 
Every entity has as one of its attributes some 
type. This is a special attribute that corresponds 
to the kinds of properties that are typically real- 
iT, ed by head nouns; an example is (type, dog). 
• The knowledge base may organize some attribute 
values in a subsumption taxonomy (e.g., as is done 
in KI:ONE \[BS85\] and related KR systems). Such 
a taxonomy might record, for example, that an- 
imM subsumes dog, and that red subsumes scar- 
let. For such taxonomically-organized values, 
the knowledge-base or an associated user-model 
should specify which level of the taxonomy is 
basic-level for the current user. 
We require that the following interface functions be 
provided: 
value(object,attribute) returns the value (if any) that 
an attribute has for a particular object. Value 
should return the most specific possible value for 
this attribute, e.g., chihuahua instead of dog, and 
scarlet instead of red. 
taxonomy-children(value) returns the immediate chil- 
dren of a value in the taxonomy. For example, 
taxonomy-children(animal) might be the set {dog, 
cat, horse, ... }. 
basle-level-value(object.attribute) returns the basic~ 
level value of an attribute of an object. FOr exam- 
ple, basic-level-value(Garfield, type) might be cat. 
The knowledge-representation system should in 
principle allow different basic-level classes to be 
specified for different users \[Ros78,Rei91\]. 
user-knows(object, attribute-value-pair) returns true 
if the user knows or can easily determine (e.g., by 
direct visual perception) that the attribute-valuc 
pair applies to the object; false if the user knows or 
can easily determine that the attribute-value pair 
does not apply to the object; and unknown other- 
wise. FOr exmnple, if object x had the attribute- 
value pair (type, chihuahua), and the user was ca- 
pable of distinguishing dogs from eats, then user- 
knows(x, (type, dog)) would be true, while user- 
knows(x, (type, cat)) would be false. If the user 
was not, however, capable of distinguishing differ- 
ent breeds of dogs, and had no prior knowledge 
of x's breed, then user-knows(x, (type, chihuahua)) 
and user~knows(x, (type, poodle)) would both re- 
turn unknown, since the user would not know or 
be able to easily determine whether x was a chi- 
huahua, poodle, or some other breed of dog. 
Finally, we a~ume that the global variable 
*p~eferred-attributes* lists the attributes that human 
speakers and hearers prefer (e.g., type, size, shape, 
and colour in the ~.,~embly task transcript mentioned 
above). These attributes should be listed in order of 
preference, with the most preferable attribute flint. 
The elements of this list and their order will vary 
with the domain, a~ld should be determined by em- 
pirical inv~tigation. 
Inputs to the Algorithm 
In order to construct a reference to a particular em 
tity, tile host system must provide: 
- a symbol corresponding to the intended referent; 
and 
• a list of symbols correspondiug to the members of 
the contrast set (i.e., the other entities in focus, 
besides the intended referent). 
The algorithm returus a list of attribute-value pairs 
that correspond to tim romantic content of the refer- 
ring expression to be realized. This list can then be 
converted into an SPL \[K&~9\] term, as is done in the 
II)AS implementation; it can also be converted into a 
recoverable semantic structure of the kind used 
in Daie's EPICOltE system \[Da188,Dai89\]. 
The Algorithm 
In general terms, the algorithm iterates through 
the attributes in *preferred-attributes*. For each at- 
tribute, it checks if specifying a value for it would 
rule out at least one member of the contrast set that 
has not already becu ruled out; if so, this attribute is 
added to the referring ~t, with a value that is known 
to the User, rules out as many contrast set mem- 
bers as possible, and, subject to these constraints, 
is as cl(~e as possible to the basic-level value. The 
process of adding attribut~value pairs continues mt- 
til a referring expression has been formed that rules 
out every member of the contrast set. There is no 
backtracking; once an attribute-value pair has been 
added to the referring expression, it is not removed 
even if the addition of subsequent attribute-value 
pairs make it unnecessary. A head noun (i.e., a value 
for tim type attribute) is always included, even if it 
Acres DE COLING-92, NANTES, 23-28 AO~n' 1992 2 3 5 PROC. OV COTING-92, NANTES, AUG. 23-28, 1992 
l make-referring-expression(r, C,P) I 
L*-- {} 
D,-C 
for each member A~ of list P do 
V = flnd-best-value(A~, baslc-level-value(r, A~)) 
IfV ~ nil A rules-out((A~, V)) ~ nil 
then L ~ L U {(AI, V)} 
D ~ D - rules-out((At, V)) 
endlf 
If D = {} then 
if (type, X) (: L for some X 
then return L 
else return L U {(type, basic-level-value(r, type))} 
endif 
endif 
next 
return failure 
I find-best-value(A, initial-valse) l 
ff user-knows(r, (A. initial-value)) = true 
then value ~-- initial-value 
else value ~ nil 
vndlf 
for v~ E taxonomy-children(initial-value) 
lfv~ subsumes value(r, A) A 
(new-value ~ find-best-value(A, vi)) ~ nil A 
(value = nll Y 
Irules-out((A, new-value)) I > Irules-out((a, valse) )l) 
then value ~ new-value 
endif 
next 
return value 
\[ ,ul;s-out(<A, v>)\[ 
return {x : x E D A user-knows(x, (A, V)) = false} 
Figure 1: The Algorithm 
has no discriminatory power (in which ease the basic 
level value is used); other attribute values are only 
included if, at the time they were under considera- 
tion, they had some discriminatory power. 
More precisely, the algorithm is as shown in Figure 1. 
Here, r is the intended referent, C is the contrast set, 
P is the list of preferred attributes, D is the set of 
distractom (contrast set members) that have not yet 
been ruled out, and L is the list of attribute-value 
pairs returned, a 
make-referring-expression is the top level function. 
This returns a list of attribute-value pairs that 
specify a referring expression for the intended ref- 
a For simplicity of expo6ition, the algorithm as described 
here returns failure if it is not pesaible to rule out all the 
mernbem of the contrast set. A more robust algorithm 
might attempt to pur~m other strategies here, e.g, gen- 
erating a referring expression of the form one of the Xs, 
or modifying the contrast set by adding navigation infor- 
mation (navigation is discussed in the section on Future 
Work). 
erent. Note that the attributes are tried in the or- 
der specified in the *preferred-attributes* list, and 
that a value for type is always included, even if 
type has no discriminatory power. 
find-best-value takes an attribute and an initial 
value; it returns a value for that attribute that is 
subsumed by the initial value, accurately describes 
the intended referent (i.e., subsumes the value the 
intended referent possesses for the attribute), rules 
out as many distractors as possible, and, subject 
to these constraints, is as close as possible in the 
taxonomy to the initial value. 
rules-out takes an attribut~.~value pair and returns 
the elements of the set of remaining distractom 
that are ruled out by this attribute-value pair. 
An Example 
Assume the task is to create a referring expression 
for Objectl in a context that also includes Object2 
and Object3: 
• Object1: (type, chihuahua), (size, small), (calour, 
black) 
• Object2: (type, chihuahua), (size, large), (colour, 
white) 
* Object3: (type, siamese-cat), (size, small), (colour, 
black) 
In other words, r = 0bjectl and (7 = {Objest2, 
Object3}. Assume that P = {type, colour, size, ... }. 
When make-referring-expression is called in this con- 
text, it initializes L to the empty set and D to C, i.e., 
to {Object2, Object3}. Find-best-value is then caUed 
with A = type, and initial-value set to the basic-level 
type of Object1, which, let us assume, is dog. 
Assume user-knows(Object1, (type, dog)) is true, i.e., 
the user knows or can easily perceive that Objectl 
is a dog. Find-best-value then sets value to dog, 
and examines the taxonomic descendants of dog to 
see if any of them are accurate descriptions of Ob- 
ject1 (this is the subsumption test) and rule out 
more distractors than dog does. In this case, the 
only accurate child of dog is chihuahua, but (type, 
chihuahua) does not have more discriminatory power 
than (type, dog) (both rule out {Object3}), so find- 
best-value returns dog as the best value for the type 
attribute. Make-referring-expression then verifies that 
(type. dog) rules out at least one distraetor, and 
therefore adds this attribute-value pair to L, while 
removing rules-out((type, dog)) = {Object3} from D. 
This means that the only remaining distraetor in 
D is Object2. Make-referring-expression (after cheek- 
ing that D is not empty) calls find-best-value again 
with A = colour (the second member of P). Find- 
best-value returns Objectl~s basic-level colour value, 
which is black, since no more specific colour term 
has more discriminatory power. Make-referring- 
expression then adds (colour, black) to L and removes 
rules-out((colour, black)) = {Object2} from D. D 
is then empty, so the generation task is completed, 
ACTES DE COLING-92, NANTES, 23-28 AO~r 1992 2 3 6 PROC. OF COLING-92, NArCrEs. AUG. 23-28. 1992 
and make-referring-expression returns {(type, dog), 
(celour, black)} , i.e., a specification for the refer- 
ring expression the black day. Note that if P had 
been {type, size, colour, ... } instead of {type, cnlour, 
size, ...}, make-referring-expeession would have rc~ 
turned {(type, dog), (size, small)} instead, i.e., the 
sraall do#. 
Implementation 
The algorithm is currently being used within the 
n)AS system \[RML92\]. ll)hS is a natural lm~guage 
generation system that generates on-line documen- 
tation and help texts from a domain arid linguistic 
knowledge base, lining user expertise models, user 
task models, and discourse models. 
IDAS uses a KL-ONE type knowledge repr~entation 
system, with roles corresponding to attributes and 
lillem to values. The type attribute is implicit in 
the position of an object in the taxonomy, and is 
not explicitly represented. The value and taxonomy- 
children functions are defined in terms of standard 
knowledge-base access functions. 
A knowledg~base author can specify explicit basic- 
level attribute values in IDAS user models, but IDAS is 
also capable of using heuristics to guess which value 
is basic-level. The heuristics are fairly simple (e.g., 
"nse the most general value that is not in the upper- 
model \[BKMW90\] and has a one-word realization"), 
but they seem (so far) to be at least somewhat effec- 
tive. A *preferred-attributes* list has been crcated 
for IOAS's domain (complex electronic machinery) 
by visual inspection of the equipment being docu- 
mented; its first members are type, colour, and la- 
bel. The user-knows function simply returns true if 
the attributc~value pair is accurate and false other- 
wise; this essentially assumes that the user can visu- 
ally perceive the value of any attribute in *preferred- 
attributes*, which may not tie true in general. 
The referring expression generation model seems rea- 
sonably successful in IDAS. In parLieular, the algo- 
rithm lure proven to be useful because: 
1. It is fast. The algorithm runs in linear time in 
the number of distractors, which is probably im- 
possible for any algorithm that includes an ex- 
plicit brevity requirement (e.g., the algorithms of 
\[Da189,Rei90a\]). Of equal importance, its run- 
time is independent of the number of potential 
attributes that could be used in the referring ex- 
preszion. This ks a consequence of the fact that 
the algorithm does not attempt to find the at- 
tribute with the highest discriminatory power, but 
rather simply takes attributes from the *preferred- 
attributes* list until it has built a successful refer- 
ring expression. 
2. It allows human preferences and capabilities to 
be taken into consideration. The *preferred- 
attributes* list, the preference for basic-level val- 
ues, and the user~knows function are all ways of 
biasing the algorithm towards generating referring 
expressions that use attributes and values that hu- 
taan hearers, with all their perceptual limitations, 
lind easy to process. 
Almost all referring expressions generated by IDAS 
contain a head noun and zero, one, or perhaps at 
most two modifiers; longer referring expressions are 
rare. The most important task of the algorithm is 
therefore to quickly generate easy-to-understand re- 
ferring expre~mions in such simple cases; optimal han- 
dling of more complex referring expressions is lees 
important, although the algorithm should be robuat 
enough to generate something plausible if a long re- 
ferring expression is needed. 
Future Work 
Navigation 
As mentioned in the introduction, the algorithm pre- 
sented here assumes that the intended referent is in 
the context set. An important question we need to 
address is what action should be taken if this is not 
the c~.se, i.e., if the intended referent is not in the 
current focus of attention. 
Unfortunately, we have very little data available ou 
which to bose a model of the generation of such refer- 
ring expressions. Psyclmlinguistic researchers seem 
to have paid relatively little attention to such eases, 
and the transcripts we have (to date) examined have 
contained relatively few instances where the intended 
referent was not already salient. 
ltowever, we take the view that, in the general case, 
a referring expression contains two kinds of informa- 
tion: navigation and discrimination. Each de~ 
scriptor used in a referring expression plays one of 
these two roles. 
• Navigational, or attention-directing informa- 
tion, is intended to bring the intended referent into 
the hearer's focus of attention. 
• Discrimination information is intended to distin- 
guish the intended referent from other objects in 
the hearer's focus of attention; such information 
has been the subject of this paper. 
Navigational information is not needed if the in- 
tended referent is already in the focus of attention. 
If it is needed, it frequently (although not always) 
takes the form of loeational information. The IDAS 
system, for example, can generate referring expres- 
sions such as tl~e black power supply in the equipment 
rack. In this case, in the equipment rack is navigation 
information that is intended to bring the equipment 
rack and its components into the hearer's focus of 
attention, while black power supply is discrimination 
information that is intended Ix) distinguish the in- 
tended referent from other members of the context 
~t (e.g., the white power supply that is also present 
in the equipment rack). 
The navigation model currently implemented in If)AS 
is simplistic and not theoretically well-justified. We 
hope to do further research on building a better- 
ju~stified model of navigation. 
AcrEs DE COLING-92, NAbrlns. 23-28 ^o~r 1992 2 3 7 PRoc, OF COLING-92, NANTES, AtJcl. 23-28, 1992 
Relative Attribute Values 
As mentioned previously, the transcript analysis 
shows that human speakers and hearers often pre- 
fer relative instead of absolute attribute values, e.g., 
small instead of one inch. Knowledge bases some- 
times explicitly encode relative attribute values (e.g., 
(size. small)), but this can cause difficulties when re- 
ferring expressions need to he generated in different 
contexts; a one-inch screw, for example, might be 
considered to be small in a context where the other 
screws were all two-inch screws, but large in a context 
where the other screws were all half-inch screws. 
A better solution is for the knowledge base to record 
absolute attribute values, and then for the genera- 
tion algorithm to automatically convert absolute val- 
ues to relative values, depending on the values that 
other members of the context set pussc~ for this at- 
tribute. Thus, the knowledge base might record that 
a particular screw had (size. one-inch), and the gen- 
eration system would choose to call this screw small 
or/arye depending on the size of the other screws in 
the context set. We hope to do further research on 
determining how exactly this process should work. 
Conclusions 
We have presented an algorithm for the generation 
of referring expressions that is substantially simpler 
and faster than the algorithms we have proposed 
in previous work \[Da189,Rei90a\], largely because it 
performs much less length-oriented optimization of 
its output. We have been guided in this simplifica- 
tion effort by psycholinguistic findings and transcript 
analyses, and believe that the resulting algorithm is 
a more practical one for natural language generation 
systems than the ones we proposed previously. 
ACIds DE COLING-92, NANI~S, 23-28 AOUT 1992 2 3 8 PROC. OF COL1NG-92, NANI'ES, AUG. 23-28, 1992 

References 

lapp85\] Douglas E Appelt. Planning English Sentences. 
Cambridge Univemity Press, New York, 1985. 

\[BKMWgO\] John Bateman, Robert T Kasper, Johanna 
D Moore and Richard A Whitney. A General 
Organization of Knowledge for Natural Language 
Processing:. The Penman Upper Model. Unpub- 
lished technical report, Information Sciences Insti- 
tute/University of Southern California, 1990. 

\[BS85\] Ronald Brachman and James Schmolze. An 
overview of the KL-ONE knowledge representation 
system. Cognitive Science 9:171-216, 1985. 

\[Cru77\] D. Cruse. The pragmatics of lexieal specificity. 
Journal of Linguistics , 13:153-164~ 1977. 

\[Da188\] Robert Dale. Generating Referring Expressions in 
a Domain of Objects and Processes. PhD Thesis, 
Centre for Cognitive Science, University of Edin- 
burgh, 1988. 

IDol89\] Robert Dale. Cooking up referring expre~ens. 
In Proceedings of the $7th Annual Meeting of the 
Association for Computational Linguistics, pages 
68-75. 1989. 

\[DH91\] Robert Dale and Nicholas Haddock. Content de- 
termination in the generation of referring expres- 
sions. Computational Intelligence, 7(4), 1991. 

\[Don66\] Kelth Donnellan. Reference and definite descrip- 
tion. Philosophical Review, 75:281-304, 1986. 

\[17075\] William Ford and David Olsea The elaboration 
of the noun phrase in children's de~riptton of ob- 
jects. Journal of Ezpemmentol Child Psychology, 
19:371-382, 1975. 

\[GJW83\] Barbara Gresz, Aravind Jeshi, and Scott Wein- 
stein. Providing a unified account of definite noun 
phrasesin discourse. In Proceedings of the £1st An- 
nual Meeting of the A ssoeiation for Computafional 
Linguistics, pages 44-50. 1983. 

\[Gri75\] H. Paul Grice. Logic and conversation. In P. Cole 
and J. Morgun~ editors, Syntax and Semantics: Vat 
3, Speech Acts, pages 43-58. Academic Prese, New 
York, 1975. 

\[GS86\] Barbara Grenz and Candaen Siduer. Attentlon~ 
intention, and the structure of discourse. Compu- 
tatioual Linguistic, 12:175-208, 1986. 

\[Kns89\] Robert Kasper. A flexible interface for linking 
applications to Penman's sentence generator. Pro- 
ceedings of the 1989 DARPA Speech and Natural 
Language Workshop~ pages 153-158. 

\[Kro86\] Amichai Kronfeid. Donnsllan's distinction and a 
computational model of reference. In Proceedings 
of the ~.~th Annual Meeting of the Association for 
Computational Linguistics, pages 185-191. 1986. 

\[Lev89\] Willera Levelt. Speaking: bq~om Intention to Ar- 
ticulation. MIT Pre~s, 1989. 

\[Pec89\] Thomas Pechmann. Incremental speech produc- 
tion and referential overspeeificatlon. Linguistics, 
27:89-110, 1989. 

\[Rdg0a\] Ehud Reiter. The computational complexity of 
avoiding conversational implicatures. In Pea~.2A- 
ings of the ~8th Annual Meeting of the Association 
for Computational Linguistics, pages 97-104. 1990. 

\[Relg0b\] Ehud Reiter. Generating Appropriate Natural 
Language Object Descriptions. Phi3 thesis, Aiken 
Computation Lab, Harvard Univeraity, 1990. Also 
available as Aiken Computation Lab technical re- 
port TK-10-go. 

\[Rei91\] Ehud Reiter. A new model of lexical choice for 
nouns. Computational Intelligence, 7(4), 1991. 

\[RML92\] Ehod Relter, Chris Mellish, and John Levine. 
Automatic generation of on-line documentation In 
tbe IDAS project. In Proceedings of the Third Cor~ 
ferenee on Applied Natural Language Pn~.p_.~sing, 
pages 64-71. 1992. 

IRes78\] Eleanor Rcech. Principles of categorization. In 
E. Resc~ and B. Lloydj editors, Cognition and Cat- 
egorization, pages 27-48. Lawrence Erlbaum, Hills- 
dale, NJ, 1978. 

\[Sid81\] Canda~e Sidner. Focusing in the comprehension of 
definite anaphora. In M. Brady and R. Berwick, 
editors, Computational Models of Discourse, pages 
267-330. MIT Press, Cambridge, Mass, 1981. 

\[Son85\] Susan Sonnenschein. The development of rder- 
ential communication skills: Some situations in 
which speakers give redundant messages. Journal 
of Paycholingnistic Research, 14:489-508, 1985. 

\[Wh176\] Graver Whitehurst. The development of commu- 
nicatiea: Changes with age and modeling. Child 
Development, 47:473-482, 1876. 
