Integrating Referring and Informing in NP Planning 
Michael O'Donnell, Hua Cheng, 
Department of Artificial Intelligence, 
University of Edinburgh, 
80 South Bridge, Edinburgh, Scotland, EH3 9JF. 
Janet Hitzeman 
Human Communication Research Centre, 
University of Edinburgh 
2 Buccleuch Place, Edinburgh, Scotland, EH8 9LW. 
Abstract 
Two of the functions of an NP are to refer 
(identify a particular entity) and to inform 
(provide new information about an entity). 
While many NPs may serve only one of these 
functions, some NPs conflate the functions, not 
only referring but also providing new informa- 
tion about the referent. For instance, this deli- 
cious apple indicates not only which apple the 
speaker is referring to, but also provides inform- 
ation as to the speaker's appreciation of the 
apple. 
This paper describes an implemented NP- 
planning system which integrates informing into 
the referring expression generation process. The 
integration involves allowing informing to influ- 
ence decisions at each stage of the formation 
of the referring form, including: the selection 
of the form of the NP; the choice of the head 
of a common NP; the choice of the Deictic in 
common NPs; the choice of restrictive modifi- 
ers, and the inclusion of non-referring modifi- 
ers. The system is domain-independent, and is 
presently functioning within a full text genera- 
tion system. 
1 Introduction 
Two of the functions of an NP are to refer 
(identify a particular entity) and to inform 
(provide new information about an entity). In 
most cases, a given NP may serve only one of 
these functions. However, in some cases, the 
writer/speaker may choose to conflate the func- 
tions, providing an NP which not only refers but 
also provides new information about the refer- 
ent. For instance, this delicious apple indicates 
not only which apple the speaker is referring to, 
but also provides information as to the speaker's 
appreciation of the apple. 
Most of the work on NP planning has con- 
sidered only the referring function of the NP 
(e.g., Dale 1988, 1989; Reiter 1990; Reiter & 
Dale 1992; Horacek 1995). Appelt (e.g., Ap- 
pelt 1985; Appelt & Kronfeld 1987) has con- 
sidered the question of integrating referring and 
informing, although rather briefly, and without 
much detail. This paper will extend upon his 
discussion, and describe its role in ILEX, a text 
generation system which delivers descriptions of 
entities on-line from an underlying knowledge- 
base (see Mellish et al. 1998). ILEX is at 
present generating descriptions in the museum 
domain, in particular, that of 20th Century jew- 
ellery. 
Our focus on this topic has grown out of 
the need to integrate two strands of research 
within ILEX. One strand involves the work on 
anaphora by Janet Hitzeman. She implemen- 
ted a module to construct contextually appro- 
priate referring expressions within ILEX, based 
on Centering Theory (Grosz et aI. 1986). See 
Hitzeman et al. 1997. 
The second strand involves the aggregation 
module (implemented by Hua Cheng, see Cheng 
et al. 1997). The task of this module is to re- 
package discrete informational units into single 
complex sentences. She is presently exploring 
the aggregation of information into the NP, for 
instance this gold and silver ring, designed by 
King. 
These two functions, the referring and the in- 
forming, interfere with each other, to the extent 
that each wishes to control the construction of 
the NP form. These tasks thus need to pay 
regard to each other, and this paper, and the 
implementation it describes, are an attempt to 
answer this need. 
Appelt's approach seems to be to build an NP 
46 
~.. O~...O-. Q. O. Q cl 
1 • . • • u • uul u S 
• • • I a i ~ FACTS 
R~.-ATH)N$ 
Figure 1: The Information Graph 
for referring, then either modify the elements 
(e.g., substitution of the head noun) or fill un- 
used structural slots with non-referring inform- 
ation. However, we have found that the two 
tasks of referring and informing can be more 
highly integrated, with each decision within the 
construction of the NP taking into account the 
needs of both tasks, rather than satisfying the 
referring function first, then looking to the in- 
forming function. In other words, we follow an 
integrated rather than pipeline approach. 
Section 2 will describe how information is rep- 
resented in ILEX. Section 3 describes the inter- 
face between the text-planner and the NP- plan- 
ner, the input specification for the NP-planner. 
Section 4 discusses the syntactic structure of the 
NP, and which syntactic positions allow inform- 
ing without interfering with referring. Section 5 
details the referring expression generator which 
integrates referring and informing goals. An ex- 
ample of the generation process is given in sec- 
tion 6 and section 7 summarises. 
2 Information Representation 
To properly describe our NP-planning process, 
we need to describe how information is represen- 
ted in ILEX. Domain knowledge is represented 
in terms of an information graph, which rep- 
resents the entities of the domain, the inter- 
relation between these entities (called facts); 
and the relations between these facts (e.g., a 
causal relation between two facts). Figure 1 
shows an abstract representation of an informa- 
tion graph. At present, relations between facts 
are not used in tim NP-planner, so will not be 
discussed further here. 
Initially, the information graph representa- 
tion was developed for text planning. However, 
following a suggestion from Alistair Knott, we 
have found it useful to use it for NP-planning 
as well. 
2.1 Entities 
Entities represent the objects of the domain. In 
the Museum domain, this includes not only the 
museum artifacts, but their designers, the ma- 
terials they are made from, the styles they are 
made in, the periods they belong to, the loca- 
tions of their manufacture, etc. 
Entities are typically specific entities: real- 
world individuals. However, some of the entit- 
ies will be generic entities, those representing 
classes of entities, such as Art-Deco jewellery. 
We also cater to sets of entities, which can be 
realised through either plural anaphora, or con- 
junctive NPs. 
2.2 Facts 
A fact in ILEX represent a relation between two 
entities. These relations may be processual, e.g., 
that X made Y: maker(J-999, King01); or 
stative (e.g., that X is a type of Y: isa(j-999, 
generic-brooch). 
Each fact is represented as an attribute-value 
structure as below: 
P red: "maker" 
Argl: J-999 
Arg2: King01 
Polarity: positive 
Status: indefeasible 
Assimilation: 0 
Importance: 6 
Interest: 8 
Note that apart from the predicate and ar- 
gument information, several other fields qualify 
the informational status of the fact, including 
the polarity (whether or not the relation holds), 
and de feasibility (distinguishing between hard 
facts about the entity, and those which are only 
tendencies, e.g., Art-Deco jewellery tends to be 
made of enamel (see Knott et al. 1997 for dis- 
cussion of defensibility in ILEX). The remaining 
fields, having a stronger affect on NP-planning, 
include: 
• Assimilation: the degree to which the 
system considers the user to have under- 
stood the information. This is of particu- 
lar importance to reference, since adequate 
reference usually requires the user to know 
the information used for reference (see later 
for exceptions}. 
47 
Importance/Interest: the degree to 
which the fact is considered important for 
the system to deliver to the user, and the 
system's estimate of the degree of interest 
to the user. These values are represented 
for each predicate type as a whole, and vary 
for different user models. These values are 
used when selecting the facts to use to pro- 
duce a unique reference. 
3 NP Specification 
One of our goals in the design of the ILEX NP- 
planner was to provide a clean interface between 
text-planning and NP-planning, such that the 
text planner can specify what it wants from 
the NP without needing to know about syntax 
at all. To this end, we have developed a two- 
level specification of the NP, one at the semantic 
level, and one at the syntactic level. The text- 
planner specifies the NP only at the semantic 
level, leaving details of syntax totally to the NP- 
planner. 
3.1 The NP Specification Interface 
The interface between the text-planner and the 
NP-planner is in the form of an attribute-value 
matrix, the attributes of which are: 
1. Cat: the function of the NP being pro- 
duced. The NP-planner allows a wide 
range of NP functions, not only referring, 
shown in figure 2 and discussed below: 
(a) Referring: an NP which uniquely or 
non-uniquely refers to the referent. 
More delicate options can be specified, 
such as refer-by-name, refer-by-type, 
or refer-by-pronoun; and also whether 
the reference should be unique or not. 
(b) Describing: an indefinite NP giving an 
arbitrary number of the entity's attrib- 
utes without attempting to be definit- 
ive. 
(c) Classifying: an indefinite NP which 
provides only the superclass of the 
item, e.g., this is a brooch 
(d) Defining: for generic entities, an NP 
which provides the entities defining 
characteristics, e.g., a necklace is an 
item ofjewellery worn around the 
neck 
. 
. 
. 
. 
. 
(e) Eliciting: a wh- NP for the referent. 
Eliciting can be selective, e.g., which 
designer or non-selective, e.g., Who. 
If referring-np is selected, various sub-types 
of reference can also be preselected through 
this slot (the specification of the Cat can be 
logically complex). 
If no preselection is made by the text- 
planner, the system will decide NP function 
on the basis of constraints and defaults. 
For instance, in the usual case, the text- 
planner will specify only referring-np, and 
leaves it up to the NP planner to choose 
the exact type of reference. 
In some cases, the system will override 
the preselection if it is incompatible with 
the referring environment. For instance, 
if the text-planner specifies refer-by-name, 
but no name information is provided for the 
entity, then a choice between refer-by-type 
and refer-by-pronoun will be made. The 
NP-specification can thus be seen to offer a 
flexible interface, allowing the text-planner 
to determine the amount of control it de- 
sires to assert. 
Sem: the referent of the NP, an entity in 
the information graph (or a set of such en- 
tities if desired, realised as either a coordin- 
ated NP, or using plural anaphora). 
Syn: the slot to be constructed by the NP- 
planner, a syntactic structure. 
Orth: the slot to hold the eventual sur- 
face string for the NP. If the sentence plan- 
ner provides a filler for this slot, then NP- 
planning is avoided and the string is used. 
ILEX thus allows canning of NPs when 
needed. 
Agenda: a list of fact-ids which are to be 
incorporated into the NP if possible. The 
aggregation module uses this slot to state 
its requirements from the NP, which facts 
the NP is to express. 
Restrictions: a list of fact-ids which 
should not be used in the NP. For instance, 
we might wish to avoid generating the sen- 
tence Other jewels designed by Jessie M. 
King include a brooch designed by Jessie M. 
King. To avoid such sentences, we place 
48 
-descnbing-np 
-classifying-np -defining-np 
NP-N~ JDE- --elicidng-np 
np-node ~/t - referring-np 
EL|CrrATION- r selective-elicitafi°n 
SE~.,.SCTWtTf ~ nonselecfive-elicimtion 
REEERF, J~CE-s.rRATEGy f nonposidonal-rcfercnce 
~,- posi\[ional-re fcrence 
NON~ Isrn{,NAL- REIF,.,R ~CE -SI'RA'r~; Y 
,_F 
  - refer-by-name... 
t... refer-by-pronoun 
Figure 2: System Network for Nominal Function 
the fact-id of the 'designer' fact into the 
Restrictions field for the mention of the 
brooch. The designer fact will not then be 
used as a referential restrictor. 
A sample NP-specification is shown below: 
Cat: unique-reference" 
Sem: J-999 
Agenda: (FN-48 FN-56) 
Restrictions: (FN-59) 
...where FN-48 and FN-56 are facts to include 
in the reference, and FN-59 is a fact to avoid. 
Problems of Modularity: One of the 
problems of a clean separation between NP- 
specification and NP-planning is that it might 
not be possible to incorporate all facts on the 
informing agenda into the NP. However, given 
that NPs syntactically allow any number of 
non-referring post-modifiers, our planner will 
handle any arbitrary number of facts on the 
agenda. However, in terms of intelligibility, too 
many post-modifiers will produce unintelligible 
NPs. We make the simplifying assumption that 
the text planner uses some metrics (simple or 
complex) to avoid overloading the NP. We are 
merely providing a mechanism to support NP- 
planning once the agenda is specified. 
3.2 Focus Spaces 
Apart from the text-planner's specification of 
the NP-form, the text-planner also maintains 
some variables concerning the current referring 
environment, mainly in terms of various focus 
spaces. These various spaces are: 
. 
. 
. 
Mentioned Entities: entities mentioned 
at some point within the discourse. 
Shared Entitles: entities which the 
system assumes the addressee to know 
about. These entities include world- 
knowledge (e.g., Ronald Reagan), but also 
entities mentioned previously in the dis- 
course (mentioned-entities), and entities in 
the immediate context (focal-objects). En- 
tities in this space are potential confusers 
for definite reference with the. 
Focal Objects: the focal space includes 
a set of entities which may potentially be 
referred to as this z. Firstly, we have the 
Prior-Cb (backward looking centre, usually 
the subject of the prior sentence). Entities 
directly related to this may also be focal. 
This is also called the local focus in our 
system. Then there is the page-focus, the 
focus of the current object description in 
the ILEX system, e.g., this brooch. Other 
objects are also focal by being part of the 
immediate context of the reader/writer. In 
a web-browsing environment, this might in- 
clude the current page (this page), or parts 
of the page (this picture). 
49 
In addition to the focal spaces, there are also 
variables holding individual focal objects, in- 
cluding the Cb and Prior-Cb, Page-Focus and 
Discourse-Focus. We allow pronominalisation 
only when the object being referred to is Prior- 
Cb, which seems to produce coherent reference. 
4 NP Structure for Referring and 
Informing 
The NP-Planner has distinct procedures for 
each of the NP functions, one for classifying, one 
for referring, one for eliciting, etc. Due to lack 
of space, we will focus from now only on NPs 
primarily serving a referring function. Other 
nominal functions will be covered in a later pa- 
per. 
The issue of how to construct a referring NP 
is well explored. The issue remaining is how 
non-referring information can best be incorpor- 
ated into referring NPs. This section will look at 
the locations in the NP which can express non- 
referring information, without interfering in the 
reference of the NP. The following section will 
describe the algorithm which allows referring 
and informing to be integrated. 
4.1 Nucleus and Satelite Composition 
of the NP 
We consider the basic structure of the NP to fall 
into two components: a nucleus, which performs 
the nominal function of the NP, and optional 
satelites, where additional information can be 
placed. 1 The nucleus of the NP consists of all 
slots before the head, and the defining post- 
modifying slots (e.g., defining relative clauses, 
or prepositional phrases). The satelite elements 
are typically realised by indefinite NPs, or non- 
defining relative clauses (or complexes of such), 
e.g., \[Nue: this brooch\], \[Sat: designed by Jessie 
M. King\], or \[guc: King\], \[Sat: a Scottish de- 
signer\]. 
In our model, all referring is performed by 
the nucleus - the satelite(s) are non-defining, 
i.e., perform only an informing role. However, 
as will be discussed below, the nucleus can also 
contain non-referring information. 
1Nucleus and sateli£e are terms taken from RST 
(Rhetorical Structure Theory, e.g., Mann & Thompson, 
1987), although usually applied to the relations between 
sentences. 
4.2 The Structure of the Nucleus 
The range of slots in a systemic analysis of the 
NP, in the order they typically appear (after 
Halliday 1994), appears below, and figure 3 
shows a typical NP structure: 
(Deictic) " (Numerator) " (Epithet*) " 
(Classifier*) " Thing " (Qualifier*) 
Key: () - optionality 
* - any number of this slot may 
occur 
4.3 Informing within the Nucleus 
While primarily for referring, non-referring in- 
formation can sometimes be included in the nuc- 
leus without interfering with the referring func- 
tion. For instance, we can add information to an 
already uniquely-referring NP, making the ref- 
erence more explicit: The \[granny smith\] apple 
on the table; this \[enjoyable\] book. The degree 
to which informing and referring can be so in- 
tegrated varies from domain to domain. 
The major constraint we seem to face is that 
there is a degree of expectation under conversa- 
tional implicature that the speaker refers using 
information known to the addressee (see Dale 
& Reiter 1996). Thus, in a situation where 
only one apple is visible, if I say pass me the 
Spanish apple, the addressee might be confused 
by the inclusion of the superfluous information, 
and perhaps think there must be another apple 
somewhere. 
However, in some registers this form of ref- 
erence seems to offer no problems. Appelt 
(1985) mentions the case of the speaker pointing 
at some implement and saying use the wheel- 
puller. The addressee, not knowing the name, 
but having the item identified through point- 
ing, accepts the naming. We thus have an NP 
whose head-noun is not serving a referring func- 
tion, but rather an informing function, since the 
referring function was otherwise fulfilled. 
The newspaper genre is particularly strong on 
this type of reference, as shown by the newspa- 
per article below: 
Student fights for life after flat fire: 
A young student was today fighting for 
50 
"Cat: 
Deictic: 
Numerator: 
Epithet: 
Classifier: 
Thing: 
Qualifier: 
nominal-group I 
Cat: definite-determiner| 
Lex: the-det \[ 
Orth: "the" j 
I Cat: ordinal-adjective 1 Lex: seven-adj | Orth: "seven" J 
"Cat: adjectival-group 
"Cat: intensifier\] 
Intensif: Lex: very-adv | 
Orth: "very" J 
"Cat: adjective' 
Head: Lex: large-adj 
Orth: "large" 
"Cat: nominal-group 
Cat: adjective" 
Epithet: Lex: used-adj 
\[Orth: "used" 
'Cat: noun 
Thing: Lex: car-noun 
\[Orth: "car" 
noun 
saleman-noun 
"salesman" 
Cat: 
Lex: 
Orth: 
"Cat: prep-phrase 
'Cat: preposition" 
Marker: Lex: from-prep 
\[Orth: "from" 
Cat: nominal-group 
Head: \[Thing.Orth: "Perth" 
Figure 3: A Sample NP Structure 
her life after fire ripped through her Ed- 
inburgh flat. Nicola Graham is in a "seri- 
ous but stable" condition at the specialist 
burns unit in St John's Hospital, Living- 
ston. Firefighters suspect the blaze may 
have been started by a dropped cigar- 
ette in Miss Graham's bedroom. The 19- 
year-old was transferred from Edinburgh 
Royal Infirmary to St John's for emer- 
gency burns treatment .... " 
The sequence of references to the student suc- 
cessively add new information: A young stu- 
dent: Age and occupation; Nicola Graham: 
Name; Miss Graham: Marital status; the 19- 
year-old: Age. This writer is not depending on 
assimilated information to refer, but, depending 
on the lack of potential confusors, is successfully 
referring with new information. While this style 
is more typical of newspaper reporting, where 
compact information delivery is important, it is 
still an issue which needs to be addressed in any 
NP-planner. 
In the register of museum object descriptions, 
it seems that the degree to which new inform- 
ation can be included in the nucleus is limited. 
New information seems not to be appropriate in 
the Deictic, Classifier, Thing or Qualifier slot, 
but is generally allowed in the Numerative and 
Epithet slots. This makes some degree of sense, 
since these slots are the least restrictive. The 
Numerative can be used restrictively when used 
contrastively, e.g., the five cups (butnot the set 
of three), but this is rare. Epithets generally 
add qualitative information, and are thus less 
restrictive. 2 
Another approach is to examine the se- 
mantic types of pre-modifier elements, to see 
which, when inserted for informing reasons, 
seem to interfere with the referring function. 
We have found some of our fact-predicates in- 
terfere more, some less. As a result of this, 
we maintain a list of fact-predicates which are 
judged, for the current domain, to be suitable 
for pre-modifier slots without interfering with 
reference. This allows us to produce, for in- 
stance, this \[important\] designer; the |gold and 
enamel\] brooch designed by King; the \[quite in- 
fluential\] Art-Deeo style. 
! 
5 The Planning Algorithm 
The tension in planning definite reference de- 
rives from the need to serve both the referring 
and informing functions. The referring function 
is mainly concerned with the Sere slot of the 
NP-specification: the task is to point uniquely 
at the filler of this slot, distinguishing it from 
all other entities. The planner may need to use 
any or all of the syntactic slots of the nucleus 
2A study of 20 randomly chosen museum descriptions 
, undertaken by Cheng, from four museums and galleries 
revealed that only 1/3 of Epithets act restrictively. 
51 
to do this. 
On the other hand, the informing function is 
primarily concerned with the Agenda slot - the 
function is satisfied if all the facts in this slot are 
expressed somewhere in the NP (whether nuc- 
leus or satelite). While these facts can be placed 
in satelite position, it is often more coherent to 
place them within the nucleus. 
As such, the two functions are in competition 
for the syntactic slots, and structural decisions 
good for one function may be sub-optimal for 
the other. The usual approach is to allow the 
referring component to go first, generating the 
desired referring form. Then, the needs of the 
informing component are fitted into this struc- 
ture. 
However, we have found it far more rewarding 
to allow all syntactic decisions to be mutually 
negotiated between the two functions. Below, 
we describe the definite description algorithm 
as used in ILEX. 
5.1 Construction of NP 
The steps of building the NP are as follows: 
Build Nucleus: Since we wish to choose 
a referring expression which opportunistically 
serves some of the informing function, we will 
build the nucleus on the basis of i) the refer- 
ential context, and ii) the agenda of things to 
say: 
1. Location of Assimilated Agenda: the facts 
on the agenda need not all be new inform- 
ation (unassimilated) - the text-planner 
may place previously given information on 
the agenda, perhaps for some pragmatic 
reason, e.g., to ensure that the addressee 
is aware of some fact at this point of the 
discourse. 
Assimilated facts on the agenda play an im- 
portant role in our algorithm, since we will 
use them preferentially for referring. 
2. Choose Referential form: we need to choose 
between refer-by-name, refer-by-pronoun 
and refer-by-type, on basis the referential 
context. Where the referential context al- 
lows more than one choice, we refer to the 
assimilated-agenda to help. For instance, 
if a Name fact is the agenda (and assim- 
ilated), we might favour refer-by-name, if 
Gender is the sole fact on the agenda, pro- 
nominal might be favoured. With several 
(assimilated or unassimilated) facts on the 
agenda, a common-group is preferred since 
it offers more opportunities for inclusion of 
facts (although proper-np expression also 
supports non-referring post-modification). 
3. Choose Head Noun: for common noun- 
phrases, the head noun will be chosen from 
the most specific assimilated isa fact about 
the object. However, an assimilated isa 
fact on the agenda is allowed to override 
the default. 
4. Choose Determiner. for common noun- 
phrases, the determiner will be chosen on 
the basis of the objects focal status (e.g., 
this if focal, the otherwise). If a fact spe- 
cifying the owner of the entity being ex- 
pressed is on the agenda (and assimilated), 
then a genitive deictic will be used. 
5. Choose Restrictive Modifiers: if the 
common-noun-phrase form was selected, 
then we need to determine which modifi- 
ers are to be included to produce a unique 
reference. For instance, if we have the Style 
fact of a jewel on the agenda, and it is as- 
similated, then that will be preferred as a 
restrictive modifier. See below for more de- 
tail. 
6. Fill in Unused Slots: When we have a func- 
tioning referring form, then we can add in- 
formation from the agenda into the unused 
slot, e.g., this book -b enjoyable ~ this en- 
joyable book. The system is provided with 
a list of fact-predicates which can be ex- 
pressed in pre-modifier slots, e.g., in the 
Jewellery domain, Materials - this gold and 
enamel brooch, Fame- the famous designer 
called Jessie M. King. 
Add $atelites: Any information which was 
not consumed in the nucleus can now be placed 
into non-defining satelites, e.g., \[Nuc: Jessie M. 
King\], \[Sat: a Scottish designer\]. 
5.2 The Confusor Set 
The confusor set is the set of entities which a 
partially constructed NP unintentionally refers 
to. For instance, if we have only selected the 
head noun, brooch, then the confusor set is all 
52 
brooches known to the system apart from the 
intended referent. 
The process of building a referring expression 
can be seen as successively reducing the con- 
fusor set until it is empty. Assuming a common- 
np, the steps in this reduction are: 
1. Set Initial Confusors based on focus status: 
• Sere is member of Focal-objects: 
Deixis: proximal (this/these). Con- 
fusors: Focal-objects, without Sere. 
• Entity has an assimilated Owner fact: 
Deixis: Owner. Confusors: other en- 
tities owned by Owner. 
• Entity has been mentioned already on 
this page: Deixis: nonproximal (the). 
Confusors: Other entities mentioned 
on page. 
• Default: Deixis: nonproximal (the). 
Confusors: Shared-Entities. 
2. Restrict on class: Choose a head noun for 
the item, and eliminate all confusors which 
do not take the class. 
3. Add Restrictive Modifiers: Choose a sub- 
set of assimilated facts which eliminates all 
confusors (see next sub-section). 
4. Insert "One of" if needed: (not yet imple- 
mented) If insufficient assimilated facts to 
eliminate all confusors, insert "one of' or 
"another off into the Pre-Deictic slot. "an- 
other of' is used if the confusor is already 
mentioned on the page. 
5.3 Choosing Restrictive Modifiers 
There are a number of strategies used to se- 
lect the optimal set of restrictive modifiers to 
produce unique reference. There seems to be 
two main approaches. One attempts to select 
the smallest subset of modifiers which uniquely 
refers (e.g., Reiter 1990; Dale 1989). A solution 
which offers better ~omputational complexity is 
based on the premise that some fact-types are 
better suited as restrictive modifiers than oth- 
ers, and thus restrictive modifiers are chosen by 
incrementally taking the next modifier from the 
list (e.g., Reiter & Dale 1992). 
In ILEX, we follow the incremental approach, 
adding restrictors in order from our (domain- 
dependent) list (but only if the restrictor elim- 
Information 
Class: apple 
Owner: John 
Color: Red 
Variety: Granny Smith 
Position: on table 
Taste: good 
On Agenda? 
no 
yes 
no 
yes 
yes 
no 
Assim.? 
yes 
no 
yes 
no 
yes 
no 
Figure 4: An Example Information Base 
inates some confusors). We have found that or- 
dering restrictors in terms of goodness, the NPs 
we generate are of better quality. 
The need to integrate informing into the pro- 
cess changes the process slightly. As stated 
above, the text-planner is allowed to place as- 
similated, as well as unassimilated, informa- 
tion on the agenda. If this has not happened, 
then we use the standard incremental strategy. 
However, if the text-planner has placed assimil- 
ated information on the agenda, then our plan- 
ner places these at the front of the preferred- 
restrictors list. 
We note however, that there are cases where, 
while the text-planner may want the fact to be 
included, the fact is not a suitable restrictor. 
For instance, including the place-of-wearing fact 
on the agenda could result in an NP like the gold 
necklace that is worn around the neck. However, 
since the place-of-wearing does not actually dis- 
criminate (given all necklaces are worn around 
the neck), the fact was not used restrictively, 
and was later realised in a satelite of the NP, 
e.g., the gold necklace, which is worn around 
the neck. However, there may be facts which 
are partially restrictive, but nevertheless poor 
candidates for restriction. Our algorithm does 
not cater to these cases as yet. : 
6 An Example 
Agenda: Assume we are talking about an 
apple, and have the information as in figure 4 to 
express. In short, the facts on the agenda are: 
Owner, Variety, and Position. 
Referential Context: Assume also that we 
have several red apples, but only one on the 
table. The apple above has been mentioned, 
but not for a while, with other apples mentioned 
since. 
Stage 1: Building the Nucleus: 
53 
1. Choose Referential form: Since the item is 
not the Cb, we cannot use a pronoun. Since 
it doesn't have a proper-name, proper-noun 
reference is also out. We axe forced to use 
a common noun-phrase. 
2. Choose Restrictive Modifiers: We have 
a set of potential referential restrictors 
of: (Class Owner Color Variety Position 
Taste). Of these, we can only refer using as- 
similated roles, so we can use: (Class Color 
Position). We also have the agenda role-list 
of: (Owner Variety Position). of which the 
assimilated items axe: (Position). Since the 
Class fact is assimilated, we automatically 
take the class as the head of the referring 
NP, e.g., the apple. This is not however 
unique, so we need to add more restrictions. 
We use the first (and only) item in the as- 
similated agenda: Position: the apple on 
the table. This happens now to be unique, 
so we have a functional referring NP. 
3. Fill in Unused Slots: This leaves two facts 
unexpressed: Owner and Variety. The 
Owner predicate can normally be expressed 
in one of two slots of the nucleus: 
• the Deictic slot e.g., John's apple on 
the table; or, 
• the Qualifier slot (after the Head noun, 
e.g., the apple that John owns on 
the table. (I assume here that non- 
restrictive relative clauses are always 
satelites, discussed below). 
In both of these slots, the inclusion of unas- 
similated Owner information seems to mess 
up the reference, seemingly because it im- 
plies the reader should already know the 
ownership. We thus leave the Owner role 
for expression in a satelite position (realised 
as a non-restrictive relative clause, e.g., the 
apple on the table, which john owns). 
The variety fact can be realised best 
through the Classifier slot, e.g., the Granny 
Smith apple on the table. This does not 
seem to interfere with the referring func- 
tion, so this fact-type would occur on our 
list of facts which can appear in a pre- 
modifier slot without interfering with the 
referring function. 
This stage thus ends with the referring slot 
consisting of: the Granny Smith apple on the 
table. We have only one item left on the agenda, 
the Owner fact. 
Stage 2: Adding Satelites The Owner fact 
can be incorporated into the NP as a satelite (as 
a non-referring relative clause), e.g., the Granny 
Smith apple on the table, which John owns. 
7 Conclusions 
We have improved on the integration of refer- 
ring and informing within NP generation by al- 
lowing informing to influence decisions at each 
stage of the formation of the referring form. 
Previous np-generation systems only satisfy in- 
forming goals after the referring form has been 
determined. 
The points of intervention in the referring 
process include: the selection of the form of the 
NP; the choice of the Deictic in common NPs; 
and choice of restrictive modifiers. Information 
remaining on the agenda at this point is ex- 
pressed in non-referring slots of the NP, in pax- 
ticular, the Epithet slot, or non-referring post- 
modifier slots. The use of an Agenda slot in 
the NP-specification is the main addition, which 
allows the Aggregation component to interface 
with the referring expression generator. 
8 Acknowledgements 
The ILEX project is supported by EPSRC grant 
GR/K53321. Thanks to Alistair Knott for ideas 
contribing to the implementation, and to Ren- 
ate Henschel for valued comments on this paper. 

References 
Appelt, D E. 1985. "Planning English Referring 
Expressions". Artificial Intelligence, 26, pp 1- 
33. 
Appelt, D and Kronfeld A. 1987. "A Compu- 
tational Model of Referring". In Proceedings 
off the Tenth International Joint Conference on 
Artificial Intelligence, Milan, Italy, August 23- 
28, 1987, pp 640-647. 
Cheng, Hua,& Chris Mellish. 1997. "Aggrega- 
tion in the Generation of Argumentative texts". 
Proc. of PhD Workshop on Natural Language 
Generation, 9th European Summer School in 
Logic, Language and Information (ESSLLI97). 
Aug. 1997, France. 
Dale, Robert. 1988. Generating Referring 
Expressions in a Domain of Objects and Pro- 
cesses. Ph.D. Thesis, Centre for Cognitive Sci- 
ence, University of Edinburgh. 
Dale, Robert. 1989. "Cooking up referring ex- 
pressions". Proceedings of A CL-89. Vancouver, 
pp 68-75. 
Dale R. and E. Reiter. 1996. "The Role 
of the Gricean Maxims in the Generation of 
Referring Expressions". Working Notes for 
the AAAI Spring Symposium on Computational 
Implicature, Stanford, 1996, pp 16-20. 
Grosz, Barbara J., Aravind K. Joshi and Scott 
Weinstein. 1995. "Centering: A Framework 
for Modeling the Local Coherence of Discourse". 
Computational Linguistics, Volume 21, Number 
2, June 1995, pp 203-225. 
Halliday, M.A.K. 1994. Introduction to Func- 
tional Grammar. 2nd edition. London: Edward 
Arnold. 
Hitzeman, Janet, Chris Mellish & Jon Ober- 
lander. 1997. "Dynamic Generation of Mu- 
seum Web Pages: The Intelligent Labelling Ex- 
plorer". Proceedings of the Museums and the 
Web Conference, Los Angeles, March 1997. 
Horacek, Helmut. 1995 "More on Generating 
Referring Expressions". Proceedings of the 5th 
European Workshop on Natural Language Gen- 
eration. Leiden, The Netherlands. 
Knott, Alistair, Michael O'Donnell, Jon Ober- 
lander, Chris Mellish. 1997. "Defensible Rules 
in Content Selection and Text Structuring". 
Proceedings of the 6th European Workshop on 
Natural Language Generation. March 24 - 26, 
1997 Gerhard-Mercator University, Duisburg, 
Germany. 
Mann, William & Sandra Thompson, 1987. 
"Rhetorical Structure Theory: A Theory of 
Text Organization". Technical Report ISI/RS- 
87-190. 
Mellish, C., O'Donnell, M., Oberlander, J. and 
Knott, A. 1998 "An architecture for opportun- 
istic text generation". Proceedings of the 9th 
International Workshop on Natural Language 
Generation. 5-7 August 1998. Prince of Wales 
Hotel, Niagara-on-the-Lake, Ontario, Canada. 
Reiter, E. 1990 Generating appropriate natural 
language object descriptions. PhD Thesis, Hat- 
yard University. 
Reiter, E. and Dale R. 1992 "A Fast Algorithm 
for the Generation of Referring Expressions". 
Proceedings of COLING-92. Nantes, 1992. 
