EXPERIMENTS USING STOCHASTIC SEARCH FOR TEXT PLANNING 
Chris Mellish, Alistair Knott, Jon Oberlander and Mick O'Donnell 
Department of Artificial Intelligence and Human Communication Research Centre, 
University of Edinburgh 1 
Abstract 
• Marcu has characterised an important and difficult problem in text planning: given a set of facts 
to convey and a set of rhetorical relations that can be used to link them together, how can one 
arrange this material so as to yield the best possible text? We describe experiments with a number 
of heuristic Search methods for this task. ' 
1 • Introduction: Text Planning 
1.1 The Task 
This paper presents some initial experiments using stochastic search methods for aspects of text 
planning. The work was motivated by the needs of the ILEX system for generating descriptions of 
museum artefacts (in particular, 20th Century jewellery) \[Mellish e t al 98\]. We present results on 
examPles semi-automatically generated from datastructures that exist within ILEX. 
Forming a set of facts about a piece of jewellery into a structure that yields a coherent text is 
a non-trivial problem. Rhetorical Structure Theory\[Mann and Thompson 87\] claims that a text 
is coherent just in case it can be analysed hierarchically in terms of relations between text spans. 
Much work in NLG makes the assumption that constructing something like an RS tree is a necessary 
step in the planning of a text. This work takes as its starting point Marcu's \[Marcu 97\] excellent 
• formalisation of RST and the problem of building legal RST trees, and for the purposes of this 
paper the phrase "text planning" will generally denote the task characterised by him. In this task, 
one is given a set of facts all of which should be included in a text and a set of relations between 
facts, some of which canbe included in the text. The task is to produce a legal RS tree using the 
facts and some relations (or the "best" such tree). 
Following the original work on RST and assumptions that have been commonly made in sub- 
sequent work, we wil ! assume that there is a fixed set of possible relations (we include "joint" as a 
second-class relation which can be applied to any two facts, but whose use is not preferred). Each 
relation has a nucleus and a satellite (we don't consider multiple nuclei or satellites here, apart 
from the case of "joint", which is essentially multinuclear). Each relation may be indicated by a 
distinctive "cue phrase", with the nucleus and satellite being realised in some fashion around it. 
Each relation has applicability conditions which can be tested between two atomic facts. For two 
complex text spans, a relation holds exactly when that relation holds between the nuclei of those 
spans. Relations can thus hold between text spans Of arbitrary size. 
Figure 1 shows an example of the form of the input that is used for the experiments •reported 
here. Each primitive "fact" is represented in terms of a subject, verb and complement (as well 
as a unique identifier). The "subject" is assumed to be the entity that the fact is "about". The 
approaches reported here have not yet been linked to a realisation component, and so the entities 
a 80 South Bridge, Edinburgh EH1 1HN. Email: {chrism, micko}~dai, ed. ac. uk, {alik, j on}ecogsc£, ed. ac. uk 
• 98 
I 
I 
I 
I 
I 
I 
•( !| 
:! 
I 
' " I 
I 
I 
i 
fact( 
fact( 
fact( 
fact( 
fact( 
'this item','is','a figurative jewel',f6). 
bleufort,'was','a french designer',f3). 
shiltredge,'was','a british designer',fT). 
'this item','was made by',bleufort,f8). 
titanittm,'is','a refractory metal',f4). 
rel (contrast, f7, f3, \[\] ). 
rel(elab,Fi,F2, \[\]) :- 
mentions (FI, O), 
mentions (F2, O) , 
\+ FI=F2. 
Figure 1: Example Input 
are represented Simply by canned phrases for readability (it is assumed that each entity in the 
domain has a fixed distinctive phrase that is always used for it). Relations are represented in 
terms Of the relation name, the nucleus and satellite facts and a list (in this example, empty) 
of precondition facts which need to have been assimilated before the :relation can be used (this 
represents an extension to Marcu's chcracterisation). This example uses the definition of (object- 
attribute) "elaboration" that we will be using consistently, namely that one fact can elaborate 
another if they have an entity in common (of course, there are other kinds of elaborations, but we 
would want to model them differently). 
1.2 Controlling Search in Text Planning 
There seem to be three main approaches to controlling the search for a good RS tree (or something 
similar). One is to restrict what relations can appear in the nucleus and satellite of others (for 
instance, using Hovy's \[Hovy 90\] idea of "growth points"): This is astep towards creating "schemas" 
for larger pieces of text. It can therefore be expected that it will produce very good results in 
restricted domains where limited text patterns are used, but that it will be hard to extend it to 
freer text types. The second idea is to use information about goals to limit possibilities. This 
is an element of Hovy's work but is more apparent in the planning work of Moore and Paris 
\[Moore and Paris 93\]. This second approach will work well if there are strong goals in the domain 
which really can influence textual decisions. This is not always the case. For instance, in our ILEX 
domain \[Mellish et al 98\] the system's goal is something very general like "say interesting things 
about item X/subject to length and coherence constraints". 
The third approach, most obviously exemplified by \[Marcu 97\], is to Use some form of explicit 
search through possible trees, guided by heuristics about tree quality. Marcu first of all attempts 
to find the best ordering of the facts. For every relation that could be indicated, constraints are 
generated saying what the order of the two facts involved should be and that the facts should be 
adjacent. The constraints are weighted according to attributes of rhetorical relations that have 
been determined empirically. A standard constraint satisfaction algorithm is used to find the linear 
sequence such that the total weight of the satisfied constraints is maximal. Once the sequence of 
facts is known, a general algorithm \[Marcu 96\] is used to construct all possible RS trees based on 
those facts. It is not clear how the best such tree is selected, though clearly theadjacency and 
order constraints-could in principle be reapplied in some way (possibly with other heuristics that 
Marcu has used in rhetorical parsing) to select a tree. 
We are interested in further developing the ideas of Marcu, but seek to address the following 
problems: 
1. It is not clear howscalable the approach is. Constraint satisfaction in general is intractable, 
99 
and having weighted constraints seems to make matters worse. Enumerating all RS trees 
that can be built on a given sequence of facts also has combinatorical problems. Marcu's 
approach may not be much better than one that builds all possible trees. Yet if there are 
enough relations to link any pair of facts (which, given the existence of elaboration, may often 
be nearly the case), the number of trees whose top nucleus are a specified'fact grows from 
336 to • 5040 to 95040 as the number of facts grows from 5 to 6 to 7. In our examples, we have 
more like 20-30 facts. 
2. As Marcu points out, the constraints on linear order only indirectly reflect requirements on 
the tree (becaus e related facts need not appear consecutively). Though in fact we will use - 
the idea of planning via a linear sequence later, we would like to experiment using measures 
of quality that are applied directly to the trees. We also have a number of factors that we 
would llke to take account of in the evaluation (see section 3 below). 
2 Stochastic Search 
Building a good RS tree is a search problem. Stochastic search methods are a form of heuristic 
search that use the following generic algorithm: 
i. Construct a set of random candidate Solutions. 
2. Until some time limit is reached, 
Randomly pick one or more items from the set, in such a way as to prefer items with 
the best "scores". 
Use these to generate one or more new random variations. 
Add these to the set, possibly removing less preferred items in order to keep the size 
constant. 
Examples Of stochastic search approaches are stochastic hillclimbing, simulated annealing and evol- 
utionary algorithms. The approaches differ according to factors like the size of the population of 
possible solutions, that is maintained, the operations for generating new possibilities and any spe- 
cial mechanisms for avoiding local maxima. They are similar toone another (and different from 
constraint satisfaction and enumeration approaches) in that they are heuristic (not guaranteed to 
find optimal solutions) and they are "anytime". That is, such an algorithm can be stopped at 
any•point and it will be able to yield at that point a result which is the best it has found so far. 
This is important for• NLG applications where interface considerations mean that texts have to be 
produced within a limited time. 
3 Evaluating RST trees• • 
A key requirement for the use of any stochastic search approach is the ability to: assess the quality 
of a possible solution. Thus we are forced to confront •directly the task of evaluating RST trees. 
We assign a candidate tree a score which is the sum of scores for particular features the tree 
may have. A positive score here indicates a good feature and a negative one indicates a bad one. 
We cannot make any claims to have the best way of evaluating RS trees. The problem is far too 
complex and our knowledge of the issues involved so meagre that only a token gesture can be made 
i00 - 
I 
i 
',I 
I 
i.l 
at this point. We offer the following evaluation scheme merely so that the basis of our experiments 
is clear and because we believe that some of the ideas are starting in the right direction. Here are 
the features that we score for: 
Topic and Interestingness We assume that the entity that the text is "about"is specified with 
the input. It is highly desirable that the "top nucleus" (most important nucleus) of the text be 
about this entity. Also we prefer texts that use interesting relations. We score as follows: 
-10 for a top nucleus not mentioning the subject of the text 
-30 for a joint relation 
+21 for a relation other than joint and elaboration 
• Size of Substructures - Scott and de Souza \[Scott and de Souza 90\] say that the greater the 
amount of intervening text between the propositions of a relation, the more difficult it will be to 
reconstruct its message. We score as follows: 
-4 for each fact that will come textually between a satellite and its nucleus 
Constraints on Information Ordering Our relations have preconditions which are facts that 
should be conveyed before them. we score as follows: 
-20 for an unsatisfied precondition for a relation 
Focus Movement We do nothave a complex model of focus development through the text, 
though development of such a model would be worthwhile. As McKeown and others have done, we 
prefer certain transitions over others. If consecutive facts mention the same entities or verb, the 
prospects for aggregation are greater, and this is usually desirable. We score as follows: 
-9 for a fact (apart from the first) not mentioning any previously mentioned entity 
-3 for a fact not mentioning any entity in the previous fact, but whose subject is a 
previously mentioned entity 
• +3 for a fact retaining the subject of the last fact as its subject • 
+3 for a fact using the same verb as the previous one 
Object Introduction When an entity is first introduced as the subject of a fact, it is usual for 
that to be a very general statement about the entity. Preferring this introduces a mild schema-like 
influence to the system. We score as follows: 
+3 for the first fact with a given entity as subject having verb "is" 
4 Using Stochastic Search for Text Planning 
Using the above evaluation metric for RS trees, we have experimented with a range• of stochastic 
search methods. Space does not permit us to discuss more than one initial experiment in this 
section. In the next section, we describe a Couple of methods based on genetic algorithms which 
proved more productive. 
I01 
4.1 Subtree Swapping 
The subtree swapping approach produces new trees by swapping random subtrees in a candidate 
solution. It works as follows: 
1. Initialise with a tree for each combination of interesting (non-elaboration) relations, with any 
fact only appearing in one. Make into a complete tree by combining together these relations 
and any unused facts with "joint" relations (or better ones if available). 
2. Repeatedly select a random tree and swap over two random subtrees, repairing all relations. 
Add the new tree to the population. 
When two subtrees are swapped over in an RS tree, some of the relations indicated in the tree 
no longer apply (i:e. those higher relations that make use of the nuclei of the subtrees). These 
are "repaired" by in each case selecting the "best" valid relation that really relates the top nuclei 
(i.e. a non-elaboration relation is chosen if possible, otherwise an elaboration if that is valid, with 
"joint" as a last resort). 
We investigated variations on this algorithml including having initial random balanced trees 
(including the "best" relation at each point) and focussing the subtree swapping On subtrees that 
contributed to bad scores, :but the above algorithm was the one that seemed most successful. 
4.2 Initial Results• .... : : 
Figure 2 shows an example tex t generated by subtree swapping. Note that we have taken liberties 
in editing by hand the surface text (for instance, by introducing better referring expressions and 
aggregation). For clarity, coreference has been indicated by subscripts. The ordering of the material 
and the use of rhetorical relations "are the only things which are determined by the algorithm. 
Results for subtree swapping are shown together with later results in Figure 5 (the example text 
shown for subtree swapping is for the item named j-342540). The most obvious feature of these 
results is the huge variability of the results , which suggests that there are many local maxima in 
the search space. Looking at the texts produced, we can see a number of problems. If there is only 
• one way smoothly to include a fact in the text, the chance of finding it by random subtree swapping 
is very low. The Same goes for fixing other local problems in the text. The introduction of "the 
previous jewel" is an example of this. This entity can only be introduced elegantly through the fact 
that it, like the current item, is encrusted with jewels. The text is also still suffering from material 
getting between a satellite and its nucleus. For instance, there is a relation (indicated by the colon) 
between "It is encrusted with jewels" and "it has silver links encrusted asymmetrically...", but this 
is weakened by the presence of "and is an Organic style jewel" in the middle). 
The trouble is that subtree swapping needs incrementally to acquire all good features not 
present in whichever initial tree develops into the best solution. It can only acquire these features 
"acCidentally" and the chances of stumbling on them are small. Different initial trees will contain 
• different good fragments, and it seems desirable to be able to combine the good parts of different 
• solutions. This motivates using some sort of Crossover operation that can combine elements of two 
solutions into a new one \[Goldberg 89\]. But it is not immediately clear how crossover could work 
on two RS trees, tn particular, two chosen trees will rarely have non-trivial subtrees with equal 
fringes. Their way of breaking up the material may be so different that it is hard to imagine how 
one could combine elements of both. i .- !!i 
I 
• : 'i 
I 
,I 
I 
I 
\[l 
!1 
I 
I 
I 
i 
I 
l 
This jewel/ is made from diamonds, yellow metal, pearls, oxidized white metal and 
opals. 
It~ was made in 1976 and was made in London. 
This jewe4 draws on natural themes for inspiration: itl uses natural pearls. 
Iti was made by Flockinger who is an English designer. 
Flockinger lived in London which is a city. 
This jeweli is a necklace and is set with jewels. 
Iti is encrusted with jewels and is an Organic style jewel: iti has silver links encrusted 
asymetrically with pearls and diamonds. 
Indeed, Organic style jewels are usually encrusted with jewels. 
Organic style jewels usually draw On natural themes for inspiration and are made up of 
asymmetrical shapes. 
Organic style jewels usually have a coarse texture. 
• This jewel/is 72.0 cm long. 
The previous \]ewelj has little diamonds scattered around its edges and has an encrusted 
bezel. Itj is encrusted with jewels: itj features diamonds encrusted on a natural shell. 
Figure 2: Example Text from Subtree Swapping 
5 Restricting the Space of RST Trees 
As a way of making a crossover operation conceivable, our first step has been to reduce the planning 
problem to that of planning the sequential order of the facts (in a way that echoes Marcu's approach 
to some extent). We have done this by making certain restrictions on the RS trees that we are 
prepared to build. In particular, we make the following assumptions: 
• 1. The nucleus and satellite of a non-joint relation can never be separated. 
2. "Joint" relations are used to connect unrelated paragraphs. 
With these assumptions, an RS tree is characterised (almost) by the sequence of facts at its leaves. 
Indeed, we have an algorithm that almost deterministically builds a tree from a sequence of facts, 
according to these principles. • (The algorithm is not completely deterministic, • because there may 
be more than one non-elaboration relation that can be used with two given facts as nucleus and 
satellite - our evaluation function won't, of course, differentiate between these). 
The algorithm for building a tree from a sequence essentially makes a tree that can be processed 
by a reader with minimal short-term memory. The tree will be right-branching and if the reader 
just remembers the last fact at any point, then they can follow the connection between the text so 
far and the next fact 2 Interestingly, Marcu uses "right skew" to b disambiguate between alternative ~- 
tree s produced in rhetorical parsing. Here we are setting it as a much harder constraint. The only 
2In fact, there is local left-branching for (non-nested) relations whose satellite is presented first. Such relations 
are often presented using embedded clauses in a way that signals the deviation from right-branching clearly to the 
reader. 
103 
exception is "joint" relations, which can join together texts of any size, but since there is no real 
relation involved in them there is no memory load in interpreting them. 
The first two assumptions above make fundamental use of the order in which facts will appear 
in the text. For simplicity, we assume that every relation has a fixed order Of nucleus and satellite 
(though this assumption could be relaxed). The approach i s controversial in that it takes into 
account realisati0n order in the criterion for a legal tree. It is likely that the above assumptions 
will not apply equally well to all types of text. Still, they mean that the planningproblem can :be 
reduced to that of planning a sequence. The next experiments were an attempt to evaluate this 
idea. 
• 6 Using a Genetic Algorithm 
The genetic algorithm we used takes the following form: 
1. Enumerate a set of random initial sequences by loosely following sequences of facts 
where consecutive facts mention the same entity. 
2. Evaluate sequences by evaluating the trees they give rise to. 
- 3. Perform mutation and crossover on the sequences, with mutation • having a relatively 
small probability. 
4. When the "best'/ sequence has not changed for a time, invoke mutation repeatedly 
until it does. 
5. Stop after a given number of iterations, and return the tree for the "best"• sequence. 
Notice that although the algorithm manipulates sequences, the evaluation is one that operate s on 
trees. Mutation is a unary operation which, given one sequence, generates a new one. Crossover is 
binary in that it generates new solution(s ) based on two existing ones. The choice of mutation and 
crossover operations depends on how the sequences are internally represented and should facilitate 
the exchange of useful subparts of solutions. Two different representations have been tried so far. 
The relevant features are summariSed in Figure 3. 
6.1 Ordinal Representation 
The ordinal representation \[Michalewicz 92\] assumes that ~ there is an initial canonical sequence of 
facts (in the figure, this is assumed to be 1,2,3,4). A given sequence is represented by a sequence 
of numbers, where the ith element indicates the position of the ith element of the sequence in 
that canonical sequence with all previous elements deleted. So the ith element is always a number 
between 1 and n + 1 - i, where n is the length of the sequence. Mutation is implemented by a 
change of a random element to a random legal value. •Crossover (here) is implemented by two-point 
crossover - the material between two random points •of the sequences (the same points for both)is 
swapped over, yielding two new sequences. The ordina ! representation has been used extensively 
for tasks such as the travelling salesman problem , and it has the advantage that the crossover 
operation is particulariy simple. 
6.2 Path Representation 
in many ways, this is a more obvious encoding, though the operations are chosen to reflect the. 
' intuition that order and adjacency information should generally be maintained from old solution(s) 
.104 
I 
i 
I 
I 
I 
I 
!1 
it 
!1 
!1 
| 
• !,1 
Ordinal Encoding 
1.3.2., , I112 II I1 I 
Second remaining item 
Mutation 
111211111 " li1311111 
Random position changes to a random legal value 
Crossover 
. e,, 
I112 I1 I1 I •1312•12 I1 I--'-I11212 II I 
Exchange material between two random positions 
131211 II 
Path Encoding 
I;3,2,4 
Mutation 
111213 I, I 
v A 
Crossover 
• Ill 1312t,1 
II 13 It 12 I 
Slide random element to random place 
121~ 13 I11 " I~ 12 13 Ill 
1 
Insert sequence at random point, deleting duplicates outside 
Figure 3: Ordinal and Path Representations 
to the new ones they give rise to. A sequence of facts is represented simply as that sequence. 
Mutation selects a random element, removes it from the sequence and then inserts it again in 
a random place. Crossover inserts a random subsequence of one solution into another, deleting 
duplicates that occur outside the inserted subsequence. 
6.3 Results 
• Figure 4 shows an example text produced using the path encoding operations (for j-342540, after 
2000 iterations, just under 2 minutes, score I80). The same remarks about hand editing apply as 
before. Figure 5 summarises the results for subtree swapping and the two genetic algorithms on a 
set of examples. These results summarise the mean and standard deviations of the scores of the 
system run 10 times. The system was tried with a limit of 2000 and 4000 •iterations around the 
main loop of the algorithm. These took about 2 and 4 minutes respectively. With each example 
problem we have specified the number of facts, the number of elaboration relations and the number 
of non-elaboration relations. Note that there is not a very clear basis for comparison between 
105 
This jewel/is made from diamonds and yellow metals. 
It/was made by Flockinger, who was an •English designer. 
• Flockinger lived in London, which is a city. 
This jeweli was made in London. 
It/is a necklace. 
Iti is made from oxidized white metal, pearls and opals. 
It/is set with jewels. 
This jewel/ is encrusted with jewels: it/ has silver links encrusted asymetrically with 
pearls and diamonds. - 
This jewel/was made in 1976. 
Iti is an Organic style jewel and is 72.0 cm long. 
Iti draws on natural themes for inspiration: it/ uses natural pearls. Indeed, Organic 
style jewels usually draw on natural themes for inspiration. 
Organic style jewels usually have • a coarse texture, are usually made up of asymmetrical 
• shapes and are usually encrusted with jewels. 
The • previous jewelj is encrusted With jewels: itj features diamonds encrusted on a 
natural shell. 
. Itj has little diamonds scattered around its edgesand an encrusted bezel. 
Figure 4: Text Planned by GA 
algorithms, since each algorithm performs different operations during an "iteration". Nevertheless, 
since iterations take roughly the same amount of time one can get a rough idea of the relative 
performance. 
The •example text is now in a single paragraph, with a clear link from each sentence to the 
• previous ones. From the numerical results, one can see that there is imuch less variability than 
before. This is mainly because the rigid tree-building constraints prevent really bad trees • being 
built and so the worst results are less bad. The results are also significantly better than for subtree 
swapping, with the edge-sensitive representation clearly winning. • 
7 Discussion 
It is necessary to be careful in evaluating:these results, which are 0nly as good as the evaluation 
function. This is certainly flawed in major ways. The texts are of a specific type, there are only 
three of them and we have not used all rhetorical relations. Some independent evaluation by human 
readers is imperative at this point. The texts are especially limited by the fact that there is no 
account taken Of the possibilities for aggregation , •embedding etc. in the trees that are produced: 
NevertheleSs • the approach looks promising enough that it is a real candidate to be used with the 
ILEX syste m. Future work needs to look at improving the characterisation of good trees and if 
possible •introducing more natural crossover/mutation operations. Future work could also consider 
extending the scope of the algorithm to deal with aspects of content determination as well as• 
structuring. 
106 
I 
I 
!l 
i 
il 
,I 
i 
i 
! 
\] 
!.~ 
! 
it 
i 
Subtree Swapping 2000 Iterations 4000 Iterations 
Item facts elabs rels Mean S.D. Mean S.D. 
j-342540 28 298 13 -38.9 27.7 -15.0 39.3 
j-990302 25 297 13 18.5 32.6 31.6 27.9 
j-990811 24 274 6 -50.7 33.6 -2.2 27.6 
Ordinal Representation 2000 Iterations 4000 Iterations 
Item facts elabs rels Mean S.D. Mean S.D. 
j-342540 28 298 13 110.2 25.6 127.3 26.1 
j-990302 25 297 13 109.2 13.6 115.0 18.7 
j-990811 24 274 6 57.0 17.6 66.7 17.8 
Path Representation 2000 Iterations 4000 Iterations 
Item facts elabs rels Mean S.D. Mean S.D. 
j-342540 28 298 13 158.4 22.7 171.3 20.1 
j-990302 25 297 13• 175.0 19.3 192.9 13.7 
j:990811 24 274 6 90.7 11.4 104.0 17.3 
Figure 5: Results for 3 Algorithms •. 
8 Acknowledgements 
The ILEX project is supported by EPSRC grant GR/K53321. We acknowledge the valuable as- 
sistence of the National Museums of Scotland and the useful advice of Andrew Tuson. 

References 
\[Goldberg 89\] Goldberg, D., Genetic Algorithms in Search, Optimization and Machine Learning, Addison- 
Wesley, 1989. 
\[Hovy 90\] Hovy, E., "Unresolved Issues in Paragraph Planning", in Dale, R., Mellish, C. and Zock, M., 
• Current Research in Natural Language Generation, Academic Press, 1990, pp17.45. 
\[Mann and Thompson 87\] Mann, W. and Thompson, S., "Rhetorical Structure Theory: Description and 
Construction of Text Structures", in Kempen, G., Ed., Natural Language Generation: New Results in 
Artificial Intelligence, Psychology and Linguistics, Dordrecht: Nijhoff, 1987. • 
\[Marcu 96\] Marcu, D., "Building up Rhetorical Struicture Trees", Proceedings of AAAI-96, American As- 
sociation for Artificial Intelligence, 1996, pp1069-1074: 
\[Marcu 97\] Marcu, D., "From Local to Global Coherence: A Bottom-up Approach to Text •Planning", 
Proceedings of AAAI-97, American Association for Artificial Intelligence, 1997, pp629-635. 
\[Mellish et a198\] Mellish, C., O'Donnell, M., Oberlander, J. and Knott, A., "An Architecture for Oppor- 
tunistic Text Generation", Proceedings of INLGW-98, 1998. 
\[Michalewicz 92\] Michalewicz, Z., Geneti c Algorithm 4- Data Structures = Evolution Programs, Springer 
Verlag, 1992. - 
\[Moore and Paris 93\] Moore, J. and Paris, C., "Planning Texts for Advisory Dialogues: Capturing Inten .... 
tional and Rhetorical Information", Computational Linguistics Vol 19, No 4, 1993, pp651-694. 
\[Scott and de Souza 90\] Scott, D. and de Souza, C., "Getting the Message Across in RST-Based Text Gener- 
ation", in Dale, R., Mellish, C. and Zock, M., Eds., Current Research in Natural Language Generation, 
Academic Press, 1990, pp47-73. 
