A Computational Model 
for Generating Referring Expressions 
in a Multilingual Application Domain 
Elena Not 
IRST 
Loc. Pant~, 1-38050 Povo - Trento, Italy 
not@irst.itc.it 
Abstract 
In this paper we analyse the problem 
of generating referring expressions in a 
multilingnal generation system that pro- 
duces instructions on how to fill out 
pension forms. The model we pro- 
pose is an implementation of the the- 
oretical investigations of Martin and is 
based on a clear representation of the 
knowledge sources and choices that con- 
tribute to the identification of the most 
appropriate linguistic expressions. To 
cope effectively with pronominalization 
we propose to augment the Centering 
Model with mechanisms exploiting the 
discourse structure. At every stage of 
the referring expressions generation pro- 
cess issues raised by multilinguality are 
considered and dealt with by means of 
rules customized with respect to the out- 
put language. 
1 Introduction 
An automatic generation system that is to pro- 
duce good quality texts has to include effective 
algorithms for choosing the linguistic expressions 
referring to the domain entities. The expressions 
have to allow the reader to easily identify the re- 
ferred objects, avoiding ambiguities and unwanted 
implications. They have to conform to the expec- 
tations of the reader according to his evolving flow 
of attention and they have to contribute to the co- 
hesion of the text. 
In this paper, we describe a component, devel- 
oped inside the GIST project 1, building referring 
1The GIST consortium includes academic and in- 
dustrial partners -IRST (Trento, Italy), ITRI (Uni- 
versity of Brighton, Great Britain), OFAI (Vienna, 
Austria), Quinary (Milano, Italy), Universidade Com- 
plutense de Madrid (Spain)- as well as two user groups 
collaborating actively to the specification and evalua- 
tion of the system -INPS (the Italian National Security 
Service) and the Autononm Province of Bolzano. 
expressions for automatically generated multilin- 
gum (English, German, Italian) instructions in 
the pension domain. The overall decision making 
mechanism for the referring expressions choices is 
based on the theoretical investigations of Martin 
(Martin, 1992), for which we propose a possible 
implementation. The implemented model proved 
to be particularly suitable to work in a multilin- 
gum domain. For the generation of pronouns we 
define an extension of the Centering Model ex- 
ploiting the contextual information provided by 
the rhetorical structure of discourse (Not and Zan- 
canaro, 1996). 
At every stage of the referring expressions gen- 
eration process issues raised by multilinguality are 
considered and dealt with by means of rules cus- 
tomized with respect to the language. In section 
2 we first present the results of observations made 
on the corpus texts with the aim of identifying the 
typical referring expressions occurring in our do- 
main. Section 3 details the solutions implemented 
in the GIST system. Specifications for the imple- 
mentation are given in terms of data structures 
and required algorithms. 
2 Referring expressions in 
multilingual pension forms 
Ore" work on the specification for the referring ex- 
pressions component started from an analysis of 
the collected multilingual (English, German, Ital- 
ian) corpus of texts containing instructions on how 
to fill out pension forms. From this study, a gen- 
eral typology of entities referred to in the domain 
emerged together with an indication of how such 
entities are typically expressed in the different lan- 
guages (see figure 1). The classification includes: 
Specific entities. These are extensional entities: 
individuals or collections of individuals (plu- 
rals). In KL-One knowledge representation 
languages they are represented as instances 
in the A-box. 
Generic entities. These entities are intensional 
descriptions of classes of individuals and 
are often mentioned in administrative doc- 
848 
__ specific 
entitles 
unique 
individuals ~ retbrence (INPS, DSS) 
1_ variable reference 
(a benefit) 
unique < reference 
(the states phmds in the EU) 
variable reference 
(some benefits) 
__ anchored f_ cntitics 
unique reference (the applicant, the form) 
variable reference (one of applicant's 
previous jobs) 
-- generic 
entities (/,ousewives, widows) 
Figure 1: Types of entities referred 
uments, since the entities (persons or inan- 
imate objects) addressed in this kind of texts 
are not usually specific individuals in the 
mind of the public administrator but rather 
all the individuals that belong to a certain 
class, as in the following example: 
(l) Married women should send their marriage 
certificate. 
In K1,-()ne knowledge representation lan- 
guages generic entities are represented as con- 
cepts in the T-box. 
Anchored entities. These are entities that, al- 
though generic in nature, can be interpreted 
us specific whet, they are considered in the 
specific communicative situation in which the 
actual applicant reads the instructions to 
complete the pension form he has in his 
hands. Consider {br example the following 
text: "The applicant has to provide all the re- 
quested information". In this situation, the 
specific person who is reading the form in- 
stun,tares the generic applicant. All the enti- 
ties directly related to the applicant or to the 
form itself can be considered anchored, as for 
example: the applicant's name, the applican- 
t's spouse, any applicant's previous job, see- 
,ion 3 of the h)rm,.... The plausibility of this 
anchoring operation is confirmed by the fact 
that the linguistic realization choices made 
for anchored entities (definite tbrms, singular 
indefinite forms, ...) resemble very much the 
linguistic choices made tbr specific entities. 
Further investigations on the corpus texts have 
been conducted to identify language-dependent 
referring phenomena and general heuristics for the 
choice of the most appropriate linguistic realiza- 
tion. In general, we found that language style has 
great influence on the realization choices. When 
an informal style is used (like in most English doc- 
uments and in some recent Italian/German forms) 
the personal distance between the interlocutors 
(the citizen and the public institution) is reduced 
using direct references to interlocutors, by means 
of personal pronouns ("you", "we"). When the 
language is more formal, impersonal forms or in- 
direct references are preferred ("the applicant", 
"INPS", "DSS"). 
Apart from style differences, there do exist also 
differences in realization that depend on the out- 
put language. For example, in administrative 
forms, in full sentences, for entities anchored to 
the reader English and German typically use pos- 
sessive noun phrases (like "your spouse") whereas 
Italian prefers simple definite forms (e.g. "il cent- 
age" \[the spouse\]). 
3 The adopted approach 
The linguistic expressions that refer in the text. to 
tile (tomain entities have to fulfill several proper- 
ties: 
• they must allow the non-ambiguous iden- 
tification of the entities 2 ; 
• they should avoid redundancies that 
could hamper lluency; 
• they should contribute to the cohesion 
of the text by signaling semantic links 
between portions of text; 
• they should conform to the formMity and 
politeness requirements imposed to the 
output texts. 
When we choose to realize a referring expression 
with an anaphora we fulfill a double fnnction: we 
introduce some form of economy for the reference, 
avoiding the repetition of a long linguistic expres- 
sion, and we enhance the coherence of the text 
since we signal meaning relations (cohesive ties) 
between portions of the discourse. 
The choice of the correct referring expression 
depends on two major factors: 
(A) the cohesive ties that we want to signal to 
improve the cohesion of the text; 
(B) the semantic features that allow the identi- 
fication of the object in the domain (distin- 
guishing semantic features). 
Another relevant factor is the pragmatic setting 
of the discourse (formality and politeness). 
To decide on (A), data structures are main- 
tained that keep track of the evolving textual con- 
text (discourse structure and focus history) and 
record the seato-cultural background of the reader 
2In some genres the use of ambiguous references 
may be possible or desirable, for exantple in jokes, but 
in administrative genre clearness and unambiguity are 
the primary goals. 
849 
IDENTIFICATION 
reference < 
~ generic reference 
anchored reference 
specific reference 
_1  individual reference plural reference ( 
-- presenting 
presuming 1~ 
( asserting 
questioning 
total 
partial 
-- variable reference 
unique reference 
variable reference 
nominal 
{ pronominal interlocutors (INPS / we the applicant / ~ou) non-interlocutors 
(Gianni Billia) 
pronominal (he~she) ~ proximate (this~these..) 
directed I-- 
nominal -~ distant (that~those ..) 
t_ undirected (the..) 
Figure 2: How semantic features combine to identify the entity in the context 
(user model). Inquiries on these data structures 
are performed to verify whether the identity of 
the entity can be recovered from the context or 
whether there exist semantic relations with other 
cited entities that; are worth being signaled (e.g. 
cotnparative relations). 
Once the ties have been determined, the distin- 
guishing semantic features are identified. These 
semantic features depend on the entity type - 
whether generic, anchored or specific - and on the 
relationships between the entity and the context - 
whether the entity is new with respect to the con- 
text (presenting) or its identity can be recovered 
(presuming). Figure 2 illustrates a fine grained 
distinction of semantic features whose combina- 
tion specify how a referring expression can be 
built. This network of choices is an adaptation to 
the GIST application domain of the results pre- 
sented in (Martin, 1992). 
'\['he total/partial opposition is used to distin- 
guish references to sets of elements from references 
to portions of sets. The linguistic form of the ex- 
pression also varies according to the type of speech 
act that is to be realized, and this justifies the as- 
serting~questioning distinction. 
Entities may be presented as new in the dis- 
course context through references composed by a 
nominal expression or a pronoun (presenting). 
A presupposed element (presuming) may be- 
long to the cultural/social context, and therefore 
be described with a unique reference, or it may 
belong to the textual context. The presuming- 
variable option corresponds to a textual anaphora. 
In this case a pronoun or a definite expression 
can be used. In our system, pronominalization 
is decided according to new rules extending the 
Centering Model, as explained in the following 
section 3.3. Definite expressions are built select- 
ing the appropriate determiner (the, this, that... ) 
and the information (head, modifiers) to put in 
the noun phrase. This latter information is deter- 
mined through the algorithm explained in section 
3.2. 
3.1 The global algorithm 
The submodule for the generation of referring ex- 
pressions is called during the final stage of the 
text planning process, when the so called micro- 
planning (or sentence planning) takes place (Not 
and Pianta, 1995). The global algorithm imple- 
mented has been derived from the network of 
choices presented above, as emerging from the 
corpus analysis. The formal approach adopted 
proved to be particularly suitable to cope with 
multilinguality issues, since the tests performed at 
the various choice points can be easily customized 
according to the output language. The algorithm 
is activated on each object entity to be referred 
and accesses the following available contextual in- 
formation: 
Background - the cultural and social context of 
the reader. At present this is represented by a 
list of all the entities the reader is supposed to 
know (e.g. the Department for Social Security, 
the anchored entities); 
850 
RT - the rhetorical tree, specifying how the se- 
lected content units will be organized in the final 
text and which are the semant.ic relations between 
text spans that will be signaled to enhance the co- 
herence; 
AlreadyMentioned - the history of Mready men- 
tioned entities; 
StylePars - the l)arameters that define l,he style 
of the output text,; 
FocusState - the state of the attention of tile 
reader, organized as (tetailed in section 3.3. 
To model the rhetorical structure of discourse 
we consider the Rhetorical Structure Theory as 
developed in (Mann and Thompson, 1!)87). Ac- 
cording to this theory, each text can be seen as 
a sequence of clauses linked together by (seman- 
tic) relations. These relations may be grammati- 
cMly, lexically or graphically signaled. About 20 
such relations have t)een identified by (Mann and 
Thompson, t987), e.g. I!\]I,AI~OH, ATION, which 
occurs when one clause provides more details for 
a topic presented in the previous clause, C()N~ 
TI{AST which links two clauses describing similar 
situations differing in few respects, and so on. 
llere follows a sketch of the globM algorithm 
implemented (Not, 1995). ~\[b make the reading 
easier, labels in italics have been introduced to 
identify the steps of the algorithm corresponding 
to tile main choice points in figure 2. 
Prelimina.ry step: 
• (For English) if e iv an anchored entity treat it as 
if it was a specific entity in Background 
• (For Italian and German) if e is an anchored entity 
inside a concept description 
then treat it as a presenting of ~ generic entity 
with a nominM expression (goto presenting- nominal-generic) 
else treat it as if it w;m a specific entity in Back- 
ground 
Ill case: 
• e is referred to in a title ~md is anchored to the 
reader 
(English, German) if *|'ormMity* ~ inlbrmM 
then use a noun phrase with the possessive 
adjective "your" 
else use a b;Lre nou\[\[ phrase 
(ItMian) use a bare noun phrase 
• e is referred to in a title (but is not anchored to 
the reader) or in a label 
use a bare noun phrase (singular or plural ac- 
cording to the nunlber of e) 
• e C- AlreadyMentioned U Background 
then \[presuming\]: if e C Background 
then \[unique\]: if e is-a interlocutor 
then \[interlocutor\]: if *formMity* = in- 
\[orma.l 
then use a pronoun 
Mse;,use a proper noun (if it exists) or 
a definite description 
else \[non-interlocuto@ 
(English, German) if *formality* = 
informal and e is anchored to the 
reader 
then use a noun phrase with the pos- 
sessive adjective "your" 
else use a proper name or a definite 
description 
(Italian) use a proper name or a deft- 
nite description 
else \[variable\]: attempt pronominalization 
using the algorithm described in section 
3.3 accessing FocusState and RT. If e is 
pronominalizable 
then \[pronominal\]: use a pronoun 
else \[nominal: build an anaphoric ex- 
pression. Test FocusState to identify 
the ntost appropriate determiner for 
the noun phrase. Compute the head 
and the modifiers using the algorithm 
described in section 3.2. 
else \[presenting\]: if e stands for a generic person 
(collection of persons) without any specifi(: 
property 
then \[pronominal\]: use im indefinite pro- 
l!.oun 
else \[nominal: build a noun phrase, choos- 
ing the appropriate linguistic form 
If e is a: 
- specific entity, build an indefinite sin- 
gular description or an indefinite plu- 
ral description according to the nutn- 
her of e 
- generic entity, in ease: 
* e is a concept whose meaning is be- 
ing defined by syntesis, use the bare 
singular term 
. e is a concept 1)eing defined 
through a listing of its components, 
use a definite singular noun phrase 
. e appears in a list inside a concept 
definition, 
(German, Itdian) use a bare singu- 
lar or bare plural noun phrase 
(English) use a definite singular or 
definite plural noun phrase 
. e is in a question, use a singular 
indefinite noun phrase 
e e is used in procedural descrip- 
tions, 
(Italian, German) use a definite phl- 
rM description. 
(English) use a bare plural. 
3.2 Generating nominal expressions 
In this section we focus on the choice of the head 
and the modifiers tbr noun phrases. (Dale and 
Reiter, 1995) contains the following list of require- 
ments for a referring expression to obey to Grice's 
Maxims of conversational irnplicature: 
851 
1. The referring expression should not in- 
clude unnecessary information (the Maxim of 
Quantity). 
2. The referring expression should only spec- 
ify properties that have some discriminatory 
power (the Maxim of Relevance). 
3. The referring expression should be short (the 
Maxim of Brevity). 
4. The referring expression should use basic- 
level and other lexically preferred classes 
whenever possible (Lexical Preference). 
l~equirement (4) suggests that the head of the 
noun phrase should be chosen among terms of 
common use or, more in general, among terms 
that the user is likely to know. In our domain, 
however, often technical terms can not be avoided 
since the precise type of document or legal re- 
quirement have to be specified. Therefore, for the 
choice of the head of non-anaphoric expressions 
the GIST system adopts the strategy of using the 
most specific superconcept of the entity that has 
a meaningful lexical item associated (e.g. the spe- 
cific term "decree absolute" is used instead of the 
more basic term "certificate"). 
Requirements (1),..,(3) suggest that the modi- 
fiers in the noun phrase should not introduce un- 
necessary information that can hamper the text 
fluency and yield false implications. The task of 
selecting the correct modifiers for a non-anaphoric 
expression is not an easy task, since in the Knowl- 
edge Base attributive and distinguishing (restric- 
tive) properties are mixed. In GIST, the se- 
mantic relations that are relevant in the defi- 
nition of distinguishing descriptions have been 
identified through an accurate domain analysis. 
For example, we have chosen relations like has- 
partnership, owned-by or attribute-of, characteriz- 
ing distinguishing descriptions like "the applican- 
t's spouse" or "the applicant's estate". 
When an anaphora occurs but a pronoun can 
not be used, a nominal anaphoric expression is 
built. The head and the modifiers included in the 
noun phrase have to allow the identification of the 
entity among all the ones active in the reader's at- 
tention (potential distractors). In GIST we adopt 
an algorithm which is a simplified variation of the 
one Dale and l:teiter call the "Incremental Algo- 
rithm" (Dale and Reiter, 1995): whenever a new 
nominal anaphoric expression has to be built, dis- 
criminant modifiers are added to the expression 
until the set of the potential distractors (contrast 
set) is reduced to an empty set. 
3.3 Generating pronouns 
For the generation of pronouns an extension to the 
Centering Model (Grosz et M., 1995) has been de- 
fined that captures how the rhetorical evolution 
of the discourse influences the flow of attention 
of the reader. The choice of this solution has 
emerged from the observation that anaphora plays 
two roles in the discourse: it is not sufficient that a 
pronoun identifies unambiguously its referent but 
it has to reinforce the coherence of the text as well, 
supporting the user's expectations. 
In the Centering Model for each utterance U,~ 
a list of forward looking centers, Cf(Un), made up 
of all the entities realized in the utterance, is asso- 
ciated. This list is ordered according to the likeli- 
hood for the elements of being the primary focus 
of the following discourse. The first element in the 
list is called the preferred center, Cp(U,~). Among 
the centers another significant entity is identified: 
the backward looking cen~er, Cb(Un). This repre- 
sents the primary focus of Un and links the current 
sentence with the previous discourse. 
The basic constraint on center realization is for- 
mulated in the following rules: 
RULE 1 : If any element of Cf(U,~) is realized 
by a pronoun in U,+I then the Cb(U,~+I) must be 
realized by a pronoun also. (Grosz et al., 1995) 
RULE 1' : If an element in Cf(Un+I) is coref- 
erent with Cp(U,~) then it can be pronominalized. 
(Kehler, 1993) 
These rules can be used to constrain pronominal- 
ization in the text generation process. 
The Centering Model was first conceived for En- 
glish, a language where pronouns are always made 
explicit. But as soon as we consider languages 
that allow null pronominMization (like Italian) 
new extensions to the original model have to be 
designed in order to deal with pronouns with no 
phonetic content. For Italian, we defined the fol- 
lowing rule (Not and Zancanaro, 1996) which is 
compatible with the results of empirical research 
presented in (Di Eugenio, 1995): 
RULE 1" : If the Cb of the current utter- 
ance (Cb(U,,+I)) is the same as the Cp of the 
previous utterance (Cp(U~)) then a null pronoun 
should be used. If, instead, Cb(U,~+I) # Cp(U,~) 
and Cb(U,,+I) = Cb(U,~) then a strong pronoun 
should be used. 
3.3.1 The proposed extension to the 
Centering Model 
Unfortunately, the Centering Model does not 
capture completely the reader's flow of attention 
process since it fails to give an account of the ex- 
pectations raised by the role the clause plays in 
the discourse. For example consider the following 
sentences: 
(2) a. If you are separated, 
b. \[your spouse\]i should send us \[this part of 
the form\]j properly filled in. 
c. \[They\]i should use \[the enclosed envelope\]k. 
d. ek does not need a stamp. 
According to the Centering rules it would not 
be possible to use a pronoun to realize ek since 
the main center of utterance d. (the envelope) is 
852 
different from the main center of utterance c. (the 
spouse). But the use of a definite noun phrase to 
refer back to the envelope would sonnd rather odd 
to a native speaker. 
Itowever, the rhetorical structure of the text, 
providing information on the semantic links be- 
tween utterances, helps understanding how the 
content presentation progresses. Therefore, we 
claim that it can be used to explain exceptions 
to the Centering rules and used to define repairing 
strategies (Not and Zancanaro, 1996). The advan- 
tage of this solution is that it allows us to treat 
with a uniform approach different types of excep- 
tions that in literature are solved with separated 
ad-hoc solutions (e.g. parallelism, empathy). 
For exa.inl)le, in (2) above sentence d. is an 
evident ELABORATION on the envelope that ap- 
pears in sentence e. When elaborating the de- 
scription of an object the focus of attention moves 
onto the objecL itself. Therefore, the rhetorical re- 
lation that links e. and d. signals that among the 
elements in Cf(c) the envelope is the best can- 
didate to be the primary focus of the following 
sentence d. This means that the rhetorical infor- 
mation ('an "project" the default ordering of the 
elements in /;he potential focus list Cf(c) onto a 
new order tha~ reflects more closely the content 
progression. 
From a computational point of view, the re- 
suiting algorithm h)r pronominalization can be 
sketched as follows. The reader's attentional state 
is recorded in two stacks: the Centers tlistory 
Stack and the BackwaTq Centers Stack collect- 
ing respectively the Cf and the Cb of the already 
produced utterances. Whenever a new utterance 
is processed, the corresponding Cf and Cb are 
pushed on the top of the two stacks. The Cf list is 
ranked according to the default ranking strategy: 
clause theme > actor > benefic.> actee > others 
possibly modified by a "projection" imposed by 
the rhetorical relation. Rules 1' (for English and 
German) and Rule 1" (for Italian) arc then used 
to decide wllethcr a pronoun ('an be used or not. 
4 Conclusion 
We have presented the computational model im- 
plemented in the GIST system for referring ex- 
pressions generation. The model is based on a 
clear distinction of the various knowledge sources 
that come into play in the referring process and 
provides an implementation tbr Martin's theoret- 
ical invesl~igatious. An extension of the Centering 
Theory has been proposed to deal with pronom- 
iualization effectively, exploiting the information 
provided by the discourse structure on how the 
reader's flow of attention progresses. Issues of 
multilinguality are treated by ('ustomizing the se- 
lection rules according to the output language. 
5 Acknowledgments 
The global approach to the generation of 
anaphoric expressions presented in this paper, and 
in particular the treatment of pronominalization, 
has been developed together with Massimo Zan- 
canaro, whose help I gratefully acknowledge. 
References 
Robert Dale and Ehud Reiter. 1995. Computa- 
tional Interpretations of the Gricean Maxims in 
the Generation of Referring Expressions. Cog- 
nitive Science, 19:233-263. 
Barbara Di Eugenio. 1995. Centering in Italian. 
In Ellen Prince, Aravind Joshi, and Lyn Walker, 
editors, Centering in Discourse. Oxford Univer- 
sity Press. 
Barbara J. Grosz, Aravind K. Joshi, and Scott 
Weinstein. 1995. Centering: A Framework 
for Modelling the Local Coherence of Discourse. 
Computational Linguistics, 21(2), June. 
Andrew Kehler. 1993. lntrasentential Constraints 
on Intersentential Anaphora in Centering The- 
ory. In Proceedings of Workshop on Centering, 
University of Pennsylvania. 
William C. Mann and Sandra A. Thompson. 
1987. Rhetorical Structure Theory: A Theory 
of Text Organization. In Livia Polanyi, editor, 
The Structure of Discourse. Ablex Publishing 
Corporation. 
J. R. Martin. 1992. English Text. System and 
Structure. John Benjamins Publishing Com- 
pany. 
Elena Not and Emanuele Pianta. 1995. Issues of 
Multilinguality in the Automatic Generation of 
Administrative Instructional Texts. In M. Gori 
and G. Soda, editors, Topics in Artificial ht- 
telligenee, Proceedings of the Fourth Congress 
of the Italian Association for Artificial Intelli- 
gence, Lecture Notes in Artificial Intelligence. 
Springer. Also available as IRST Technical Re- 
port #9505-17, May 1995. 
Elena Not and Massimo Zancanaro. 1996. Ex- 
ploiting the Discourse Structure for Anaphora 
Generation. in Proceedings of the Discourse 
Anaphora and Anaphor Resolution Colloquium, 
DAARC96, Lancaster University, 17-18th July. 
Elena Not. 1995. Specifications for the Refer- 
ring Expressions Component. TechnicM Report 
GIST LRE Project (062-09), Deliverable TSP- 
4, IRST, September. 
853 
