Capturing the Interaction between Aggregation and Text 
Planning in Two Generation Systems 
Hua Cheng and Chris Mellish 
Division of Informatics, University of Edinburgh 
80 South Bridge, Edinburgh EH1 1HN, UK 
huac, chrism@dai, ed. ac. uk 
Abstract 
In natural  generation, different gener- 
ation tasks often interact with each other in a 
complex way. We think that how to resolve the 
complex interactions inside and between tasks 
is more important to the generation of a co- 
herent text than how to model each individual 
factor. This paper focuses on the interaction be- 
tween aggregation and text planning, and tries 
to explore what preferences exist among the fea- 
tures considered by the two tasks. The prefer- 
ences are implemented in two generation sys- 
tems, namely ILEX-TS and a text planner us- 
ing a Genetic Algorithm. The evaluation em- 
phasises the second implementation and shows 
that capturing these preferences properly can 
lead to coherent text. 
1 Discourse coherence and 
aggregation 
hi NLG, theories based on domain-independent 
rhetorical relations, in particular, Rhetorical 
Structure Theory (Mann and Thompson, 1987), 
are often used in text planning, whose task 
is to select the relevant information to be ex- 
pressed and organise it into a hierarchical struc- 
ture which captures certain discourse prefer- 
ences such as preferences for the use of rhetori- 
cal relations. 
In the theory of discourse structure developed 
by Grosz and Sidner (1986), each discourse seg- 
ment exhibits two types of coherence: local co- 
herence among utterances inside the segment, 
and global coherence between this segment and 
other discourse segments. Discourse segments 
are connected by either a dominaTzce relation or 
a satisfaction-precedence relation. 
There has been an effort to synthesise tile 
two accounts of discourse structure. X loser and 
Moore (1996) argue that the two theories have 
considerable common ground, which lies in the 
correspondence between the notion of domi- 
nance and nuclearity. It is possible to map 
between Grosz and Sidner's linguistic structure 
and RST text structure, and relation-based co- 
herence and global coherence capture similar 
discourse properties. 
Oberlander et al. (1999) propose a dis- 
tinction between two types of discourse coher- 
ence: proposition-based coherence, which ex- 
ists between text spans connected by RST re- 
lations except for object-attribute elaboration, 
and entity-based coherence, which exists be- 
tween spans of text in virtue of shared entities. 
entity-based coherence captures the coherence 
among adjacent propositions, which resembles 
local coherence in Grosz and Sidner's theory. 
To generate a coherent text, the text planning 
process must try to achieve both local (entity- 
based) and global (relation-based) coherence. 
Since the task of aggregation is to combine sinl- 
ple representations together to form a complex 
one, which in the mean time leads to a shorter 
text as a whole, aggregation could affect the or- 
dering of text plans and the length of the whole 
text.. Therefore, it is closely related to tile task 
of maintaining both types of coherence. Here 
we treat embedding as a type of aggregation. 
There is no consensus as to where aggregation 
should happen or how it is related to other gen- 
eration processes (Wilkinson, 1995; Reape and 
Mellish, 1999). In many NLG systems, aggre- 
gation is a post planning process whose prefer- 
ences are only partially taken into account by 
the text planner. 
1.1 Aggregation and local coherence 
In a structured text plan produced by the text 
planner, local coherence is normally maintained 
through the ordering of the selected facts, where 
186 
certain types of center transition (e.g. cen- 
ter continuation) :are preferred :over:others (eig,. -; 
center shifting) (Centering Theory (Grosz et al., 
1995)). Aggregation may affect text planning 
by taking away facts from a sequence featuring 
preferred center movements for embedding or 
subordination. As a result, the preferred cen- 
ter transitions in the original sequences could 
be cut off. For example, comparing the first 
two descriptions of.a necklace in Figure 1, 2 is 
less coherent than 1 because of the shifting from 
the description of the necklace to that of the de- 
signer, which is a side effect of embedding. 
Since the centers of sentences are normally 
NPs and embedding adds non-restrictive com- 
ponents into an NP, it could affect the way a Cb 
is realised (e.g. preventing it from being a pro- 
noun). As pointed out in (Grosz et al., 1995), 
different realisations (e.g. pronoun vs. definite 
description) are not equivalent with respect to 
their effect on coherence. Therefore, embedding 
could influence local coherence by forcing a dif- 
ferent realisation from that preferred by Center- 
ing Theory. There is an obvious need to balance 
the consideration for local coherence and stylis- 
tic preferences. 
1.2 Aggregation and global coherence 
Different types of aggregation need to be com- 
patible among themselves, in particular, embed- 
ding and semantic parataxis and hypotaxis. Us- 
ing the abstraction of RST, semantic parataxis 
concerns facts related by explicit multi-nuclear 
semantic relations (e.g. sequence and contrast) 
or by implicit connections like parallel common 
parts. If two facts have at least two identi- 
cal parallel components, we say that a conjunct 
or disjunct relation exists between them, and 
these relations are multi-nuclear relations. Se- 
mantic hypotaxis concerns facts connected by 
nucleus-satellite relations (e.g. cause). Seman- 
tic parataxis and hypotaxis feature in relation- 
based coherence and they depend on the text 
planner to put the related facts next to each 
other in order to perform a combination. 
(Cheng, 1998) describes interactions that 
need to be taken into account in aggrega- 
tion. Firstly, complex embedded components 
like non-restrictive clauses may interrupt tile 
semantic connection or syntactic similarity be- 
tween a set of clauses. Secondly, the possibilities 
of other types of aggregation should be consid- 
ered for both the main fact and the fact to be 
-embedded .during .:embedding .decision. maki ng... 
And thirdly, performing parataxis inside a hy- 
potaxis could convey wrong information. 
We argue that the effect of aggregation is not 
limited to the particular NP or sentence where 
aggregation happens, but to the coherence of 
the text as a whole. The complex interactions 
demand the features of aggregation to be eval- 
uated .together with other coherence~ features 
and aggregation to be planned as a part of text 
structuring. This requires better coordination 
between aggregation and other generation tasks 
as well as among different types of aggregation 
than is present in current NLG systems. 
In this paper, we describe how to capture the 
above interactions as preferences among related 
features, and the implementation of the prefer- 
ences in two very different generation architec- 
tures to produce descriptions of museum objects 
on display. 
2 Preferences among coherence 
features 
We claim that it is the relative preferences 
among features rather than the absolute magni- 
tude of each individual one that play the crucial 
role in the production of a coherent text. In this 
section we discuss the preferences among fea- 
tures related to text planning, based on which 
those for embedding can be introduced. 
2.1 Preferences for global coherence 
A semantic relation other than conjunct or dis- 
junet is preferred to be used whenever possible 
because it usually conveys interesting informa- 
tion about domain objects and leads to a coher- 
ent text span. If a conjunct relation shares a fact 
with a semantic relation, the conjunct should 
be suppressed. For example, in 3 of Figure 1. 
apart from other relations, there is an amplifica- 
tion relation signalled by indeed and a conjunct 
between the last two propositions. Compared 
with 3, 4 is less preferred because it misses tile 
amplification and the center transition from the 
necklace to an Arts and Crafts style jewel is not 
so smooth, whereas 3 expresses the amplifica- 
tion explicitly and the conjunct implicitly. 
However, a semantic relation can only be used 
if the knowledge assumed to be shared by the 
hearer is introduced in the previous discourse 
(Mellish et al.. 1998a). \Ve assume the strategy 
187 
1. This necklace is in the Arts and Crafts style. Arts and Crafts style jewels usually have an elaborate 
design. They tend to have floral motifs. For instance, this necklace has floral motifs. It was designed 
by Jessie King. King was Scottish. She once lived in London. 
2. This necklace, which was designed by Jessie King, is in the Arts and Crafts style. Arts and 
Crafts style jewels usually have an elaborate design. They tend to have floral motifs. For instance, 
this necklace has floral motifs. King was Scottish. She once lived in London. 
3. The necklace is in the Arts and Crafts style. It is set with jewels in that it features cabuchon 
stones. Indeed, an Arts and Crafts style jewel usually uses cabuchon stones. It usually uses oval 
stones. 
4. The necklace is in the Arts and Crafts style. It is set. with jewels in that it features cabuchon 
stones. An Arts and Crafts style jewel usually uses cabuchon stones and oval stones. 
Figure 1: Aggregation examples 
of (Mellish et al., 1998a) which uses a joint re- 
lation to connect every two text spans that do 
not have a semantic relation other than object- 
attribute elaboration and conjunct/disjunct in 
between. Although joint is not preferred when 
other relations are present, it is better than 
missing presuppositions or embedding a con- 
junct relation inside a semantic relation. There- 
fore, we have the following heuristics, where 
"A>B" means that A is preferred over B. 
Heuristic 1 Preferences among features for 
global coherence: 
a semantic relation > Conjunct/Disjunct > 
Joint > presuppositions not met 
Joint > Conjunct inside a semantic relation 
2.2 Preferences for local coherence 
One way to achieve local coherence is to con- 
trol center transitions among utterances. In 
Centering Theory, Rule 2 specifies preferences 
among center movement in a locally coherent 
discourse segment: sequences of continuation 
are preferred over sequences of retaining; which 
are then preferred over sequences of shifting. 
Brennan et el. (1987) also describe typical 
discourse topic movements in terms of center 
transitions between pairs of utterances. They 
argue that the order of coherence among the 
transitions is continuing > retaining > smooth 
shifting > abrupt shifting. Instead of claiming 
that these are the best models, we use them 
simply as an example of linguistic models being 
used for evaluating features of text planning. 
A type of center transition that appears fre- 
quently in descriptive text is that the descrit)- 
tion starts with an object, but shifts to associ- 
ated objects or perspectives of that object. This 
is a type of abrupt shifting, but it is appropriate 
as long as the objects are highly associated to 
the original object (Schank, 1977). This phe- 
nomenon is handled in the system of (Grosz, 
1977), where subparts of an object are included 
into a focus space as the implicit foci when the 
object itself is to be included. 
We call this center movement an associate 
shifting, where the center moves from a trig- 
ger entity to a closely associated entity. Our 
informal observation from museum descriptions 
shows that associate shifting is preferred by hu- 
man writers to all other types of center move- 
ments except for continuation. There are two 
types of associate shifting: where the trigger 
is in the previous utterance or two entities in 
two adjacent utterances have the same trigger. 
There is no preference between them. 
Heuristic 2 summarises the above preferences. 
We admit that these are strict heuristics and 
that human texts are sometimes more flexible. 
Heuristic 2 Preferences among center transi- 
tions: 
Continuation > Associate shifting > RetaiTI- 
ing > Smooth shifting > Abrupt shifting 
2.3 Preferences for both types of 
coherence 
Two propositions can be connected in differ- 
ent ways, e.g. through a semmxtic relation or a 
smooth center transition only. Since a semantic 
relation is always preferred, we have the follow- 
ing heuristic: 
Heuristic 3 Preferences among semantic rela- 
tions and center transitions: 
a semantic relation > Joint ÷ Continuation 
188 
2.4 Preferences for embedding Good embedding > Normal embedding > 
We distinguish between.a.-good,.rwrmal,and-bad Joint > Bad embedding ..... =:--..~ .:-- ~ .--:.: ........ 
embedding based on the features it bears. We do Continuation + Smooth shifting + Joint > 
not claim that the set of features is complete. 
In a different context, more criteria might have 
to be considered. 
A good embedding is one satisfying all the fol- 
lowing conditions: 
1. The referring part is an indefinite, a demon- 
strative or a bridging description (as de- 
fined in (Poesio et al., 1997)). 
2. The embedded part can be realised as an 
adjective or a prepositional phrase (Scott 
and de Souza, 1990). 
3. In the resulting text, the embedded part 
does not lie between text spans connected 
by semantic parataxis and hypotaxis. 
4. There is an available syntactic slot to hold 
the embedded part. 
A good embedding is highly preferred and 
should be performed whenever possible. A nor- 
mal embedding is one satisfying condition 1, 3 
and 4 and the embedded part is a relative clause 
which provides additional information about 
the referent. Bad embeddings are all those left, 
for example, if there is no available syntactic 
slot for the embedded part. 
Since semantic parataxis has a higher priority 
than embedding (Cheng, 1998), a good embed- 
ding should be less preferred than using a con- 
junct relation, but it should be preferred over a 
center continuation for it to happen. 
To decide the interaction between an embed- 
ding and a center transition, we use the first two 
examples in Figure 1 again. The only difference 
between I and 2 is the position of the sentence 
"This necklace was de.signed by Jessie King", 
which can be represented in terms of features of 
local coherence and embedding as follows: 
the last three sentences in 1: Joint + Contin- 
uation + Joint + Smooth shifting 
the last two sentences plus embedding in 2: 
Joint + Abrupt shifting + Normal embedding 
1 is preferred over 2 because the center inoves 
more smoothly in 1. The heuristics derived from 
the above discussions are summarised below: 
Heuristic 4 Preferences among features for 
embedding and center transition: 
Abrupt shifting + Normal embedding 
Good embedding > Continuation + Joint 
Conjunct > Good embedding 
The '+' symbol can be interpreted in different 
ways, depending on how the features are used 
in an NLG system. In a traditional system, it 
means the coexistence of two features. In a sys- 
tem using numbers for planning, it can have the 
same meaning as the arithmetic symbol. 
3 Capturing the preferences in ILEX 
The architecture of text planning has a great 
effect on aggregation possibilities. In object de- 
scriptive text generation, there lacks a central 
overriding communicative goal which could be 
decomposed in a structured way into subgoals. 
The main goal is to provide interesting infor- 
mation about the target object. There are gen- 
erally only a small number of relations, mainly 
object-attribute elaboration and joint. For such a 
genre, a domain-dependent bottom-up planner 
(Marcu, 1997) or opportunistic planner (Mel- 
lish et al., 1998b) suits better than a domain- 
independent top-down planner. In these archi- 
tectures, aggregation is important to text plan- 
ning because it changes the order in which infor- 
mation is expressed. The first implementation 
we will describe is based on ILEX (Oberlander 
et al., 1998). 
ILEX is an adaptive hypertext generation 
system, providing natural  descriptions 
for museum objects. The bottom-up text plan- 
ning is fulfilled in two steps: a content selection 
procedure, where a set of fact nodes with high 
relevance is selected from the Content Potential 
(following a search algorithm), and a content 
structuring procedure, where selected facts are 
reorganised to form entity-chains (based on the 
theory of entity-based coherence), which repre- 
sent a coherent text arrangement. 
To make it possible for the ILEX planner to 
take into account aggregation, we use a revised 
version of Meteer's Text Structure (Meteer, 
1992; Panaget, 1997) as the intermediate level of 
representation between text planning and sen- 
tence rcalisation to provkte abstract syntactic 
constraints to the planning. We call this sys- 
tem ILEX-TS (ILEX based on Text Structure). 
189 
In ILEX-TS, abstract referring expression de- 
termination and.aggxegation are performed dur-..: 
ing text structuring. For each fact whose Text 
Structure is being built, if an NP in the fact can 
take modifiers, the embedding process will find 
a list of elaboration facts to the referent and 
make embedding decisions based on the con- 
straints imposed by the NP form. The decisions 
include what to embed and what syntactic form 
the embedded part should use. 
Heuristic 1, 2 and 3 are followed naturally ~ 
by the ILEX text planner, which calculates the 
best RS tree and puts facts connected by the 
imaginary conjunct relation next to each other. 
It tries to feature center continuations as often 
as possible. When it needs to shift topic, it uses 
a smooth shifting. 
ILEX-TS has a set of embedding rules, where 
those rules featuring good embedding are al- 
ways used first, then a rule featuring a normal 
embedding. Bad embedding is not allowed at 
all. To coordinate different types of aggrega- 
tion, the algorithm checks parataxis and hy- 
potaxis possibilities for each nucleus fact and 
the fact to be embedded before it applies an 
embedding rule. These realise most of Heuris- 
tic 4 (except for the second set). However, be- 
cause the various factors are optimised in order 
(with no backtracking), there is no guarantee 
that the best overall text will be found. In addi- 
tion, complex interactions between aggregation 
and center transition cannot be easily captured. 
4 Text planning using a GA 
Although most heuristics can be followed in 
ILEX-TS, some interactions are missing, for ex- 
ample, 9 of Figure 1 will probably be generated. 
For better coordination, we adopt the text plan- 
ner based on a Genetic Algorithm (GA) as de- 
scribed in (Mellish et al., 1998a). The task is. 
given a set of facts and a set of relations between 
facts, to produce a legal rhetoricalstrncture tree 
using all the facts and some relations. 
A fact is represented in terms of a subject, 
a verb and a complement (as well as a unique 
identifier). A relation is represented in terms of 
the relation name, the two facts that are con- 
nected t) 3" the relation and a list of precondition 
facts which need to have been mentioned before 
the relation can be used i. 
1As this is an experimental system, the ability of the 
A genetic algorithm is suitable for such a 
problem.because,:the..numher-.of.-possihle-com- 
binations is huge and the search space is not 
perfectly smooth and unimodal (there can be 
many good combinations). Also the generation 
task does not require a global optimum to be 
found. What we need is a combination that is 
coherent enough for people to understand. 
(Mellish et al., 1998a) summarises the genetic 
algorithm roughly as follows: 
1. Enumerate a set of random initial se- 
quences by loosely following sequences of 
facts where consecutive facts mention the 
same entity. 
2. Evaluate sequences by evaluating the 
rhetorical structure trees they give rise to. 
3. Perform mutation and crossover on the se- 
quences. 
4. Stop after a given number of iterations, and 
return the tree for the "best" sequence. 
The advantage of this approach is that it pro- 
vides a mechanism to integrate planning factors 
in the evaluation function and search for the 
best combinations of them. So it is an excellent 
framework for experimenting with the interac- 
tion between aggregation and text planning. 
In the algorithm, the RS trees are right- 
branching and are almost deterministically built 
from sequences of facts. Given two sequences, 
crossover inserts a random segment from one 
sequence in a random position in the other to 
produce two new sequences. Mutation selects 
a random segment of a sequence and moves it 
into a random position in the same sequence. 
To explore the whole space of aggregation. 
we decide not to perform aggregation on struc- 
tured facts or on adjacent facts in a linear se- 
quence because they might restrict the possibil- 
ities and even miss out good candidates. In- 
stead, we define a third operator called embed- 
ding mutation. Suppose we have a sequence 
\[U1,U2,...,Ui,...,U.\], where we call each element 
of the sequence a unit, which can be either a fact 
or a list of facts or units with no depth limit. 
For a list, we call its very first fact the main fact, 
system is limited in all aspects. It does not have a real 
realisation component, so the parts we are less interested 
in are realised by canned phrases for readability. 
190 
Features/Factors 
Semantic relations 
a joint 
a conjunct or disjunct 
a relation other than joint, conjunct or disjunct 
a conjunct ,inside other semantic relations 
a precondition not satisfied 
Focus moves 
a continuing 
an associate shifting 
a smooth shifting 
resuming a previous focus 
Embedding 
a good embedding 
a normal embedding 
a bad embedding 
Others 
topic not mentioned in the first sentence 
Values (raters) 
1 \] 2 .. 
-20 -46 
10 11 
21 69 
-50 -63 
-30 -61 
20 7 
16 1 
14 -3 
6 -43 
6 3 
3 0 
-30 -64 
-10 -12 
Table 1: Two different raters satisfying the same constraints 
/ x I % / \ 
ii ~--~t ! x x i I x 
I x 
I I ~ x i I 
/I x l 
_m 
, -h-- ..... X ...... g .... -i~ ....... T ....... i; ....... ~ .......... ,o 
Figure 2: Scores for four museum descriptive texts 
into which the remaining facts in the list are to 
be embedded. The embedding mutation ran- 
domly selects a unit Ui from the sequence and 
an entity in its main fact. It then collects all 
the units mentioning this entity and randomly 
chooses one Uk. The list containing these two 
units \[Ui,Uk\] represents a random embedding 
and will be treated as a single unit in later op- 
erations. It takes the. position of Ui to produce 
a new sequence \[U~,U2,...,\[Ui,Uk\],...,U,\] and all 
repetitions outside \[Ui,U~:\] are removed. This 
sequence is then evaluated and ordered in the 
populat ion. 
The probabilities of apl)lying the three opera- 
tots are: 65% for crossow'r. 30% for embedding 
mutation and 5% for normal umtation. This is 
because the first two are more likely to produce 
sequences bearing desired properties by either 
combining the good bits of two sequences or 
performing a reasonable amount of embedding, 
whereas normal mutation is entirely random 2. 
5 Justifying the GA evaluation 
function 
The linguistic theories discussed in Section 2 
only give evidence in qualitative terms. For a 
GA-based planner to work, we have to come up 
with actual numbers that can be used to evalu- 
2The values for crossover and mutation rate used m 
our algorithm are fairly standard. 
191 
The small portable throne from the time of the Qianlong Emperor 1736-95 is mad e .... 
of ~acquer~d`wo~d~.wit~T.de~rati~n-in~-g~d`~and.red~It-was-use~-in-the-private.apartrn~r~`~ ..... " ......... 
of the Imperial Palaces. The cover from the reign of Jiaquing, 1796-1820 is woven in yellow 
silk, which is the imperial colour of the Qing Dynasty,1644-1911. It would have covered 
the throne when not in use. 
The design on the seat is a imperial five clawed dragon in a circular medallion. On the 
inside of the arm pieces are small shelves. Precious possessions can be placed in small shelves 
and can be studied as an aid to contemplation. 
Figure 3: A generated text scored the highest, with the embedded parts highlighted 
Score2 Score3 Score4 Score5 Score6 
Score1 .9567 .9337 .9631 .9419 .9515 
Score2 .9435 .8819 .9280 .9185 
Score3 .8650 .8462 .9574 
Score4 .9503 .8940 
Score5 .8486 
Table 2: Correlations between six raters 
ate an RS tree. Mellish et al. (1998a) present 
some scores for evaluating the basic features of a 
tree, but they make it clear that the scores are 
there for descriptive purpose, not for making 
any serious claim about the best way of evalu- 
ating RS trees. 
The methodology we adopted was that we 
took the existing evaluation function and ex- 
tended it to take into account features for local 
coherence, embedding and semantic paratmxis. 
This resulted in rater 1 in Table 1, which sat- 
isfied all the heuristics mentioned in Section 2. 
We manually broke down four human written 
museum descriptions into individual facts and 
relations and reconstructed sequences of facts 
with the same orderings and aggregations as in 
the original texts. We then used our evaluation 
flmction to score the RS trees built from these 
sequences. In the mean time. we ran the GA 
algorithm for 5000 iterations on the facts and 
relations for 10 times. The results are shown in 
Figure 2, where the four line styles correspond 
to the four texts. The jagged lines represent-the 
scores of the machine generated texts and the 
straight lines represent the scores for the corre- 
sponding human texts. 
All human texts were scored among the high- 
est and machine generated texts can get scores 
very close to human ones sometimes. Since the 
human texts were written and revised bv mu- 
seum experts, they can be treated as "'nearly 
best texts". The figure shows that the evalu- 
ation function based on our heuristics can find 
good and correct combinations. The reason for 
a relatively bad text being generated sometimes 
might be that really bad sequences were pro- 
duced at the beginning. This could be improved 
by using certain heuristics to get better initial 
sequences. Also when the number of facts be- 
comes larger, more iterations are needed to get 
readable texts. Figure 3 gives a text generated 
using rater 1. 
To justify our claim that it is the preferences 
among generation factors that decide the coher- 
ence of a text, we fed the preferences into a con- 
straint based program. If a feature can take a 
range of values, the program randomly selects 
a number in that range. A number of raters 
compatible with the constraints were generated 
and one of them is given in Table 1 as rater 2. 
We then generated all possible combinations, in- 
cluding embedding, of seven facts from a human 
text and used six randomly produced raters to 
score each of them. 
The .qualities .of the generated texts are nor- 
real distributed according to all raters. The 
raters distinguish between good and bad texts 
and they classify the majority of texts as of 
moderate quality and only very small percent- 
ages as very good or very bad texts. The be- 
haviours of the raters are very similar as the 
histograms are of roughly the same shape. 
192 
To see to what extent the six raters agree with 
each other, we calculated the Pearson correla- 
tion coefficient between them, which is shown 
in Table 2. We can claim from the table that 
for this data, the scores from the six raters cor- 
relate, and we have a fairly good chance to be- 
lieve that the six raters, randomly produced in 
a sense, agree with each other on evaluating the 
text and they measure basically the same thing. 
Daniel Marcu. 1997. From local to global coherence: 
. A_ bottom=up.approach' to. text:planning.. 'In Pro- 
ceedings of the Fourteenth National Conference on 
Artificial Intelligence, pages 629-635, Providence, 
Rhode Island. 
Chris Mellish, Alistair Knott, Jon Oberlander, 
and Mick O'Donnell. 1998a. Experiments using 
stochastic search for text planning. In Proceed- 
ings of the 9th International Workshop on Natural 
Language Generation, Ontario, Canada. 
6 Conclusions and future work 
..... Chris:MeUish,: Mick O'_Donnell,:.Jon Oberlander, and 
Alistair Knott. 1998b. An architecture for op- 
This paper describes an experiment with the 
preferences among features concerning aggrega- 
tion and text planning, in particular, we present 
an mechanism for how relevant features can be 
scored to contribute together to the planning of 
a coherent text. The statistical results partially 
justify our claim that it is the preferences among 
generation features that decide the coherence of 
a text. 
Our experiment could be extended in many 
ways, for example, validating the evaluation 
function through empirical analysis of human 
assessments of the generated texts, and us- 
ing more texts to test the correlation between 
raters. The architecture based on the Genetic 
Algorithm can also be used for testing interac- 
tions between or within other text generation 
modules. 

References 
Susan Brennan, Marilyn Walker Friedman, and Carl 
Pollard. 1987. A centering apporach to pronouns. 
In Proceedings of the 25th Annual Meeting of the 
Association for Computational Linguistics, pages 
155-162, Stanford, CA. 
Hua Cheng. 1998. Embedding new information into 
referring expressions. In Proceedings of COLING- 
A CL '98, pages 1478-1480, Montreal, Canada. 
Barbara Grosz and Candace Sidner. 1986. Atten- 
tions, intentions and the structure of discourse. 
Computational Linguistics, 12:175-204. 
Barbara Grosz, Aravind Joshi, and Scott Weinstein. 
1995. Centering: A framework for modelling the 
local coherence of discourse. Computational Lin- 
guistics, 21(2):203-226. 
Barbara Grosz. 1977. T-he representation and use of 
focus in dialogue understanding. Technical report 
151, SRI International. 
William Mann and Sandra Thompson. 198.7. 
Rhetorical structure theory: A theory of text or- 
ganization. Technical Report ISI/RR-87-190, In- 
formation Sciences Institute, University of South- 
ern California. 
Daniel Marcu. 1997. From local to global coherence:
A bottom-up approach to text planning. In Proceedings
of the 14th National Conference on 
Artificial Intelligence, pages 629-635, Providence,
Rhode Island.
Chris Mellish, Alistair Knott, Jon Oberlander,
and Mick O'Donnell. 1998a. Experiments using
stochastic search for text planning. In Proceedings
of the 9th International Workshop on Natural
Language Generation, Ontario, Canada. 
Chris Mellish, Mick O'Donnell, Jon Oberlander, and
Alistair Knott. 1998b. An architecture for op-
portunistic text generation. In Proceedings of the 
9th International Workshop on Natural Language 
Generation, Ontario, Canada. 
Marie Meteer. 1992. Expressibility and The Prob- 
lem of Efficient Text Planning. Communication 
in Artificial Intelligence. Pinter Publishers Lim- 
ited, London. 
Megan Moser and Johanna Moore. 1996. Toward a 
synthesis of two accounts of discourse structure. 
Computational Linguistics, 22(3):409-419. 
Jon Oberlander, Mick O'Donnell, Ali Knott, and 
Chris Mellish. 1998. Conversation in the mu- 
seum: Experiments in dynamic hypermedia with 
the intelligent labelling explorer. New Review of 
Hypermedia and Multimedia, 4:11-32. 
Jon Oberlander, Alistair Knott, Mick O'Donnell, 
and Chris Mellish. 1999. Beyond elaboration: 
Generating descriptive texts containing it-clefts. 
In T Sanders, J Schilperoord, and W Spooren, 
editors, Text Representation: Linguistic and Psy- 
cholinguistic Aspects. Benjamins, Amsterdam. 
Franck Panaget. 1997. Micro-planning: A uni- 
fied representation of lexical and grammatical re- 
sources. In Proceeding of the 6th European Work- 
shop on Natural Language Generation, pages 97- 
106. 
Massimo Poesio, Renata Vieira, and Simone Teufel. 
1997. Resolving bridging references in unre- 
stricted text. Research paper hcrc-rp87, Centre 
for Cognitive Science, University of Edinburgh. 
Michael Reape and Chris Mellish. 1999. Just what 
is aggregation anyway? In Proceedings of the 7th 
European Workshop on Natural Language Gener- 
ation, pages 20-29, Toulouse, France. 
Roger Schank. 1977. Rules and topics in conversa- 
tion. Cognitive Science, 1(1):421-441. 
Donia Scott and Clarisse Sieckenius de Souza. 1990. 
Getting the-message across in rst-based text gen- 
eration. In R. Dale, C. Mellish, and M. Zock, edi- 
tors, Current Research in Natural Language Gen- 
eration, pages 47-73. Academic Press. 
John Wilkinson. 1995. Aggregation in Natural Lan- 
guage Generation: Another Look. Technical re- 
port, Computer Science Department, University 
of \Vnterloo. 
