Lexical cltoice in context: generating procedural texts 
Agn6s Tntin l, Richard Kitu-edge 
l)6partetnent tie lingnislique, 
Universit6 de Montrdal 
C.I'. 6128, Sacc "A", 
Montrdal P.Q. If3C 3J7 Canada 
Abstract 
This paper shows how lexical choice during text generation 
depemls on linguistic context. We argue that muking c(Irrect 
lexieal choice in rite textual context requires distinguishing 
properties of concepts, which are ntme or less independent of 
file language, from language-specific representations of text 
where lexemes and their semantic and symantic relations are 
represented. In particular, l.exical Fnnctions are well-suited to 
tormalizing anaphoric lexical links in text, including the 
introduction of superordinates. This sheds new light on the 
notion of "basic level", which has recently been applied to 
lexieM selection in genaration. Some consh'aints governing rite 
generation of lexical nnd grammatical anaphora are proposed 
for procedural text, using examples front tim sublanguage of 
recipes. 
O. Introduction 
Lexical choice cannot be made dining text generation 
wifltottt taking into account the linguistic context, both 
the lexical context of inmmdiately surrounding words 
and rite larger textual context. 
a) Lexical context consists of the words (or lather, the 
lexicnl specifications of nascent words) that enter into 
syntactic relations with the lexical item being generated. 
This intrn-clausal context is crncial lor lk)rmnlating 
collocational constraints, which restrict ways of 
expressing a preci~ meaning to certain lexical items, lot 
example as in expressions like pay attention, receive 
attention or narrow escape. The importance of 
eollocational constraiuts hlts IR3en emphasized in the 
literature on text generation aud inachine translation 
(Bateman & Wanner 1990, lordanskaja et al. 1991, 
Nirenbnrg & Nirenbnrg 1988, Held & Raab 1989). 
b) Textual conlext consists of the linguistic content of 
previous and subsequent clauses. This context is tile 
scope for cohesive links (ttalliday & Hasan 1976) with 
the lexical items m be generated in the current clause. 
The gn'eat majority of cohesivc links are anaphoric in 
nature 2. A textnal element T is an attaphor with reslmct 
to an antecexlent A (previtmsly introdacexl in the text) if 
the semantic or referential interpretation of T ~ depends 
on the interpretation of A. When genmating auallhttrs, it 
is Iberefot~ the previotts context that ntust bc taken inlo 
acconn\[, ItS ill: 
I) Now in Centre National d'Etudes des T61dcommunications, 
LAA/SLC\]AIA, Route de Trdgnstel BP 40 22301 l..unifion 
Cddex, Frmice. 
2) In the case of ca!alphorn, as in rite following sentence, 
If you want to, peel and chop the lualatoes. 
tile subsequent context mr,st be taken into account. 
3) Reference of a textual elemeut is file association betwean n 
textual eleme, nt and extra-linguistic reality. 
(1) ihel)am tile carrots, the celery mid the asparagus. 
Cook tile vegetables in I~iling wat,~" for ten minutes. 
Two textual clement are coreferential if they t-eli:r to the 
same exlralinguistic reality. Coreierential elentenls in 
our examples are often written ill italics, or indicated by 
identical snbsctipts. 
Faihue Rt ch~.R)se all appropriate anaphoric expression 
daring generation typically lcads to awkward or 
nnacceptable text such as (2): 
(2) a. l:'repme the t:aTmts, the celery and the asparagus. 
b. Cook tile caners, the celcly and tile aslmragos in N)iling 
wateL 
c. Take tile carrots, tile celery and the asparagus otlt at'tel 
tell ninltltes. 
Ill this paper, we examine the mechanisms required 
fox' making natural lexical choice as a lunetion of 
preceding text tlud its reference to extralingttistic objtx:ls 
or concepts. In particular, we are interested ill lexical 
anal)hera, wllere open-class lexical items or exl)ressions 
provide a coroference link to one or nlore snch items ill 
preceding clanses. For example, ill (1) vegetables is a 
lexicat corefercntial anaphor of the carrots, the celery 
and the asparagus.. 
ht what fallows, we ainl to show that correct 
lexicalization in context reqnires access to both the 
ctntcepnml relmence and the linguistic prolmrties of 
preceding text. For it pipelined generation architcctmc 
which \[naps froln abstract representation levels towards 
text, this implies distinguishing at conceptual level, more 
or less independent of the language, ti'oln language- 
specific representation levels which encode lexemes and 
the grammatical relation between thenl. In particular, we 
illustrate the paradignlatic Lexical l'unctkms (hereafter 
l,Fs) of Mel'cnk's Explanatory Combinatorial 
Dictionary (hereafter F, CD) (Mel%nk et al. 1988; 
Mel',Suk & Polgn~re 1987). 
1. Varieties of lexical anapbora 
Beft)re reviewing conshaints on tile introduction of 
lexical anal)hera during generation, we give examples of 
inlportant tYlmS of cm'eferential anaphoric links 4. 
We consider an anallher 10 13(3 lexical only if we can 
establish a semantic link between the attaphor and its 
antecedent. Therefore, in the following example: 
(3) Edith Cresson arrived Monday ut 9:00. At 11:00, the 
Prbne Miniater of France gave a pxess conference. 
4) We will not treat here non -comlErantial anaphora like: 
Marie threw away all her old dre~es because she wanted 
to buy lleV¢ Ollt2S. 
ActTzs DE COLING-92, NANTt'S, 23-28 Ao~r 1992 7 6 3 l'rttx:. OF COLIN(;-92, NAt, arES, AU(;. 23-28, 1992 
Prime Minister of France is a cognitive coreferential 
anaphor of Edith Cresson, but not a lexical one because 
the coreferential link between the two phrases is based 
on world knowledge, and not on linguistic semantics. 
One type of coreferential lexical anaphora is called 
"reiteration" by Halliday & Hasan (1976) with three 
subtypes: exact repetition, illusWated by (4b), synonym 
substitution (4b'), and superordinate substitution (4b"). 
We can add to this group partial repetition (4b"'): 
(4) a. 1 bottle of light red bordeaux. 
b. Pour the light red bordeaux on the meat. 
b'. Pour the lieht red bordeaux wine on the meat. 
b". Pour the ~ oct the meat. 
b'". Pour the ~ on the meat. 
Nominalization provides another way of introducing a 
coreferential link to a previous verb: 
(5) Cook the rabbit for two hours. 
Ten minutes before the end of cooking, add the spices. 
Coreferential lexical links can also be established 
between an action and its result. 
(6) Meanwhile. mix the egg yolki with the sugarj. 
Pour the milk on the mixturei+ j. 
In this example, mixture has no direct semantic link with 
its antecedents egg yolk and sugar. The link appears 
indirectly through the verb mix. 
Another type of lexical anaphora occurs with nouns 
denoting typical actants of an antecedent verb: 
(7) Marga i was lecturing to third year students k. The lecturer i 
was very interesting and the audience k quite attentive. 
In this case, lecturer is linked coreferentially with 
Marga because it is the "agent noun" of lecture, while 
audience is the corresponding "patient noun", and is 
coreferential with third year students. 
These examples illustrate some of the diversity of 
lexico-semantic resources needed to build coreferential 
links in text. Text generation therefore requires a lexicon 
which gives access to the full range of such resources 
from the "viewpoint" of the antecedent lexeme. As seen 
in the next section, LFs provide an appropriate access 
mechanism for choosing the correct anaphor. 
2. Lexical Functions of the ECD for creating 
lexical anaphora 
Lexical Functions of the ECD provide a formalism 
representing many common instances of coreferential 
anaphora. Formally defined, a Lexical Function f is a 
correspondence between a lexical item L, called the key 
word of f, and a set of lexical items f(L) - the values of f 
(Mel'~uk & Zholkovskij 1970, Mel'~uk 1988b). 
Approximately sixty standard Lexical Functions have 
been defined (for a recent description of LFs in the ECD 
in English, see Mel'~uk & Polgu~re 1987). They can be 
divided into two subsets: syntagmatic LFs and 
paradigmatic LFs. 
- Syntagmatic or collocational LFs are used to link 
unpredictable lexical cooccurrences in texts between the 
key word and its values through a specific semantic 
relation. Typical examples of syntagmatic LFs are Oper i 
(semantically empty verb which takes the i-th arrant 5 of 
the key word as its subject and the key word as its direct 
object), like Operl(attention) = pay, Oper2(attention) = 
receive or Magn(escape) = narrow. These examples 
show that these LFs convey eooccurrence relations. 
-Paradigmatic LFs are used to express semantic 
relations in the lexicon between the key word of the LF 
and its values, but not cooccurrence relations. Typical 
examples are Sl(lecture ) = lecturer (Sl:Noun of the 
first typical actant), Sloc (box) = ring (Sloc: Noun of 
typical place), So(buy) = purchase (So: Derived noun). 
Some paradigmatic LFs can be used to analyse or 
generate lexical coreference relations: 
- Syn: synonym 
- ConvUkl: conversive 
- aener: generic word 
- St: typical noun for the i-th actant 
- Slratr: noun for typical instrument 
- Sm~: noun for typical means 
- Site: noun for typical place 
- Sr~: noun for typical result 
- Stood: noun for typical mod~ 
- S a : name of action 
Syn(callhlg) = vocation 
Conv32 t 4(sell) = buy 
Gener(apple) = fruit 
SI (lecture) = lecturer 
Sinsn(palnt) = brush 
Smed(ltolsalt ) = salt 
Slo~(box) ~ ring 
S~(mix) = mixture 
Stood(write) = writing 
S0(buy ) = purchase 
Relations encoded by these LFs can appear in direct 
eoreferential relations in texts when the value of the 
function and the key word maintain a semantic 
relationship directly formalizable through a LF such as 
Sres, Gener, Syn and Convijkl, as in: 
(8) Gener(lamb) = meat 
Buy lamb. Be sure the meat is very fresh. 
LFs can be used to formalize indirect iexieal 
coreference when coreference exists between lexical 
items and a dependent. The dependent may be an actant 
as in (7) (lecturer, the Sl(lecture) is coreferential with 
the fn'st actant Marga of lecture whereas audience, the 
S2(lecture) is coreferential with the second actant of 
lecture), or an adverbial, as in the following example: 
(9) Sloc(patiner ) = patinoire 
Marguerite et Jean ant patin6 sur le canali Rideau. 
Cene patinoire i fait 8km de long. 
\[Marguerite end Jean skated on the Rid~au Canal. This 
"skating rink" is 8 km long.\] 
In (9), patinoire, Sloc(patiner) is coreferential with 
canal Rideau. 
Moreover, LFs can be combined, as we see in the 
following table: 
5)In the ECD, lecture will be described as a noun which has 
three syntactic actants: X's (actant I) lecture to Y (arrant II) on 
Z (arrant III), for example Jean's (actant I) lecture to third year 
students (actant II) on semantic causality (actant I1\]). 
ACRES nE COLING-92. NAm'ES, 23-28 AOfn" 1992 7 6 4 PROC. OF COLING-92, NXlSrrES, AUG. 23-28, 1992 
key wurd values LFs or 
eumposlthmnf Lb' 
Gener 
Gener 
Oeller 
Gener 
Gener 
S3m 
ConY3214 
SO 
SO 
Geuer o Gener 
GeIler o Syn 
Gener o Conv3214 
Gener o S O 
SO o Conv3124 
ConY3214 oS0 
achat \[lmrclu~se\] 
vente Isale\] 
transaction Ideal\] 
auto \[car\] 
voiture lear\] 
voiture lcarl 
acheter \[buy\] 
acheter \[buy\] 
vendre lsell\] 
achat \[purchase\] 
auto \[carl 
vente \[sale\] 
acheter \[buy\] 
achetcr \[buy\] 
acheter \[buy\] 
transaction \[deal\] 
transaction Ideal\] 
action \[action\] 
v6hicule \[vehicle\] 
v6hicule \[vehicle\] 
auto tear\] 
v enth'e \[sell\] 
achat \[purchase\] 
vente lsale\] 
action \[acthl(li 
vghicule \[vehicle\] 
trm~sactiou Ideal\] 
uausaction Ideal\] 
vente I sale\] 
vente \[s~de\] 
= 
Table 1: LFs aud compositions of LFs fltr direct 
coreference links 
The following facts should be notcxl about compositions: 
Composition is not commutative. Thus, 
So(Conv3214(aeheter)) = Conv3214(So(acheter)) = 
vente but Gener(So(acheter)) ~ S0(Gener(acheter)) 
because Gener(acheter) does not have a value. 
- Some compositions are reducible. For example, the LF 
Syn plays a transparent role in composition. 
In the perspective of text generation, this formalism 
appears very interesting for building coreferential 
expressions. To point back to a referent already 
introduced, LFs and compositions of I,Fs offer many 
possible ways for lexicalizing a given relerent. For 
example, let us suppo~ that after having introduced the 
following sentence, 
(10) a. L~fisser &uver laviande. \[Let the meat stema.\] 
we have to refer again to the action la viande Ftuve. We 
could try to use a noutinalization (So). But, as fltere is no 
nominalization for the verb Ftuver, we could use instead 
the nominalization of the generic term, 
So(Gener(dtuver)) = cuisa~n. We could thus produce the 
following sentence: 
(11 ) b. A la fin de la cuisson, ajouter les cpiees 
\[At tile end of cooking, add the spices\] 
In the next section, we will examine the case of a 
complex lexical anaphor: file superordilmte term. 
3. Superordinates and basic nouns 
The use of superordinate terms as anaphors raises 
several interesting questions. 
First, to the extent that a generic concept (for two or 
more specific concepts) has a simple expression in a 
language, this is not necessarily the same term as the 
superordinate term (for the term corresponding to the 
specific concepts). For example, from a conceptual point 
of view, knife and scissors are "cutting instruments". 
Nevertheless, it is not possible to naturally stthstitute 
cutting instrument for knife and scissors, as in: 
(12) a. Use a knife mid scissors to cut up the duck, 
b.? If you don't have these cutting instruments, pull the 
duck apart. 
ACIES DE COLING-92, NANTES, 23-28 AO~f 1992 7 6 5 
There is no consistently used term 12tr expressing tile 
generic concept of knife or scissors. ~lltis can be cutting 
instruments as well as instruments for cutting or cutting 
utensils. Whether or uot such a teNll exisLs varies among 
lauguages. For example, in Mandarin Chinese, file term 
d~o is fidly accepted as a superordinate term to point to 
the Chinese equivalents of knife aml scissors. Ill 
English, a term like vegetable is the superordiuate of 
carrot, tomalo Or cucumber because it is consistently 
used for tlte~ previous words in texts. 
This entails that choice of snperordinate terms as 
lexical auaphors cannot be made at tile conceptual level 
alone. 
Moreover, saperordinate terms call often bc: lnorc 
easily nsed to lexiealize reference to a uon-homegeueous 
set of elements than for reference to a single element or 
homogcueous ~t, as illustrated in (13) and (14): 
(13) a. Put tile carrots ill to boiling water. 
b. ? Remove the vegetables after l0 mimnes. 
(14) a. Throw the carrots, the leeks altd Ihe lx~tatoes in to 
boiling water. 
b. Remove tile vegetables after l0 minutes. 
However, the ease with which a snperordinale can be 
used depends ou rite particular noun. For example, in 
French, viande \[meatl can be snbsituted for bleuf Ibeefl 
even ill singular: 
(15) a. Menre le b~euf h cuire dmls l'eau I~ouillante. 
IPut the t~_ef in the I~iling waterl 
b. Retirer la viande ml bout de 20 minutes. 
\[Remove the meat aftc'r 20 minutes\] 
This somewhat suri)rising i)henomenon can be analysed 
with the help of the notion of basic level object proposed 
by Roseh et al. (1976). The imlx)rtance of the basic level 
distinction for text generation has recently been shown 
by Reiter (1990). Rosch et al. demonstrated that the 
taxonomy of concepts could be organized using a 
stnlcture with three levels: superordiuate, haste and 
subordinate. They define the basic level as follows: 
"basic objects a~e the most inclusive categories whose 
members: (a) possess significant numbers of attributes in 
cmmnon, (b) have motor programs which are similar to 
one another, (c) have similar shapes, and (d) can be 
identified from averaged shapes of members of the 
class" (Rosch et al. 1976: 382) 
It has been shown that lexemes correspomling to basic 
level objects seem to be the most natural terms to 
introduce referents already idcntified. For example, if 
one wants to refer to some champignons de Paris 
\[button mushrootos\], one would prctk:r to call them 
champignons \[nmshrooms\], provided that there is no 
potcntial amttiguity with auy other mushrooms. 
Champignons de Paris would sccm too specific in this 
context and vegetables would seem too vague. This 
choice is not made randomly: champignon is the noun 
corresponding to the highest basic level coucept to 
designate these objects. This would explain why in (15), 
one can refer to b~eufwith the superordinate viande. 
Nevertheless, rite notion of basic level object does not 
always seem well suited to explain phenomena such as 
that observed in (15). For example, it seems that the 
concept "volatile" \["fowl"\] fits perfectly the four criteria 
given by Rosch. But, volatile \[fowll does not seem a 
PROC. ol; C{1LIN(;-92, NANTES, AUG. 23-28, 1992 
natural French term for referring to a chicken, 
particularly in the sublangtmge of recitees. 
It is also problematic that the naming of basic level 
objects varies a great deal among languages. For 
example, in Mandarin Chinese, the most natural term to 
designate a knife when there is no ambiguity is the term 
d~o , which corresponds to "cutting instrument" in 
English. We could argue that conceptual representation 
differs with the mother tongue of the speaker (which is 
plausible, without joining the debate about language and 
thought) and that the lexicon reflects the conceptual 
views. Nevertheless, this position does not solve the 
problem of terms like volaille, a unuatural term for a 
basic level object. 
It is significant that this position creates practical 
problems for text generation: if conceptual 
representation is reflected too closely in the choice of 
lexemes, this representation cannot be used as an 
interliugua for multilingual generation or machine 
translation. 
In the light of this evidence, we have decidcd in favor 
of a strict theoretical separation between conceptual 
representation and lexical representation. We believe 
that an appropriate couceptual representation can be 
used for multilingual generation because it is a non 
linguistic generalization above specific lexical 
representations. We therefore distinguish the notion of 
basic level object, which belongs to cognitive science, 
from the notion of basic noun, which is a linguistic 
notion 6. We consider "viande" and "volatile" to be basic 
level objects while only viande is a basic noun. 
For lexical choice in text generation, we thus have to 
distinguish two very different processes: 
- Superordination should be used to introduce a noun 
which points back to a set of different nouns. This is the 
case in {carrots, leeks, cucumber} --> vegetables. This 
process obeys a principle of economy. 
- Basic denomination is used to introduce the most 
natural term for a given referent or a set of referents. 
This process obeys a principle of "naturalness": it 
introduces the most closely basic noun that corresponds 
to the concept to be lexicalized. Basic denomination is 
often used in texts like recipes: objects are first 
introduced with extreme precision and subsequently 
referred to with the basic term. 
4. Knowledge sources for determining 
lexical anaphors 
In the course of our work, we have proposed a series 
of algorithms for generating grammatical and lexical 
anaphora in procedural texts (Turin 1992). Contrary to 
lexical anaphora, grammatical anaphora makes use of 
closed lexical classes (determiners, pronouns and a few 
special verbs) as well as ellipsis. 
These algorithms are derived from an empirical study 
of French recipes, using a representative corpus of over 
16,000 words. Recipes serve as a good prototype of 
procedural texts for assembling complex objects from 
parts. Even this modest corpus presents a wide variety of 
lexical and grammatical anaphora which are typical of 
assembly instructions. 
6) Wierzbicka (1985) has shown in lexicographic descriptions 
that the nantes of (words for) basic level objects have special 
semantic properties., 
We describe below some of the knowledge sources 
and organization needed to generate lexical and 
grammatical anaphora. For lack of space, however, we 
leave out the model of state change management (needed 
to describe recipe ingredients being mixed together and 
transformed (Kosseim 1992)), a~ld the focus model used. 
4.1 Input 
We limit our scope to the linguistic part of generation; 
therefore, we assume that onr input is the ontput of a text 
planner, which has already grouped actions into 
discourse structures as proposed by Grosz and Sidner 
(1986) and (Dale 1988), The input is thus a sequence of 
actions and states in which participants (ingredients, 
instruments and agent) are represented by indices. 
4.2 Dictionary of concepts 
The dictionary of concepts has been inspired by 
Nirenburg and Raskin 1987; concepts are mainly 
subdivided into actions or objects. We have added a 
category of properties, needed to describe relations 
between concepts (e.g., temporal limit) or attributes (e.g. 
size). 
Relations between concepts are isa, part-o for result, 
the latter one useftd in a domain where state changes are 
frequent. Thus, one can relate the action "cut" to the 
concept "piece" which is the result of "cut". The 
dictiouary of concepts is not a copy of the language and 
there are Concepts without any corresponding 
lexicalization. Taxonomic organization is fimetional and 
depends greatly on the field for which it has been 
established. In other words, our description of concepts 
has limited value outside the domain of recipes. 
4.3 Dictionary of lexical entries 
The representation of lexical entries is strongly 
infhlenced by tile ECD (Mel'~uk & Folgu&e 1987, 
Mel'~uk et al. 1988). Two parts of tile entry are 
particularly interesting for our topic: the semantic zone 
and the LF zone. 
The semantic zone contains four types of information: 
- Tile semantic field to which the lexeme belongs. For 
example, the verb simmer would have feature/cook/. 
- The mass/count feature. 
- The "basicness" feature, if the lexical item is a noun, 
indicates whether or not the noun is a basic noun. 
- The key word(s) for which the lexeme can be a value. 
For example, for the lexeme mixture, it will be stated 
that it is the Sres of mix. 
In the LF :,.one, we simply enumerate the values of the 
lexical item as a key word. For example, the entry for 
the verb hacher \[chop\] may contain, among many 
others, hachis (Sres(hacher)) and hachoir 
(Smed(hacher) ). 
5, Constraints for generating anaphors 
We now turn to the constraints which apply to the 
choice of grammatical or lexical anaphors during text 
generation. Our aim here is to generate the most 
appropriate anaphor with respect to the textual context. 
To determine what is appropriate, we have used an 
empirical approach, rather than appeal to general 
principles such as Gricean conversational maxims (see 
Reiter 1990a & Dale 1988 for use of these notions for 
lexical choice in text generation). A detailed 
ACRES DE COLING-92, NANTES, 23-28 not'rr 1992 7 6 6 Paoc. OF COLING-92, NANTES, AUG. 23-28, 1992 
examinatiou of our corpus of cooking recipes has sttowu 
that anaphora is not governed so muct, by sU'ict rnlcs as 
by tendencies. Thus, in a given context, a set of possible 
anaphors can "compete" for selection. When choosing 
from multiple possibilities we flavor ttte most 
"economical" anaphor, i.e., the due which conveys the 
least information 7. 
Space limitations prevent a complete discussion of all 
factors required for an anaphor choice algorithm (see 
Tutin 1992). Here we give the most ilnporfant 
constrai,~ts on choice among the principal anaphoric 
devices s. 
The selection of an anaphoric device Ires two slages: 
• First, at choice is made among of grammatical devices 
(e.g. personal pronoun, verb complement ellipsis, 
coreferential definite NP, demonstrative NP). 
• Then, if a lexical NP has been chosen, the corrcct 
lexical eataphor is determined. 
5.1 Grammatical anaphnrn 
The introduction of a given grammatical anaphor 
depends mainly on 4 kinds of paranmtcrs: a) the 
conceptual nature of referents, b) distance In antecedent 
and discourse structure, c) focaliz~'~tion and d) potential 
ambiguity. 
We briefly review these different parameters tor each 
type of grammatical anaphor: verbal complement 
ellipsis, persomtl prolmml, demonstrative NP, 
corefereutial definite NP. 
Verbal complement ellipsis as in the following 
exaraple is very widespread in recipes, and characteristic 
of procedural instructions in general. 
(16) Prepare the carrots, the celery and the asparagus. Cook 
in the boiling water and take O out after 10 minutes. 
Verbal complement ellipsis is generally used to 
designate a heterogeneous .set of objects, coutrary to 
personal pronouns. "llm distance from the antecedent can 
be quite far bat focalization coustraints, in particular 
global focus - defined as the subset of the most salient 
items - play a determining role for the production of this 
anaphor. 
A personal prononn must nalne an object or a set of 
similar objects. It is governed by very strong locality 
constraints (Hobbs 1978) and, as previously noted ill the 
literature, personal pronouns often mainlain the tlmntatic 
continuity (Van Dijk & Kintseh 1983), i.e. pronoun is 
the local focus (what the clause is about) of both the 
previous and the current clauses. In fact, local focus 
generally supplies enough information for the healrer to 
correctly interpret the pronoun (as emphasized by Grosz, 
Joshi & Weinstein 1983), even if it is morphologically 
ambiguous. 
Choice of a demonstrative NP does not depend ou the 
conceptual nature of the referent, which may be either 
the local focus or the global focus. Its contrastive 
functions with respect to personal pronouns and definite 
NPs are rather complex. Since demonstratives are 
7) Anaphoric devices thus have a default (strict) order of 
priority for application. 
8) We omit the realization constraints, such as the fact that 
certain verbs do not allow their complements to be 
pronominalized. 
infrequent in onr corpus, lbey are not treated furllter 
here. 
Fur a definite NP, there is no conceptual iestriction on 
the refm'ent. A definite NP can be introduced at 
substaulial distance from its textual antecedent, and 
typically dnes not occl,r in the following clan.~, 
especially if the antecedent was the local focus of its 
clause and there is no potential ambiguity 9 . 
For each NP to be generated, potential ambiguity 
must be taken into account. This has to tlo with lexieal 
choice. For example, choice of an anlbignons NP such as 
& vin \[the wiue\] must be blocked if there is white wine 
and red wine in the context. The context in which the 
anaphoric NP must be distinctive depends on the 
annphor chosen: it is tile preceding sentence for 
demonstrative NP while, for definite NP, a larger context 
nmst be taken in account lO, 
5.2 Lexical amlphoru 
We now tun, to the constraints on choice of lexical 
anaphoL When the grammatical mechanism chosen tklr 
exprc.ssing atnaphora involves a corellareutial (definite or 
den:tonstrative) lcxical NP, these c(mstraiuts conic into 
play to pick the most appropriate lexical form. The 
anaphoric lexical devices presented here for recipes 
constitute only a subset of those that conld appear in the 
language as a whole. Nevertheless, we hylX)thesize that 
the COl)ceptual and linguistic constraints governing their 
usage are generalizable to other kinds of text. Lexical 
anaphora differs significantly on this point from 
grammatical anaphora, whnse constraints, like discourse 
structure or focalization, vary greatly according to rite 
kiud of text. Therefilrc, while a giveu kind of text might 
use only a subset of possible lexieal auaphoric devices, 
these devices are governed by the mtme constraints in all 
kinds of texts. For exautple, typical result mention (m/x - 
-> mixture) is widespread in procedural texts but 
constradltS governing thCnl are tile same ill ally kind of 
lext. lu contrast, it appears that tile constraints governing 
usage of granlmalical anaphoric devices, anti even the 
devices thenlselves, are much more depenttent on the 
variety of text. 
Given that a lexica\] NP has been chosen, as the 
general type of anaphoric device, two kinds of 
constraints, conceptual and linguistic, apply to select the 
the specific kind(s) of lexical aUalthora which may be 
nsed. In case of anthiguity, i.e. if file NP produced is not 
distinctive, addilional processing will t'.c requirexl. 
Conceptual Cmlslraittts concern mainly: 
The state of the object . l,'or example, all object 
whose state is being transfonued by an action should be 
relcrenced via its resulting state. 
- Groupings of ohjects: is the referent to be generated 
a set of identical objects, a heterogenous set, a 
homogenous sct or a single element? A heterogenons set 
is composed of elements which just have no close 
9) For example, die definite NP hi file second clause is ra)t very 
natural in French: 
Marie a rencontr6 un charcufier. ~_¢I f~it on tr~s 
bun pht& 
\[Marie met a porkbutcher. ~Ii~ makes very 
go~,d l~.l 
10) For recipes, we use Date's (1988) proposal to take the 
whole text as context, since it is usually short. This would of 
CoIIrSe not he satisfactory hlr longer texts, 
ACRES DE COLING-92, NANTES, 23-28 AO~r 1992 7 6 7 PROC. O1~ COLING-92. NANn.:S. AU¢;. 23-28, 1992 
generic concepts in common, such as, 
{ "salt" ,"knife" ,"table"}. 
Linguistic constraints involve mostly the lexical 
form and relative order of the coreferential NPs that 
have been lexicalized in the preceding text. Therefore, 
we do take advantage of referents already lexiealized in 
the previous context (which must be stacked for being 
available when lexicalizing). 
The following properties are examined: 
- The linguistic form of antecedent NP: is it a single 
noun, a compound noun or a complex NP? 
- The existence of a lexico-semantic association for the 
antecedent like the generic term or the typical result 
(which can mostly be formalized through a LF). 
- The "basicness" of the head word of the antecedent NP. 
Ambiguity constraints are used to check if the 
lexicalization is not ambiguous. 
If a unique object or a set of identical objects can not 
be lexicalized in a non ambiguous way, we lexicalize it 
the same way it has been first introduced in the text 
(Initial strict repetition). We use this ad hoc strategy 
because first mention of a referent is generally the most 
accurate. Of course, this would not always be the 
minimal distinguishing description (Dale 1988), but as 
Reiter (1990a) points out, determining a minimal 
distinguishing description may require overly complex 
processing. 
In case of potential ambiguity for a set of 
heterogeneous objects, we use "complex coordination". 
With this process, we regroup first the first level 
superordinates and apply the other devices to the 
remaining list of objects 11. 
Table 2 shows several important kinds of lexical 
anaphoric devices, with their associated conceptual, 
linguistic and non ambiguity conditions. 
Concepicual Properties Linguistic Non ambiguity " 
Lexleal anaphor Properties Constraints Examples 
Unique object'or set of Antecedent is a 'single No instance previously lapin-->lapin \[rabi~it\] Strict Repetition 
Initial Strict Repetition 
identical objects 
Unique object or set of 
noun (or fixed compo- 
und) and is a basic noun 
introduced has the same 
repetition 
No constraints Tile other devices are A small rabbit ... the 
identical objects ambiguous rabbit --> the small rabbit 
Partial Repetition Unique object or set of Antecedent is a not fixed No previously introduced petit lapin \[small rabbit\] 
identical objects compound (except "part- NP has the same partial --> lapin 
of" types) and the NP repetition 
head is a basic nt~un 
Superordlnatlon Set' of objects having a\[ Nominal heads of ante- No previously introduced {carottes, poireaux, 
close common genericl cedents have the same NP has the same 
concept common superordinate, supemrdinate term 
LF: Gener 
Basic Denomination Umque Object or set of Nominal head of NP is No previously introduced 
identical objects not a basic noun NP has the same basic 
denomination 
N°mlnallzatl°n Action '" No constraints 
Object(s) having been 
, affected by a strong 
transformation 
Set of different objects 
which have no common 
genetic concept 
Typical Result Mention 
Antecedent verb can be 
nominalized or super- 
ordinate of antecedent 
verb can be nominalized. 
LFs: S o or S O o Gener 
There is a result noun for 
file actants having been 
affected by the 
tranformation. 
LF: S,,~ 
No element of the 
coordination is 
ambiguous 
Complex 
Coordination 
No previously introduced 
NP has tbe same result 
mention 
No constraints 
tomates} \[carrots, leeks, 
tomatoes\] --> 
16~umes \[vegetables\] 
petites girolles 
\[small chanterellesl --> 
ehampignons 
\[mushrooms\] 
faire cuire le poulet 
\[cook the chicken\] --> la 
cuisson du poulet \[the 
cooking of the chicken\] 
mtlanger les patates i 
\[mix the potatoes i \] --> le 
mtlange i \[the mixmrei\] 
{petit lapin, grosses 
chanterelles } \[small 
rabbit, big chanterelles\] 
--> le lapin et les cham- 
pignons \[le lapin et les 
champignons\] 
Table 2: Constraints governing the introduction of lexical anaphora 
11) We choose here to apply supe~ordination separately to each 
instance: we do not allow regroupings of elements for 
superordinatinn or typical result mendon because, as Kosseim 
has noticed, we would have to process all the subsets to 
generate correct lexicalizatinns. 
AcrEs DE COLING-92, NANTES, 23-28 hO~" 1992 7 6 8 Paoc. OF COL1NG-92, NANTES, AUO. 23-2g, 1992 
Conclusion 
In this tmper, we have described some of the problems 
raised making lexical choice in textual context, in 
particular for coreferential lexical anaphora. We have 
showed that paradigmatic Lexical Functions are well 
suited for creating lexical coreferential links. We have 
also distinguished the selection tff superordinate term, 
which is used to point back to a set of different words, 
from selection of basic denomination, which is usetl to 
name in the most natural way a concept already 
introduced by a previous noun. 
A series of constraiuts has been formulated which can 
be implemented in an algorithm for selecting among 
natural grammatical and lexical anaphors in procedural 
texts. Most of these algorithms have been implemented 
by Kosseim 1992. The generator uses Prolog anti 
specifically Definite Clause Grammar (DCG) to produce 
text. 
We find that determination of grammatical anaphom 
is more dependent on the genre and sublanguage than is 
lexical anaphora, which appears governed by fairly 
general constraints. However, morn work needs to be 
done to check these results in other precedural texts, and 
then more broadly in less similar text types, Also, it 
would be interesting to see to what extent anaphoric 
expressions share common constraints with deictic 
expressions for which the context of interpretation is not 
the previous text, but the extra-linguistic context~ 
Acknowledgements 
We would like to thank Guy Lapalme, Igor Mel'~uk, 
Alain Polgu~re, Marga Alonso Ramos and Xiaobo Ren 
for fruitful discussions and helpful suggestions. Special 
thanks to Le'fla Kosseim, who collaborated in this 
research and with whom we shared many interesting 
discussions. The work reported in this paper was 
supported by a Government of Canada Award. 
References 
Bateman J. A. & L.Wanner (1990). Lexical Cooccurrence 
Relations in Text Generation, in Proceedings of the Fifth 
International Workshop on Natural Language Generation, 
Dawson, Pennsylvania, 31-38. 
Dale, R. (1988). Generating Referring Expressions in a 
Domain of Objects and Processes, Ph..D. thesLs, University of 
Edinburgh. 
Dtcary M. & G. Lapalme (1990). An Editor for the 
Explanatory Dictionary of Contemporary French (DECFC), 
ComputationalLinguistics, 16, 3, 145-154. 
Grosz B.J., Joshi A., Weinstein S. (1983). Providing a 
Unified Account of Definite Noun Phrases in Discourse, in 
Proceedings of the 21st Annual Meeting of the ACL, MIT, 
Cambridge, Mass., 15-17 june, 1983, 44,49. 
Grosz B.J. & C. Sidner (1986). Attention, Intentions and the 
Structure of Discourse, Computational Linguistics, 12, 175- 
204. 
Halliday M.A.K & R. Hasan (1976). Cohesion in English 
London, Longman. 
Heid U. & S. R~b (1989). Collocations in Multilingual 
Generation, in Proceedings of FACL, 130-136. 
Hobbs J. (1978). Resolving Pronoun References, Lingua, 
44, 311-338. 
Iordanskaja L., R. Kittredge & A. Polgu\[re (1991). Lexical 
Selection and Paraphrase in a Meaning-Text Generation Model 
in C.L. Paris, W. R. Swartout & W.C. Mann eds.,Natural 
Language Generation in Artificial Intelligence and 
Computational Linguistics, 293 -312. 
Kosseim L. (1992), G~n~.ration automatique de proc~dgs 
coh~sifv darts les recettes de cuisine, M.Sc. thesis, Dtpartement 
d'informatique et de recherche op&ationnalle, tJniversit6 de 
Monutal. 
McDonald D. (1991). On the Place of Words In lhe 
Generation Process in C.L. Paris, W. R. Swastout & W.C. 
Mann eds.,Natural Language Generation in Artificial 
Intelligence arm Computational Linguistics, 229-247. 
Mel'fiuk LA. et al. (1988a). Dictionnaire Explicatif et 
Combinatoire du Franga~ Conternporain. Recherches Lexico- 
sbnantiques. 11. Montrtal, Presses de l'Universit~ de Montrtal. 
Mel'~'uk I. (1988b). Paraphrase et lexique dans la Thtorie 
Sens-Texte, Cahiers de Lexicologie LII, 5-50 et LIlI, 5-53. 
Mel'c'uk I. & A. Polgu&e (1987) A Formal Lexicon ill the 
Meaning-Text Theory (or how to do Lexica with Words), 
Computational Linguistics, 13, 3"4, 1987, 261-275. 
Mel'cuk 1. & Zholkovsky, D (1970). Sur la synthtse 
s6mmltique, TA. Informationa', 2, 1-85. 
Nirenbnrg S. & I. Nirenburg (1988). A Framework for 
Lexical Selection in Natural Language Generation, in 
Proceedings of COLING 88, Budapest, 471 "475. 
Nirenburg S. & V. Raskin (1987). The Subworld Concept 
Lexicon and the Lexicon Management System, Computational 
Linguistics, 13,3"4, 276-289. 
Reiter E.B. (1990a). Generating Appropriate Natural 
Language Object Descriptions, Ph.D. Thesis., llarvard 
University. 
Reiter E.B. (1990b) A New Model for Lexical Choice for 
Open-Class Words, in Proceedings of the Fifth haernatio~gd 
Workshop on Natural Language Generation, Linden Hall 
Conference Center, Dawson, Pennsylvania, 23-30. 
Rosch E., C. B. Mervis, W. 1). Wayne, D. M. Jolmson & P. 
Boyes-Braen (1976). Basic Objects in Natural Categories, 
Cogr,;tive Psychology 8, 382,439. 
Sidner C. (1983). Focusing in the Comprehension of 
Definite Anaphora, in M. Brady & R. Berwick 
eds.,Computational Models of DL~course, Cambridge (UK.), 
Cambridge University Press, 267-33f). 
Tutin A. (1992) Etude des anaphores grammaticales et 
lexicales dane la perspective de la g~t~ration automatique 
darts des textes de procedures, Ph.D. Thesis, IMpartement de 
linguistique, Universit6 de Montrtal. 
Wierzbicka A. (1985). Lexicography and conceptual 
analysis, Ann Arbor (Mich.), Keroma Publishers inc. 
ACRES DE COLING-92, NAmeS, 23-28 Aou'r 1992 7 6 9 PRoc. o1: COLING-92, NANTES, AUG. 23-28, 1992 
