Cb or not Cb? Centering theory applied to NLG 
Rodger Kibble 
Information Technology Research Institute 
University of Brighton 
Lewes Road 
Brighton BN2 4GJ 
U.K. 
Rodger. Kibble@itri. brighton, ac. uk 
Abstract 
Centering theory (CT) has been mostly dis- 
cussed from the point of view of interpretation 
rather than generation, and research has tended 
to concentrate on problems of anaphora resolu- 
tion. This paper examines how centering could 
fit into the generation task, separating out com- 
ponents of the theory which are concerned with 
planning and lexical choice. We argue that it is 
a mistake to define a total ordering on the tran- 
sitions CONTINUE, PJZTAm, SHIFT and that they 
are in fact epiphenomenal; a partia/ordering 
emerges from the interaction between cohesion 
(maintaining the same center) and salience (re- 
all.qing the center as Subject). CT has generally 
been neglected by NLG practitioners, possibly 
because it appears to assume that the center is 
determined according to feedbsch from the sur- 
face grammar, to text planning, but we argue 
that this is an artefactual problem which can 
be eliminated on an appropriate interpretation 
of the CT rules. 
1 What is Centering? 
Centering theory (C~) is a theory of discourse 
structure which models the interaction of cohe- 
sion and salience in the internal organlsation of 
a text. The main assumptions of the theory as 
presented by (Gross et a11995 (GJW), Brennan 
et al 1987) rare: 
1. For each utterance in a discourse there is 
precisely one entity which is the centre of 
attention or center. 
2. There is a preference for consecutive utter- 
ances within a discourse segment to keep 
the same entity as the center, and for the 
most salient entity ~realised n in an utter- 
ance to be interpreted as the center of the 
following utterance. 
3. The center is the entity which is most likely 
to be pronominalised. 
These principles will be more precisely expli- 
cated in Sect. 2. 
CT has proved attractive to NLP researchers 
because of the elegance and simplicity of the 
core proposals; it provides a framework for 
analysing text without having to make tough 
decisions about what a partfcular utterance is 
"about", since all notions are defined in purely 
structural terms. 
Much research in CT has concentrated on in- 
terpretation, particularly reference resolution, 
developing algorithms to resolve anaph0ric ex- 
pressions based on the assumption that the 
text is constructed according to Rules 1 and 
2. So researchers have focussed on filling in de~ 
tails of the theory which were left unspecified: 
what counts as an utterance, and how should 
transitions be handled in complex sentences 
(Kameyama 1998; cf Suri and McCoy 1994)? 
how is salience ranking determined (Gordon et 
al 1993; Stevenson et al 1994;. Strube and Hahn 
1996)? what counts as ~r~Migstion ~ - does this 
include bridging references (Strube and Hahn 
op cit.)? how do centering tra~L~itions relate 
to discourse segment boundaries (Walker 1998, 
Passoneau 1998)? 
In fact I will leave many of these issues aside 
for the purposes of this paper; I will not ex- 
amine the empirical adequacy of CT, for which 
the reader is referred to papers cited above and 
• others collected in Walker et al (1998). I will 
take a different approach, which is to examine 
how the dements of CT can be applied to the 
planning of texts, with the rules and constraints 
interpreted as instructions to a generator rather 
than a guide for interpretation. To avoid intro- 
ducing too many complications I shall assume 
72 
Cb(Un)' = Cp(U.) 
Cb(Un) ~ Cp(Un) 
Cb(Un) =" Cb(Un-l) 
or Cb(Un-.l) undefined 
Continue 
Retain 
Cb(Un) ~ Cb(Un-\[) 
Smooth Shift 
Rough Shift 
Figure 1: Centering Transitions 
Constra/nts " 
C1. There is precisely one Cb. 
C2. Every element of Cf(Un) must be realised inUn. 
C3. Cb(U.) is the highest-ranked element of C/(U.-I) that is realised in Un. 
Rules 
RI. If some element of Cf(Un-l) is realised as a 1~ronoun in/.)', then so is Cb(Un). 
(Strong version: if Cb(Un) = Cb(Un-1), a pronoun should be used.) 
112. Continue is preferred over Retain which is preferred over Smooth Shift 
which is preferred over Rough Shift ..... 
@ 
O 
O 
O 
O @ 
O 
O 
O 
O 
O @ 
O 
O 
O 
O 
O 
O 
O @ 
O @ 
@ 
@ 
O 
Figure 2: Centering 
a Ucanonical ~ formulation of CT as outlined 
by Walker et al. (1998, Chapter 1) and the 
schematic ~consensus ~ generation architecture 
described by Reiter and Dale (Reiter 1994; Re- 
iter and Dale 1997). This consists a ~pipelins" 
of distinct tasks: 
Text Planning- deciding the content of a 
message, and organising the component 
propositions into a text tree; 
Sentence Planning - aggregating proposi- 
tions into clausal units and choosing lex- 
ical items corresponding to concepts in the 
knowledge base, including referring expres- 
sions (RE); 
Linguistic re_8!isation which takes care of 
surface details such as agreement, orthog- 
raphy etc. " 
Previous researchers have implemP.nted 
pronominalisafion decisions using CT, and 
so have located Centering as part of RE 
generation (e.g., Dale 1992, Passoneau 1998), 
while Mittal et al (1998) have a ~centering 
module" which forms part of Sentence Planning 
and seeks to realise the center as Subject in 
successive sentences. In what follows I will 
try to separate out the tasks which make up 
centering theory and argue that the way to 
implement CT is not as a discrete module but 
as a series of constraints on the various levels 
of the generation process from Text Planning 
Constraints and Rules 
to RE generation. I shall also briefly note 
points of comparison with systems discussed by 
CAhill and Reape (1998) in a survey of applied 
NLG systems, and conclude with some remarks 
on the applicability of my proposals to the 
"reference architecture n envisaged by Cahill et 
al. (1999), RAGS (1999). 
2 TrAnsition rules 
The rn~in clahns of CT are formalised in terms 
of C5, the "backward-looking center ~, C/, a 
list of ~'orward-looking centers z for each ut- 
terance Un, and Cp or ~preferred center z, the 
most salient candidate for subsequent utter* 
ances. Cf(Un) is a partial ordering on the en- 
tities mentioned (or ~lised n) in Un, ranked 
by grammatical role, e.g. SUBJ ) DIR-OBJ > 
INDIR-OBJ ~> COMP(S) ~> A.DJUNCT(S). C~(Un) 
is the highest ~,,ked member of C/(U,J (usu- 
ally susJ), and is predicted to be Cb(U,+~). 
C.f is partial/), ordered in most accounts, which 
leaves open the possibility that there is no 
-nlque Cb. AJSO, if two successive utterances 
have no referent in common the second Will ha~ 
no Cb. 
Successive pairs of utterances are charac- 
terised in terms of tmns/t/on.~ as defined in Fig- 
ure 1; for instance if two consecutive utterances 
have the same Co, and the Cb in the second ut- 
terance occurs in Subject position, this is classi- 
fied as a CONTINUE transition. A text is judged 
to be coherent to the extent that transitions fol- 
low the preference ordering given in Rule 2 (Fig 
73 ¸ 
2); and on the assumption that the text is coher- 
ent, pronominalisation is predicted to conform 
to Rule 1. 
The notion of "realisation" is subdivided into 
"direct' and "indirect": an entity is directly re- 
alised in an utterance if it is the denotation of an 
overt NP (or a zero pronoun where this is syn- 
tactically licensed), while "indirect realisation" 
covers for example subsectional anaphora, pos- 
sessive pronouns and inferential links such as 
bridging reference. Corpus~based investigations 
of CT have tended to concentrate on direct re- 
alisation, since ~nnotation of indirect anaphoric 
links depends on theoretically-based decisions 
and it may be difficult to achieve reliability in 
this area. This has obvious implications for the 
resulting measure of coherence, which I return 
to below. 
2.1 Salience and cohesion 
Transitions are defined in terms of tests which I 
shall call cohesion: Cb(Un) = Cb(U~-l), and 
salience: ~(Un) = ~v(Un). There are four 
possible combinations, which are displayed in 
Fig. 1. The most preferred case is where both 
apply, namely co~Imm, and the least pre- 
ferred is where neither apply, ItOUQH S~. For 
the intermediate cases there are three logical 
possibilities: prefer cohesion (It,"rAm), prefer 
salience (SMOOTH SHIFt) or allowboth equally. 
There is no obvious way to settle this a pr/or/, 
but Walker et al (1998, Ch. 1) ~ipulate that 
~"rAm is preferred over SMOOTH SHIFT. Evi- 
dence for this position is not conclusive, and in- 
deed di Eugenio (1998), Passoneau (1998) and 
Hurewitz (1998) all found a higher percentage 
of shifts than retains. This suggests either that 
salience is a stronger requirement than cohesion 
or that it.is easier to satisfy. "That is, the l~n. 
guages studied (English and Italian) may be suf- 
ficiently flexible that there is usually some way 
to realise 6'b as Subj (or first-mention) but on 
the other hand the same 6'b can only be main- 
tained for a- finite n,,mher of utterances. 
I suggmtthat these results should be treated 
with some caution since it is not dear that the 
authors have the same assumptions about the 
claims of CT or that what they are testing di- 
rectly reflects formulations of CT in the more 
theoretical literature. For instance Passoneau 
(1998) refers to two variantS of CT: "Version 
A" based on Brennan et al (1987) and ~Version 
B" taken from Kameyama et al (1993). Pas- 
soneau does n~t address the issue of direct vs 
indirect realisation and it appears from the ex- 
amples given that she only takes account of en- 
tities realised by a full NP or (possibly null) 
pronoun. The analysis according to Version B 
results in a count of 52% NULL transitions, i.e. 
no Cb, which gives the impression that CT is 
in fact a rather poor measure of coherence, It 
is probable that a higher measure might have 
been obtained if Passoneau had allowed entities 
to be added to the U/'s by inference, as dis- 
cussed in (Brennan et al, op cit.). It is of course 
impossible to verify this without access to the 
original texts, but it is instructive to consider 
the-following example from Strube and Hahn 
(1996): 
(1) a. Ein Reserve-Batteriepack vereorgt den 
316LT ca. 2 Minuten mit Strom. 
(A reserve battery pack - supplies- the 
316LT- ca. 2 minutes - with power.) 
b. Der Status des Ak/ms wird dem An- 
wemier anges  . 
(The status of- the accumulator- is - 
to the user- indicated.) 
S & H treat Ak/m in the (b) sentence as indi- 
rectly ~;|~ing the 316LT (a kind of computer) 
in the (a) sentence, so the latter becomes the 
Cb of (b) resulting in a CONTINtrJS transitio~ 
If the authors had only taken account of direct 
realisations this would be analysed as a NULL 
tr~nRition. 
Furthermore, a preponderance of shifts over 
continues may reflect the domain and content 
rather than the organlp-~tion of a text. In fact 
it can be seen that sequences of smooth shifts 
are rather natural in certain kinds of narrative 
or descriptive texts; see example (2). 
(2) The name of your medicine is Rhinocort 
Aqu  
It contains budmonide. 
This is one of a group of medicines called 
corticosteroid~ SMOOTH SHIFT 
These can help to relieve the symptoms of 
hay fever or rhinitis. SMOOTH SHIFT 
(pharmaceutical leaflet) • 
This does not appear to be an incoherent text, 
0 
0 @ 
@ 
@ 
0 
0 @ 
@ 
0 @ 
@ 
0 
0 
0 
0 
0 
0 @ 
@ 
@ 
@ 
@ 
@ 
@ 
0 
0 
0 @ 
@ 
0 @ 
0 @ 
@ 
@ 
@ 
@ 
@ 
O Q 
O 
O 
O 
O 
O @ 
@ 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O @ 
O 
O 
O 
O 
O 
O 
O 
O 
O @ 
O 
O 
O 
but there is no way that the c°ntent could be 
rearranged to turn the shifts into continues. 
2.2 Deconstructing the transitions 
Strube and Hahn (1996) question the canonical 
ordering of transitions, partly on the grounds 
that this ordering fails to predict the P~TAIN - 
SHIFT pattern which has been claimed by many 
researchers to signal a segment boundary or the 
introduction of a new =discourse topic". (See 
Section 3.2 below.) Recall that the Up is de- 
fined as the most salient entity realised in Un, 
which is predicted to be the Gb of U,+l. How- 
ever this "prediction" is not in fact cashed out 
in the rul~ for centering transitions, which take 
no account of whether Cp(Un) is actually re- 
alised in U,+l. S & H (op cit., p. 275) propose 
the principle of cheapness which is satisfied if 
Cp(U.) = Cb(U.+ l). "Cheapness" is claimed 
to minimise the inferential costs of processing 
sequences of utteranCes, and is proposed as a 
constraint on pairs of successive transitions as a 
replacement for the canonical orderings in Rule 
2. 
We call a tr~n.qition pair cheap if the 
backward-looldng center of the cur- 
rent utterance is correctly predicted by . 
the preferred center of the immediately 
preceding utterance, i.e~, Cb(Ui) = Cp(U,-d... 
In fact it turns out that although cheapness 
appears to be a sensible principle, it does not 
neatly partition the types of transition pairs; 
in particular, this principle does not necessar- 
Uy hold of all s~rAm - SMOOTH SHIFT se- 
quences. S & H propose to rectify this by re- 
defining the transitions, with an additional test 
Cb(Ui) = Cp(0'/-t) to subdivide CONTIN~ and 
mOOTS smFr (Strube, p.c.; Strube and HAhn 
forthcoming). 
In what follows I will also argue against the 
canonical ordering though on different grounds: 
one cannot in general predict a .preference for 
Retain over Shift, for the simple reason that 
there is no point at which the choice between 
these two alternatives arises. Rather, at dif- 
ferent points in the generation process there is 
a choice between maintaining the .~me Cb or 
choosing a new. one, and a choice of which en- 
tit), (Cbor non-Cb) to make salient. So the vari- 
ous transition types emerge in a partial ordering 
from the interaction between salience and cohe- 
sion. Note that if we also include Strube and 
Hahn's "cheapness" principle, there is poten- 
tial competition with salience in cases where 
Cb(Un) ~ Cb(Un+l). That is, we will need a 
way to decide which entity to realise as Cp in 
cases where there are two candidates~ one of 
which is the current G'b and the other the Cb 
of the following sentence. In the remainder of 
this paper I will not directly incorporate the 
"cheapness" principle but will suggest thatsim- 
ilar results are obtained with an appropriate in- 
terpretation of Constraint 3 in the context of 
generation. 
3 Architecture 
If we decompose the rules and constraints into 
separate specifications we see that they poten- 
tially fall under quite different headings in the 
,schematic architecture described above. The 
tr_~,sitions mentioned in Rule 2 are defined in 
terms of two principles which I have called co- 
hesion (maintaining the same ~ from one ut- 
terance to the next) and salience (realising the 
Cb as Cp, normally Subject). If we consider 
these as p|ann|ng operations, cohesion natu- 
rally comes under Text Planning and salience 
under Sentence pl~nnlng, while Rule I concern- 
ing pronominalisation falls under the Referring 
Expression component of Sentence Planning. 
3.1 Text Planning 
A text planner which operated according to C~ 
would seek to order clauses within ~ segment to 
m~t~m the same Cb in a sequence of clauses. 
There are two related issues which compUcate 
this project: firstly, Constraint 3 in Fig. 2 im- 
plies a requirement for/~ to determlne 
the Cb. In addition, there is a potential con- 
flict between top-dow~ hierarchical RST-type 
planning and sequential centering rules. 
3.1.1 Identifying the C'b 
Constraint 3 states that ~for each utterance U~ 
in a discourse segment D... 
The center, Cb(Ui, D) is the highest- 
ranked element of C/(Ui-I,D) that is 
realised in Ui." (Walker et al 1998:3). 
There are two different implementation strate- 
gies which could satisfy this constraint: 
75 ¸ 
1. Take the ranking of Cf(Ui-h D) as given 
and use this to identify the Cb. 
2. Take the Cb as given and plan the reali- 
sation of Ui-~ to make this entity the highest- 
ranked. 
The first strategy is clearly appropriate for in- 
terpretation (cf Brennan et al 1987) but for gen- 
e.ration the issue is less clear-cut. Either the 
generator "interprets" its own output to desig- 
nate Cb in terms of the grammatical structure 
of the previous utterance, in which case there 
have to be separate principles for deciding on 
the grammatical structure, or Cb is indepen- 
dently defined in the text plan and this infor- 
mation is used to plan the sentence structure. 
According to the pipelining principle infor- 
mation cannot flow 'backwards' between tasks. 
In a pipelined system, the ~a!isation of an en- 
tity as Cp may have the effect of setting up an 
expectation on the reader'S part that this en- 
tity will be Cb of the following utterance, but 
it Cannot influence the decision made by text 
planning. This would mean that the 6"b will no 
longer be defined as in Coustraint 3 but must 
be independently designated by the text plan- 
ner as the centre of attention in an utterance. In 
fact the resulting distribution of tasks would be 
rather similar to the Gossip system (Carcagno 
and Iordanskaja 1989) as described by Ca/aiR 
and Reape (1998): 
First, the planner produces "sentence- 
sized semantic nets" which it marks 
with theme-rheme information. 
... Furthermore, the theme/theme 
constraints influence clause ordering, 
...pronominalisation ...and lexical 
choice. 
An implementation of CT with feedback 
seems likely to fit more naturally into an in- 
¢rernenta/architecture, where generation tasks 
may. be carried out on an utterance-by- 
utterance basis in contrast to top-down gener- 
ation of a text tree for an entire discourse. In 
incremental systems it is possible in principle 
to plan the content of an utterance according to 
which entity is currently made salient by its sur- 
face grammatical role, whereas this would not 
in general be possible in a top-down pipelined 
system. In fact it turns out that incremental 
generators tend to perform sentence planning 
incrementally but not text planning. (See e.g. 
Reithinger 1991, DeSmedt and Kempen 1991.) 
Ill fact I would argue that the feedback prob- 
lem is only an artefact of an interpretation of 
Constraint 3 as an implication rather than a 
constraint. The ~implicational ~ interpretation 
is that if an entity a is Cp of Un and is realised 
in Un+l, it should be designated as Cb of Un+l. 
The declarative interpretation is non-directional 
and simply equates Cb of Un+l with the most 
salient entity rea lised in Un which is also re- 
alised in Un+l. The way to implement this while 
keeping to the pipelining principle is to assume 
that the text planner independently designates 
the "theme ", "topic ~ or intended centre of at- 
tention in each clause, which is marked aS Cb 
if it is realised in the previous clause, and to 
have the Sentence Planner promote an entity to 
salience in Un if it is Cb of Un+l. So the text 
planner should annotate each clausal node Un 
in the text tree with the following information: 
Cb of\[/. 
Cb of U._l 
Cb of Un+l 
The sentence planner will then make use of this 
information to decide on pronominali.qation ac- 
cording to the values of the current and pre- 
vious Cb, and on promotion of arguments to 
salience depending on the current and \]ollowing 
Cb. Some concrete proposals are discussed in 
Section 3.2, "Sentence Planning ~. The general 
division of labour is outlined in Fig. 3. 
3.1.2 Top-down Text Planning 
The reader may be disappointed that no inde- 
pendent definition of ~topic" or "theme" is of- 
feted. In fact we consider that this may be out- 
side the scope of CT. When used for interpreta- 
tion, CT offers a set of rules of thumb to guide 
the system in identifying the centre of atten- 
tion in each utterance and finding probable an- 
tecedents for anaphors, but the notion of ~cen- 
terhood" is not defined separately from these 
rules. When used for generation, the most C~ 
can offer is to take the topic or theme as 9/yen 
and exploit the centerln~ rules and constraints 
to construct the text in a way which foregrounds 
these entities and enables the user to correctly 
identify antecedents. So CT has to sit on top of 
an independently specified treatment of infor- 
mation structure. One candidate is Strube and 
Hahn's (1996) reformulation of the notions of 
76 
0 
0 
0 
0 @ 
0 @ 
0 @ 
0 @ 
0 @ 
0 
0 
0 
0 
0 @ 
@ 
0 @ 
0 @ 
0 
0 
0 
0 @ 
0 @ 
0 
O @ 
O 
O 
O 
e 
O @ 
O 
O 
O 
O 
O 
O @ 
O 
O 
O 
O 
e 
O @ 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
O 
Text Planning 
1. Content determination. 
2. Discourse planning: order clauses Ui (within segments) to maximise 
continuity of reference. 
For each clause Un: 
o Designate at most one argument as Ub(U~,), which must be an argument of U,z-I. 
(If intersection of Ufs(U,) and Ufs(U~_l) has only one member, that member is Ub.) 
o Annotate clause node with IDs of Ub(Un), Ub(Un-l) and Cb(Un+l). 
Sentence Planning 
1. $ente'nce aggregation. 
2. Lexica/isation: select verb form for Un so that 
a. Cb(U,+t) is most grammatically salient of intersection of Cfs(U,) and Cfs(U,+t); 
b. subject to (a), Ub(Un) is reafised in most salient available position. 
3. Referring expression generation: working hypotheses . 
Cb(U,) may be pronominalised ifi 
o Cb(U.) = cb(u._,) (QJw 83) 
o C b(U.) = cp(u._~) (Brenn~ 98) 
Figure 3: Locating centering tasks in the pipeline 
theme and theme. Another pomibi\]ity, which is 
currently under investigation, is to experiment 
with the effects of rhetorical structure on choice 
of Cb. 
I've assumed above that the task of main- 
raining continuity of reference can be located 
in the Text Planner. That is, the TP would 
be responsible both for annotating the Cb in 
each utterance, and for organising the text so 
that the same Cb is m~intained over a sequence 
of clauses. However, according to Reiter and 
Dale (1997) a more common method of structur- 
ing text is to make use of "discourse relations, 
such as those described by Mann and Thompson 
(1987), which do not explicitly take continuity 
of reference into account. Richard Power (p.c.) 
has proposed that the implementation of CT in 
an RST-ba~i text planner can be treated as an 
opt~mi-*-~tion problem. That is, the text plan is 
iuitiAlly taken to be a tree structure with dis- 
course relations defined on adjacent nodes but 
at most a partial specification of linear order. 
The problem will then consist of selecting a Ub 
for each propositional leaf node in such a way 
as to maximise the coherence of the text ac- 
cording to centering rules. This is an area of 
active research. Cheng (MS) has proposed a sim- 
ilar strategy for maintaining local coherence in 
a text planner using a genetic algorithm. 
Another issue is whether the CT rules, which 
ass-me a "flat" sequence of utterances, will re- 
main valid for a hierarchical/); structured text 
plan. In fact it is an open research ques- 
tion whether CT should operate in this man- 
her or whether the rules should be reformu- 
fated to take account of dominance in addition 
to precedence; di~erent positions are taken by 
Kameyama (1988) and Suri and McCoy (1994). 
3.2 Sentence planning 
According to the re-interpretation of CT which 
was sketched above, Sentence Planning may 
promote an entity for salience if it is the Cb 
of the current or .following utterance. There 
is dearly potential competition between these 
two factors which will be discussed shortly. The 
preference for rp~di~mg Cb as ~ can be hnple- 
mented by choosing a verb form which projects 
G'~ in subject position. Some poes~ilities are 
franker alternation: buy/sd~ gi~/receive, bor- 
row/qend etc, or pamivisation: your doctor ma~/ 
prescribe this medicine for gout vs this mtdicine 
,~y be pr~cnbed .for gout. 
If we compare the attested example (2) with 
the constructed (3) it is clear that the former 
reads more naturally: . 
O) Hypoglycaemia (Cb) may cause faintness, 
sweating, shaking, weakness and confusion. 
It m~y be due to lack of food or too high a 
dose of the medicine.CONTINU8 
It can be put right by eating or drinking 
something sugary.CONTINU~- 
77 
(pharmaceutical leaflet) 
(4) -..Hypoglycaemia may cause faintness, 
sweating, shaking, weakness and confusion. 
A lack of food or too high a dose of the 
medieinemay cause it.RETAIN 
Eating or drinking something sugary can 
put it right.RETAIN 
(modified example) 
Hurewitz (i998) examined the use of passives to 
promote cohesion and salience and found that 
in both written texts and speech approximately 
75% of passives had either the CONTINUE or 
SMOOTH-SHIFT transition. In each case the ef- 
fect is to promote Gb to Subject in accordance 
with the salience principle. (For written texts 
this proportion was not signiRcantly different 
from a control sample, whereas with the spo- 
ken passages the proportion was slightly higher 
than in the'control.) 
One system which explicitly makes Use of CT 
is the Caption Generation System (CGS) re- 
ported in (Mittal et al 1998). This system has 
a separate "centering module" which orders ar- 
guments within a clause to improve coherence 
of a text but does not influence the order of 
clauses. Thus only the salience principle is im- 
plemented and the centering task is located as 
part of Sentence Planning:. the speci-l!.~ed Cen- 
tering Module receives Control after clauses have 
• been ordered and aggregated (op cit:454). The 
strategy adopted is to keep ~the highest-ranking 
forward-looking center of the first clause of the 
segment ... as the Cp(Ui) of a/l the following 
clauses in the same segment" (op cit:456; my 
emphasis). Clearly this strategy isunlikely to 
generalise to a variety of domtt;na. 
As mentioned above there is a potential con- 
met betwe~ Co~t 3 (make Cb(U.+t) 
salient) and salience (make Cb(U,) aslient). 
As noted in Sect. 2.2 there may also be compe, 
tition between salience and Strube & Halm's 
cheapness principle, which can be seen as a 
stronger version of C3. There are different ways 
this conflict could be tadded in the cases where 
it arises aad I will consider one of them, which 
is to let C3 win out over salience. 
Consider a text with four clauses Ul - U4. 
which all have a and b among their arguments. 
Let a be the Cb of Ul- U2 and b the Cb of Us- U#. 
According to salience and C3, b will be Cp of 
/\]3 since it is Cb of that clause and the following 
one. For U2 there is competition between a and 
b to be Up, and this is decided by C3 in favour 
of b. Finally I a is chosen as Up of Ul. The 
result is as follows: 
UI : Cb f a, Cp = a 
U2:Cb=a, Cp=b 
U3 : Cb f b, Cp = b 
U4 : Cb = b, Cp = b 
In terms of the conventional transitions this 
works out as 
U~/U2: RET^IN U2/U3: 
SMOOTH SHIFT 
us/u~: COnTINUa 
This is consistent with Strube and Hahn's 
(1996) observation that "a II~rAIN transition 
ideally predicts a SMOOTH ssw'r in the follow- 
ing utterance". Brenuan et al (1987) make a 
very similar claim: 
A computational system for 9e-em- 
tion would try to plan a retention as 
a signal of an impending shift, so that 
after a retention, a shift would be pre- 
ferred rather than a continuation. 
Grosz et al (1995) give the following example of 
the ~Am - SHIFT pattern: 
(5) a. John has had trouble arranging his va- 
cation. 
b. He (Cb; John) cannot find anyone to 
take over his responsibilities. 
c. He (C/~, John) called up Mike yester- 
day to work out a plan. CONTINUB 
d. Mike hasannoyed him (Cb; John) alot 
recently, lt~rAIN 
e. He (Cb; Mike) called John at 5 am on 
Friday last week. smrr 
Under the approach outlined here, which as- 
sumes that the Cb is independently designated, 
• the system does not needto plan particular 
transition types or even to know about them; 
the desired effects come about as a result of Io- 
cal decisions by the sentence planner using in- 
formation from the text pl-nner. 
• Example (5) incidentally illustrates a limita- 
tion of CT in its Canonical version: the theory 
l'~lle w~rd a~ does not imply that these deci- 
sions are .taken in sequence. 
78 
@ 
0 
0 @ 
0 
0 
0 
0 
0 @ 
@ 
0 
0 
0 
0 @ 
@ 
@ 
0 
0 @ 
0 @ 
@ 
0 
0 @ 
@ 
@ 
0 
0 
0 
0 
0 @ 
@ 
0 
0 
0 
0 
0 
0 @ 
@ 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
0 
correctly predicts the pronominal choices in (5C 
- e) but has nothing to say about the decision 
to make Mike rather than John the Subject of 
(Sd). In fact, we can construct a variant of this 
text which follows centering rules more faith- 
fully, though it does not read any more natu- 
rally: 
(6) 
e. 
& John has had trouble arranging his va- 
cation. 
b. He cannot find anyone to take over his 
responsibilities. (Cb -- John) 
c. He called up Mike yesterday to work 
out a plan. 
(CONTINUE; Cb = John) 
d. He has been pretty annoyed with Mike 
recently. 
(colqzlm~; Cb John) 
He got a call from him (Mike) at 5 am 
on Friday last week. (CONTrol; Cb 
= John) 
If we examine the discourse structure of exam- 
ple (5), it seems that the discourse as a whole is 
about John but (Sd,e) form a parenthetical sec- 
tion which tells us something about M/ke. So 
although a blind application of centering rules 
would judge:(6) to be more coherent, (5) is in 
fact maximally coherent within the constraints 
of the structure of the discourse. 
To snmmarise: we assvme that for each utter- 
ance U, the text planner has identified Cb(U,), 
Cb(U.-1) and Cb(U.+d. The task for the sen- 
tence planner is to select a verb form or some 
other syntactic device to malise Cb(Un+l) as the 
most salient of those entities which are reAl|.~ed 
in U. and U.+t, and subject to this, to realise 
Cb(gJ'.) in the most salient position available. 
(See Figure 3,) So for example if Cb(Un) is not 
to be realised in U,÷t, it will normally be re- 
alised as Cp(U,); but if it/8 to be realised in 
Un+I then the Cb of U,+t will be at least as 
highly ranked in Un as the Cb of Un. Tlfisstrat- 
egy predicts the Ralisation of M//~ rather than 
John as Cp in (5d) above. 
3.3 Referring Expressions 
The contribution of CT to Referring Expression 
(RE) generation is to decide on pronominalisa- 
tion. Rule 1 (Fig. 2) which concerns pronom- 
inalisation has a strong and a weak formula- 
tion: tile strong one is that a pronoun should 
be used for the Cb if it is the same as the Cb 
of the previous utterance, the weak one is that 
the Cb must be pronominalised if anything is. 
In the context of generation it is probably safer 
to use the strong version. Brennan (1998) pro- 
poses, arguing from corpus analysis, that the 
Cb should be pronominalised only if it is Cp of 
the previous utterance. Robert Dale's RPICURE 
system employs the terminology of CT in Con° 
nection with RE generation, as does the ILEX 
system reported in (O'Donnell et al 1998). Both 
these systems implement a variant of Rule 1 to 
determine whether to pronominalise the center 
(Dale) or Cb (ILEX), though in neither case 
is the center identified according to the stan- 
dard apparatus of CT. In ILEX the Cb is des- 
ignated by the text planner without reference 
to the content of the previous sentence, and it 
may be pronominal|.qed if it is the same as the 
previous sentence's "Cb". Dale's IZPIOURE iden- 
tifies the center with Uthat entity which is the 
result of the previously described operation" 
(Dale 1992:170). Passoneau (1998) constructed 
input for a prototype generator by "hypothesis- 
ing" a Cb for each proposition in a text based 
on the salience of entities in a Situation. Pas- 
soneau's system uses centering constraints to 
decide whether to realise entities as definite pro- 
nouns, minimal NPs or full NPs. 
4 Conclusions and future work 
CT has developed primarily as a tool for 
analysing the structure of a given text and iden- 
tifying the most likely candidates for anaphora 
resolution. In this paper we sought to deter° 
mine whether the principles underlying the con- 
straints and rules of the theory can be Uturned 
round" and used as planning operators for gen- 
erating coherent text,. As a side-effect of this 
enterprise we have articulated a ~streamlined" 
formulation of CT in terms of the principles 
of salience and cohesion, and argued that 
the preferences for the different transition types 
emerge in a partial ordering f~om the interac- 
tion between these principles. These princi- 
ples are rather heterogenous, a fact which is 
obscured by combining them in the transition 
definitions, and can be implemented as encap- 
sulated tasks distributed between text planning, 
79 
sentence planning and RE generation. It may 
turn out that individual components such as the 
cohesion principle do not need to be explicitly 
stated but emerge as by-products from higher- 
level text planning. 
As noted at various points in this paper CT 
has never been more than partially implemented 
in NLG systems. This may be due to a belief 
on the part of NLG practitioners that CT gets 
things the wrong way round, by relying on sur- 
face grammatical realisation to determine the 
• centre of attention in an utterance. If this be- 
lld is commonly held (and anecdotal evidence 
suggests that it is), I argue that it is mistaken. 
In interpretation systems the principles of CT 
guide the system in identifying the centre of 
attention and in choosing likely referents for 
anaphors. In NLG systems, if there is a notion 
of "topic" or "theme = this should be designated 
by the text planner, while the CT rules allow 
the sentence planner to promote this entity to 
salience to keep it as the user's "centre of atten- 
tion". 
However, a more fuDd~mental explanation for 
the neglect of CT in the generation literature 
is provided by the fact that a faithful imple- 
mentation in a pipelined system turns out to 
require an independent way of designating the 
central entity in a proposition, and this itself is 
a problem which has not had much attention 
in the development of NLG systems 2. So the 
next stage in this research will Concentrate on 
developing a characterisation of Cb based on se- 
mantic content and information structure, tak- 
ing account of e.g. the proposals of Strube and 
Hahn (1996) and experimenting with optlmi- 
sation algorithms as discussed in section 3.1.2 
above. 
As mentioned above the Reiter model has 
been questioned by members of the RAGS 
project in Brighton and Edinburgh who are ac- 
tively engaged in developing a "Teference" archi 
tecture for NLG. (See Cahlll et al., 1999, RAGS 
1999.) To date' the group has concentrated 
on specifying the data structures which are re- 
quired at various stages of the generation task 
and has identified a number of discrete functions 
such as rhetorical structuring, aggregation, 
~Thk appHm at lea~ to applied systems;, see CahiU 
• and Reape 1998. One exception appears to be the GOS-. 
SIP system described in Caragno and Iordaaskaja 1989. 
coherence etc., without specifying a strict or- 
der for the execution of these functions. It is too 
early to assess how the proposals of this paper 
would fit into the RAGS scheme, but I antici- 
pate that, as with the Reiter architecture, the 
conclusion would be that referential coherence 
is not the task of a discrete module but imposes 
constraints on a number of different modules. 
It has been noted that the way the "realise" 
relation is interpreted can have significant im- 
plications for the coherence of a text as mea- 
sured by CT, and that corpus analysis has of- 
ten concentrated ondirectly realised Cb's. An 
exception is Hahn, Strube and Markert's (1996) 
treatment of bridging reference, or "textual el- 
lipsis" in their terminology. As these authors 
note there have been rather few implemented 
systems which are able to interpret bridging ref- 
erences in a principled way, and research in NLG 
is particularly weak in this area. SO, a faith- 
ful implementation of CT in generation systems 
will depend in part on progress in the genera- 
tion of bridging references. This is an area for 
future research. 
.... It is intended that the procedures described in 
"this" paper will be implemented in ICONOCLAST, 
an authoring tool which enables domain experts 
to create a knowledge base through a sequence 
of interactive choice-- and generates hiexarchi- 
tally structured text according to various stylis- 
tic constraints (See Power and Scott 1998). 
Acknowledgements 
This work was carried out as part of the 
GNOME project (Generating Nominal Expres- 
sions) which is s collaboration between ITRI in 
the University of Brighton and the H~tc in the 
~nlversities of Edinburgh and Durham, funded 
by the EPSRc under grant reference GR/L51126. 
I would like to thank ITRX and GNOME colleagues 
for helpful feedback, particularly Christy 1)o- 
ran, Renate Henschel, Richard Power and Kees 
van Deemter, as well as two anonymous referees. 

References 
S Brennan, M W, lker Priedman and C Pol- 
lard 1987. A Centering Approach to Pronouns. In 
Proc. 25th AC\[, :115-62. 
L C-hill, C Doran, R Ev-n-, C Mellish, D 
Palva, M Reape and D Scott 1999. In search of 
a reference architecture for NLG systems. In Proc. 
EWNLG'99. 
L Cahill and M Reape 1998~ Component 
tasks in Applied I:4atural Language Generation Sys- 
tems. RAGS Project Deliverable, available at 
www. i tri. brighton, ac. uk/proj ects/rags. 
D Caragno and L Iordanskaja 1989. Content 
Determination and Text Structuring in GOSSIP. In 
Eztended Abstracts o.f ENLG'89. 
H Cheng MS. Experimenting with the Interaction 
between Aggregation and Text Planning. Unpub- 
lished paper, Division of InformatiC~, University of 
Edinburgh. 
R Dale 1992, Generating Re/erring Expreasions, 
Cambridge, MA/London:MIT Press. 
K De Smedt and G Kempen 1991. Segment 
Grammar: a Formalism for Incremental Sentence 
Generation. In C Paris, W Swartout and W Mann 
(eds), Natnral Language Ge.era6on in Artificial ln- 
teUigence and Computational Linguistics. 
B Di Eugenio 1996. The discourse functions of 
Italian subjects: a centering approach. In Proe. 
GOLING96. 
B Di Eugeni0 1998. Centering in Italian. In 
Walker et al (eds) Centering Th~j in Discourse. 
P Gordon, B Grosz and L Gilllom 1993. Pro- 
nouns, Names and the Centering of Attention in Dis- 
course. Co#nitiee Scien~ 17/3:311-47. 
B Grosz, A Joshi and S Weinsteln 199.5. Cen- 
teriag: a framework for modelling the local coher- 
ence of discourse. Comput~iond Linguistics, 21/2: 
203-25. 
U Hahn, M Strube and M Markert 1996. 
Bridging Textual Ellipses. In Prec. e/COLING- 
96. 
F Hurewitz 1998. A Quantitative Look at Dis- 
course Coherence. In Wall~ et al (ecb) Centering 
Theory in Du~.rs¢ 
M Kameyama 1998. In~tential Centering: A 
Case Study. In Walker et al (eds) Centering Theory 
in Diso~rse. 
M Kameyamm, R Pammneau and M Poeslo 
1993. Temporal C4mtefinfr In Prec. o~ $1stACL. 
W M--, and S Thompson 1987. Rhetorical 
Structure Theory: A Theory of Text Organisation. 
In L Polanyi (ed.), The 3ble~ure oj r D/acoerse 
V Mittal, J Moore, G Carenini and S Roth 
1998. Describing Complex Charts in Natural Lan- 
guage: A Caption Generation System. Computa- 
tional Linguistics, 24/3:431-468. 
• R Power and D Scott 1998. Mult'flingual author- 
ing using ffedback texts. In Prec. COLING/ACL 
'98 
RAGS 1999. The RAGS Project. Towards 
a Reference Architecture for Natural Language 
Generation Systems. Technical report ITRI-99- 
14, ITRI, University of Brighton. Available at 
wwu. itri. brighton, ac. uk/proj ects/rags. 
E Reiter 1994. Has a consensus NL generation 
architecture appeared, and is it psycholinguistically 
plausible? In Proc. INLG 7.'163-70. 
E. Reiter and R. Dale (1997). Building Applied 
Natural-Language Generation Systems. Journal o.f 
Natural-Language Engineerin9 3:57-8Z 
N Reithinger 1991. POPEL - A Parallel and In- 
cremental Natural Language Generation System. In 
C Paris, W Swartout and W Mann (eds), Natural 
Languafe Generation in Artificial Intelligence and 
Computational Linguistic. 
R Stevenson, R Crawley and D Kleinman 
1994. Thematic roles, focus and the representation 
of events. Laf~3nage and Cognitive Proc_e~_es, 9:519- 
48. 
M Strube and U H,h- 1996. Fancfional Center- 
ing. In Proceedings ACL 34:270-77. 
M Strube and U Hahn 1999. Functional Cen- 
tering:. Grounding Referential Coherence in Infor- 
marion Structure. To appear in Computational Lin- 
guistics. 
L Suri and K McCoy 1994. RAFT/RAPR 
and Centering: A Comparison and Discussion of 
Probl~ Related to ProcemingComplex Sentences. 
GompeMtional linguistics 20/2:30!-17. 
M Walker 1998. Centering, Anaphora Resolution 
and Discourse Structure. In Walker et al (eds) Cen- 
tering Theory in Disanwse. 
M Walker, AK Joshl and E Prince (eds) 1998. 
Centering Theory in Disburse. Oxford: Clarendon 
Preu. 
