Introduction to the Special Issue on 
Natural Language Generation 
Robert Dale* 
Macquarie University 
Donia Scow 
University of Brighton 
Barbara Di Eugenio t 
University of Pittsburgh 
1. Introduction 
There are two sides to natural language processing. On the one hand, work in natural 
language understanding is concerned with the mapping from some surface represen- 
tation of linguistic material expressed as speech or text--to an underlying repre- 
sentation of the meaning carried by that surface representation. But there is also the 
question of how one maps from some underlying representation of meaning into text 
or speech: this is the domain of natural language generation. 
Whether our end-goal is the construction of artifacts that use natural languages 
intelligently, the formal characterization of phenomena in human languages, or the 
computational modeling of the human language processing mechanism, we cannot 
ignore the fact that language is both spoken (or written) and heard (or read). Both are 
equally large and important problems, but the literature contains much less work on 
natural language generation (NLG) than it does on natural language understanding 
(NLU). There are many reasons why this might be so, although clearly an important 
one is that researchers in natural language understanding in some sense start out with 
a more well-defined task: the input is known, and there is a lot of it around. This is not 
the case in natural language generation: there, it is the desired output that is known, 
but the input is an unknown; and while the world is awash with text waiting to be 
processed, there are fewer instances of what we might consider appropriate inputs for 
the process of natural language generation. For researchers in the field, this highlights 
the fundamental question that always has to be asked: What do we generate from? 
Despite this problem, the natural language generation community is a thriving 
one, with a research base that has been developing steadily--although perhaps at a 
slower pace because of the smaller size of the community--for just as long as work 
in natural language understanding. It should not be forgotten that much of NLP has 
its origins in the early work on machine translation in the 1950s; and that to carry out 
machine translation, one has to not only analyze existing texts but also to generate 
new ones. The early machine translation experiments, however, did not recognize the 
problems that give modern work in NLG its particular character. The first significant 
pieces of work in the field appeared during the 1970s; in particular, Goldman's work 
on the problem of lexicalizing underlying conceptual material (Goldman 1974) and 
* School of Mathematics, Physics, Computing and Electronics, Sydney NSW 2109, Australia 
t Learning Research and Development Center, 3939 O'Hara Street, Pittsburgh, PA 15260, U.S.A. 
:~ Information Technology Research Institute, Lewes Road, Brighton BN2 4GJ, UK 
(~) 1998 Association for Computational Linguistics 
Computational Linguistics Volume 24, Number 3 
Davey's work on the generation of paragraph-long descriptions of tic-tac-toe games 
(Davey 1979) were among the first to focus on issues unique to NLG. The field really 
took off, however, in the 1980s; for those working in NLG, the decade began with a 
bang, and the Ph.D. theses of McDonald (1980), Appelt (1981), and McKeown (1982) 
have had a lasting impact on the shape of the field) 
But what has happened in the last fifteen years since those major pieces of work 
first appeared? Although one does find articles on NLG in the pages of Computational 
Linguistics and other journals in the field, and papers on generation do appear at the 
major NLP conferences, the quantity and range of work being carried out in NLG 
tends to be underrepresented in these forums. Instead, the community has tended 
to present its results at the two biennial series of workshops--one European and one 
international--that have sprung up in the last ten years. Many of these workshops have 
led to books: see Kempen (1987); McDonald and Bolc (1988); Zock and Sabah (1988); 
Dale, Mellish, and Zock (1990); Paris, Swartout, and Mann (1991); Dale et al. (1992); 
Horacek and Zock (1993); and Adorni and Zock (1996). This special issue of Computa- 
tional Linguistics was inspired by discussions at the International Workshop on Natural 
Language Generation held in Herstmonceux in 1996; the aim of the volume you are 
reading is to show the wider computational linguistics community something of the 
range of activities in NLG. 
2. Some Perspectives on Natural Language Generation 
What is natural language generation about? A definition offered by McDonald (1987, 
983) over ten years ago has stood the test of time: 
Natural language generation is the process of deliberately constructing a natural 
language text in order to meet specified communicative goals. 
A more recent definition with a slightly different emphasis can be found in Reiter and 
Dale (1997, 57): 
Natural language generation is the subfield of artificial intelligence and 
computational linguistics that is concerned with the construction of computer 
systems that can produce understandable texts in... human languages from 
some underlying non-linguistic representation of information. 
Both definitions pick out some of the foci of interest that give work in NLG its distinc- 
tive flavor. From the first we note the emphasis on deliberate choice as the fundamental 
operation that underlies much work in the area, and on the generation of texts as 
opposed to single sentences; from the second, we note the emphasis on underlying 
representations of information that may be nonlinguistic in nature. Each of these points 
bears some elaboration: 
In work in NLG, a major concern is that of choosing between different 
ways of doing things, as the same content can often be expressed in 
many different ways. Although some of these choices may indeed be 
arbitrary, there is a view in NLG that a great many are not, and the 
choices between different ways to say things---different ways to structure 
a text, different ways to refer to objects, the use of different syntactic 
1 The latter two works are more widely available in revised form as Appelt (1985) and McKeown (1985). 
346 
Dale, Di Eugenio, and Scott Introduction 
constructions, and of different words to realize underlying 
concepts--need to be motivated in some way. Much research in NLG is 
oriented towards uncovering those motivations. 
There is a sense in which work in NLU tends to start with the sentence 
as the principal focus of inquiry. However, for much work in NLG, the 
primary focus is the text or discourse: although there are many important 
issues involved in the generation of sentential forms, those working in 
NLG research have long accepted that discourse-level issues are just as 
important, and probably more so. This relates to the previous point: it is 
often only by considering the context within which a sentence is being 
generated that the appropriate choice of surface form can be made. 
The input representation provided to an NLG system may be symbolic 
(for example, an expert system knowledge base) or numeric (for 
example, a database containing stock market prices) but it is generally 
nonlinguistic in nature. Early work in the field relied on the use of 
hand-crafted knowledge sources, which sometimes meant that the 
representations used embodied unspoken assumptions. More recent 
work has been able to take advantage of representations created for other 
purposes; using these as the input to the generation process reinforces 
the realization that the elements of the underlying representation may 
not correspond in a straightforward way to words and sentences. 
Much work in NLG thus concerns itself with pragmatics and discourse-level consider- 
ations. Interestingly, these too have been somewhat underrepresented in the standard 
computational linguistics forums, where the bulk of the work carried out is often in 
the area of well-specified and rigorous formal treatments of sentential phenomena. 
There is an important point here that bears emphasizing: natural language gener- 
ation is not the inverse of the process of parsing. 2 Those working in NLG generally 
break down the process of generating a text into a number of stages, and it is only 
the last of these--generally referred to as surface realization--that corresponds to any- 
thing like the inverse of parsing. If we want to seek the mirror-image of work in NLG 
within research in natural language understanding, we have to consider the entire 
analysis process, all the way through to plan recognition in multisentential discourses 
or dialogues. 
This is perhaps an appropriate place to review what the task of NLG is now 
commonly seen to involve: 
First, there is the question of content determination: deciding what to 
say. This impacts at both macro and micro levels. At the macro level 
researchers in NLG are concerned with how the content of a 
multisentential text, or of a turn in a dialogue, can be determined. At a 
micro level, researchers are concerned with how the content of 
appropriate referring expressions can be worked out. In each case the 
problem is how to select the right information from that which is 
available; it is rarely appropriate to say everything we could say. 
2 It should be noted, however, that there is a body of work that looks at the use of bidirectional 
grammars, where a common declarative representation of grarmnatical knowledge is used both for parsing and for realization; see, for example, Shieber et al. (1990). 
347 
Computational Linguistics Volume 24, Number 3 
There is also the question of text structure: texts are not just random 
collections of sentences; they exhibit a structure that plays a key role in 
conveying their meaning. Researchers in NLG are concerned with 
elucidating mechanisms for determining the most appropriate structures 
to use in particular circumstances, and with working out how the 
information to be conveyed can best be packaged into paragraph- and 
sentence-sized chunks. 
Closer to the kinds of issues that concern those working in parsing, there 
are the problems of surface realization and lexicalization: once the 
content of individual sentences has been determined, this still has to be 
mapped into morphologically and grammatically well-formed words and 
sentences. Where the underlying representation expresses informational 
elements at a granularity that does not map easily into words, decisions 
about how to lexicalize the conceptual material have to be taken. 
These are the kinds of issues that have driven much research in NLG over the 
last 15 years. Our understanding of the issues has come a long way in that time. This 
issue of Computational Linguistics contains what is no more than a snapshot of work in 
the field at the current time; it should be read against the background of the broader 
picture we have attempted to sketch here, albeit briefly. In the next section, we provide 
short summaries of the papers collected together in this special issue. 
3. An Overview of the Issue 
From the 25 papers originally submitted to the special issue, our reviewers helped 
us eventually select five. There were many more papers that, given space, we would 
have included; we hope that some of these will appear in subsequent regular issues 
of Computational Linguistics. 
3.1 Chu-Carroll and Carberry 
As we mentioned earlier, "deciding what to say" is a key issue in NLG. Chu-Carroll 
and Carberry's paper focuses on strategies for selecting the content of responses in 
collaborative planning dialogues. The authors concentrate on situations in which the 
system and the user have different beliefs that they attempt to reconcile, namely: cases 
when the system needs to gather further information in order to decide whether to 
accept a proposal from the user, and cases when the system must negotiate with the 
user to resolve a detected conflict in beliefs. In both cases, the implemented algorithms 
identify the subset of beliefs that the system believes will most effectively help solve 
either its uncertainty or the conflict in beliefs; further, the system chooses an appro- 
priate strategy and produces a response that initiates a subdialogue addressing the 
:impasse in conversation. 
Two other points deserve special note. First, the computational model is based on 
a small but convincing corpus study. Second, the authors conducted a formal, even 
if limited, evaluation of their prototype implementation. The evaluation consists of 
human raters grading the system's actual response and some distractors, obtained 
by selectively altering the system's response generation strategies. The evaluation is 
suggestive that the proposed strategies and their implementation are effective. 
348 
Dale, Di Eugenio, and Scott Introduction 
3.2 Stede 
Addressing the issue of "deciding how to say it," Stede focuses on the role of the 
lexicon and of lexical choice within an NLG system. More specifically, Stede describes 
how his approach can generate verbal alternations that change the aspectual category 
of the verb. One such alternation is the causative, as in The mechanic drained the oil from 
the engine, the causative form of The oil drained from the engine. Stede takes the lexicon to 
be the central device for mapping between domain representations and intermediate 
semantic representations. Alternations are generated by applying one or more rules 
in a predetermined order to a basic lexical form; the choice of a specific alternation is 
determined by parameters such as salience. The intermediate semantic representation 
of a sentence is the input for the surface generator--in this case, Penman (Penman 
group 1989). 
Interesting aspects of Stede's approach are the capacity of his system to generate 
fine-grained distinctions of meaning, and the attention paid to both linguistic con- 
straints and computational concerns. Although only English is discussed in this paper, 
Stede's system uses the same mechanism to generate alternations in German as well. 
3.3 Mittal, Moore, Carenini, and Roth 
Generation technology is now increasingly finding a place in applied systems. One 
such application is described by Mittal, Moore, Carenini, and Roth, who have devel- 
oped a system to generate captions to accompany the graphical presentations produced 
by SAGE (Roth et al. 1994). It is well known that the interpretation of even simple, con- 
ventional graphics can be difficult without accompanying textual pointers (e.g., keys, 
labels of axes, and the like). SAtE is innovative in its ability to produce novel graphics 
for highly abstract and complex data. The comprehension of these presentations is 
often heavily reliant on captions: extended textual descriptions of the relation of the 
presentation to the data it depicts. 
Mittal et al. show how a SAGE graphic, together with information about the per- 
ceptual complexity of its elements and the structure of its underlying data, can be used 
to generate an effective multisentential caption. This is demonstrated through exam- 
ples in the domain of housing sales; however, with the exception of the lexicon, the 
caption generator is fully domain independent. Although the system has not yet been 
formally evaluated, we are told that users of SAGE report that the generated captions 
contribute positively to their understanding of complex graphical presentations. 
3.4 Radev and McKeown 
Automated text summarization is a practical problem of increasing interest, especially 
with the ever-widening dissemination of the World Wide Web; this is an obvious area 
where NLG can contribute. Radev and McKeown describe an application of NLG 
techniques towards the end of producing summaries of a kind that are, as the au- 
thors argue, beyond the scope of current statistical summarizers. There are two main 
elements to their approach. The first is the use of "surrunarization operators" that com- 
pare data structures containing information derived from different sources and thus 
allow the system to produce summaries of several input messages; the second is the 
use of a technique for identifying proper names and related descriptions from on-line 
text so that these can be used to extend the descriptions provided in summaries. 
The data structures used as input to generate the summary texts are filled MUC- 
style templates; the task of identifying key information in source texts and extracting 
it has already been carried out. The paper thus provides an excellent application of 
established technologies, with new mechanisms being developed to complete the pic- 
349 
Computational Linguistics Volume 24, Number 3 
ture; the work shows well how NLG techniques can make a real difference to the 
important task area of summarization. 
3.50berlander 
Texts are generated to be read, and while generators can provide a range of texts for a 
given context, the question of what expression is most appropriate remains nontrivial. 
Typically nowadays, the designers of NLG systems tune their generators to produce 
expressions compatible with those found in a corpus of "good" exemplars from the 
domain in question (see, e.g., Scott and Power \[1994\] and Paris et al. \[1995\]). But this 
approach is not always possible (e.g., for novel domains), and even in cases where 
there is an available corpus, judgments of quality can often only be made by appeal- 
ing to convention. Psycholinguistic studies suggest themselves as a useful source of 
guidance but this too can be problematic: what speakers or writers typically produce 
often conflicts with their preferences as perceivers, as shown, for example, with re- 
gard to referring expressions by Stevenson and her colleagues (Stevenson, Crawley, 
and Kleinman 1994; Stevenson and Urbanowicz 1995). 
Oberlander discusses this apparent paradox with reference to the generation of 
referring expressions--in particular, the suggestion by Dale and Reiter (1995) that 
generation algorithms for definite noun phrases should be based on observations about 
human language production rather than on a strict observation of the Gricean maxims 
(Grice 1975). Oberlander calls this the Spike Lee maxim: Do the right thing--where 
"right" is that which is human and simple. He shows that, when generating referring 
expressions, we can't always tell whether the right thing is to mimic the preferences 
of language producers or language perceivers, since these preferences often conflict. 
He argues that until we develop a more sophisticated view of the expectations of 
speakers and hearers, developers of NLG systems should probably stick to the Spike 
Lee maxim: even with its known limitations it produces more natural results than are 
achieved by following a strict interpretation of the Gricean maxims--and we would 
all agree that even that is better than the Cole Porter maxim. 3 
4. Future Directions for Research in Natural Language Generation 
We said earlier that NLG research has come a long way since its beginnings, but there 
is still a long way to go. What does the future hold? Crystal-ball gazing is always a 
risky business, but on the basis of our experience and some of the issues that arise 
both in the work presented here and in other submissions to the special issue, we 
would suggest the following aspects of NLG will be seen as important areas in the 
next five years. 
Microplanning. Ever since Thompson (1977), there has been a tendency to see NLG 
as involving two problems, which Thompson characterized as being concerned with 
decisions of strategy and tactics: in short, questions about what to say and questions 
about how to say it. In the field, this translated into work in the two areas of text 
planning and linguistic realization, with researchers often declaring themselves as 
working on one or the other. In more recent years, there has been the realization that 
something is required in the middle; this was most notably expressed in Meteer's work 
on what she called "the generation gap" (Meteer 1990). This has given rise to a body 
of work that explores questions of what is often referred to as microplanning: once a 
3 This being: Anything goes. 
350 
Dale, Di Eugenio, and Scott Introduction 
text planning process has worked out the overall structure of a text and the content 
to be conveyed, how is this information packaged into sentences? Serious work here 
has only just begun: there are a great many unresolved issues, and in many cases the 
questions themselves are unclear. 
Multimodal generation. Real text is not disembodied. It always appears in context, and 
in particular within some medium--for example, on a page, on a screen, or in a 
speech stream. As soon as we begin to consider the generation of text in context, 
we immediately have to countenance issues of typography and orthography (for the 
written form) and prosody (for the spoken form). These questions can rarely be dealt 
with as afterthoughts. This is perhaps most obvious in the case of systems that generate 
both text and graphics and attempt to combine these in sensible ways. We predict that 
the World Wide Web will be a major factor in forcing some of the issues here: if systems 
are to automatically generate the text on Web pages (see, for example, Milosavljevic 
and Dale \[1996\]), then they also need to consider other elements of that container. 
Reusable resources. It may be an indication of a maturing of some subareas of NLG 
research that we are now in a position where there are reusable components for par- 
ticular tasks. Specifically, three linguistic realization packages, FUF/SURGE (Elhadad 
1993a, 1993b; Elhadad and Robin 1996), PENMAN/NIGEL (Penman group 1989), and 
its descendant KPML/NIGEL (Bateman 1997), are widely used in the field. For any- 
thing other than simple applications, it is now questionable whether it makes sense to 
build a linguistic realization component from scratch. We may expect other kinds of 
reusable components to be developed within the research community within the next 
5-10 years; it is developments of this kind that signal significant progress, since being 
able to reuse the work of others obviously has the potential to increase research pro- 
ductivity. In related developments, there is a growing interest within the community 
in defining a reference architecture for NLG; if successful, this is likely to stimulate 
further research and development in NLG through the provision of a modular baseline 
for development, comparison, and evaluation. 
Evaluation. Although there have been attempts at the evaluation of NLG techniques 
and systems in the past, formal evaluation has only recently come to the fore. For 
example, systems have been evaluated by using human judges to assess the quality 
of the texts produced (Lester and Porter 1997; Chu-Carroll and Carberry, this issue); 
by comparing the system's performance to that of humans (Yeh and Mellish 1997); 
by corpus-based evaluation (Robin and McKeown 1996); and indirectly through "task 
efficacy" measures (Young 1997). The major stumbling block for progress is determin- 
ing what metrics and methods should be used: for example, how can the quality of 
an output text be measured? Because of the different nature of the task, it is unlikely 
that methods that have been used in NLU, such as the evaluation process adopted 
in the Message Understanding Conferences, can be carried over to generation. Dale 
and Mellish (1998) suggest that the NLG community could make progress by devis- 
ing specific evaluation methods for NLG subtasks such as content determination, text 
structuring, and realization; this "glass box" approach is likely to result in a clearer 
understanding of how to evaluate NLG systems as a whole. 
The particular foci we have just outlined are specific to work in NLG. However, 
just as corpus-based methods have become very important in NLU research, we may 
expect this to happen increasingly in work on NLG too. Raw or coded text has been 
used by researchers to investigate strategies in a number of different areas of NLG, as 
demonstrated in the papers by Radev and McKeown and by Chu-Carroll and Carberry 
351 
Computational Linguistics Volume 24, Number 3 
in this issue. Given the emphasis within NLG research on text-level issues, a major 
bottleneck for work here is the encoding of corpora with semantic and discourse 
structural features; see Di Eugenio, Moore, and Paolucci (1997). These are needed to 
uncover plausible text-structuring and microplanning strategies, but annotating cor- 
pora for such features will remain a laborious manual task at least for the foreseeable 
future. This effort may be alleviated if sharable corpora become available through 
the Discourse Resource Initiative (http://www.georgetown.edu/luperfoy/Discourse- 
Treebank / dri-home.html). 
With so many rich seams to mine, natural language generation has a promising 
future. We mentioned at the outset that researchers in NLG face the unique problem of 
deciding what to generate from: Yorick Wilks is credited with pointing out that, while 
the problem of natural language understanding is somewhat like counting from one 
to infinity, researchers in natural language generation face the problem of counting 
from infinity to one. In order to make progress, researchers in NLG pick a reasonably 
high number and get to work; as researchers in NLU climb the numerical ladder from 
the other end, we can expect that some of the big numbers discovered in NLG will 
prove to be of use in NLU too. 
Acknowledgments 
We offer our grateful thanks to the body of 
reviewers who did so much work in 
helping us put this issue together. We also 
acknowledge the many fruitful discussions 
we have had with our colleagues, including 
especially Giuseppe Carenini for sharing his 
notes on evaluation in NLG and Ehud 
Reiter for his observations on the state of 
the field. 
References 
Adorni, G. and M. Zock, editors. 1996. 
Trends in Natural Language Generation. 
Lecture Notes in Artificial Intelligence. 
Springer-Verlag, Berlin. 
Appelt, Douglas E. 1981. Planning Natural 
Language Utterances to Satisfy Multiple 
Goals. Ph.D. thesis, Stanford University, 
Stanford, CA. Available as SRI Technical 
Note 259. 
Appelt, Douglas E. 1985. Planning English 
Sentences. Cambridge University Press, 
Cambridge. 
Bateman, John A. 1997. Enabling technology 
for multilingual natural language 
generation: The KPML development 
environment. Natural Language 
Engineering, 3:15-55. 
Dale, Robert, Eduard H. Hov3~ Dietmar 
R6sner, and Oliviero Stock, editors. 1992. 
Aspects of Automated Natural Language 
Generation. Lecture Notes in Artificial 
Intelligence. Springer-Verlag, Berlin. 
Dale, Robert and Christopher S. Mellish. 
1998. Towards evaluation in natural 
language generation. In Proceedings of the 
First International Conference on Language 
Resources and Evaluation, Granada, Spain, 
May 28-30. 
Dale, Robert, Chris Mellish, and Michael 
Zock, editors. 1990. Current Research in 
Natural Language Generation. Academic 
Press, New York. 
Dale, Robert and Ehud Reiter. 1995. 
Computational interpretations of the 
Gricean maxims in the generation of 
referring expressions. Cognitive Science, 
18:233-263. 
Dave~ Anthony C. 1979. Discourse 
Production. Edinburgh University Press, 
Edinburgh. 
Di Eugenio, Barbara, Johanna D. Moore, 
and Massimo Paolucci. 1997. Learning 
features that predict cue usage. In 
Proceedings of the 35th Annual Meeting, 
pages 80-87, Madrid, Spain. Association 
for Computational Linguistics. 
Elhadad, Michael. 1993a. FUF: The 
universal unifier--user manual version 
5.2. Technical Report CUCS-038-91, 
Columbia University. 
Elhadad, Michael. 1993b. Using 
Argumentation to Control Lexical Choice: A 
Unification-based Implementation. Ph.D. 
thesis, Computer Science Department, 
Columbia University. 
Elhadad, Michael and Jacques Robin. 1996. 
An overview of SURGE: A reusable 
comprehensive syntactic realisation 
component. In Proceedings of the 8th 
International Workshop on Natural Language 
Generation (Demos and Posters), 
Herstmonceux, Sussex, UK, June. 
Goldman, Neil M. 1974. Computer Generation 
of Natural Language from a Deep Conceptual 
Base. Ph.D. thesis, Stanford University. 
Dale, Di Eugenio, and Scott Introduction 
Available as Stanford AI Laboratory 
Memo AIM-247 or CS Technical Report 
CS-74-461. 
Grice, H. P. 1975. Logic and Conversation. 
In P. Cole and J. L. Morgan, editors, 
Syntax and Semantics 3: Speech Acts. 
Academic Press. 
Horacek, H. and M. Zock, editors. 1993. 
New Concepts in Natural Language 
Generation. Pinter, London. 
Kempen, Gerard, editor. 1987. Natural 
Language Generation: New Results in 
Artificial Intelligence, Psychology and 
Linguistics. NATO ASI Series No. 135. 
Martinus Nijhoff Publishers, Boston, 
Dordrecht. 
Lester, James C. and Bruce W. Porter. 1997. 
Developing and empirically evaluating 
robust explanation generators: The 
KNIGHT experiments. Computational 
Linguistics, 23(1):65-102. 
McDonald, David D. 1980. Natural Language 
Production as a Process of Decision Making 
under Constraint. Ph.D. thesis, MIT, 
Cambridge, MA. 
McDonald, David D. 1987. Natural 
language generation. In Stuart C. Shapiro, 
editor, Encyclopedia of Artificial Intelligence. 
John Wiley and Sons, pages 642-655. 
McDonald, David D. and Leonard Bolc. 
1988. Natural Language Generation Systems. 
Springer-Verlag, New York NY. 
McKeown, Kathleen R. 1982. Generating 
Natural Language Text in Response to 
Questions About Database Structure. Ph.D. 
thesis, University of Pennsylvania, 
Philadelphia, PA, May. Available as 
Technical Report MS-CIS-82-05. 
McKeown, Kathleen R. 1985. Text Generation: 
Using Discourse Strategies and Focus 
Constraints to Generate Natural Language 
Text. Cambridge University Press, 
Cambridge. 
Meteer, Marie. 1990. The Generation Gap: The 
Problem of Expressibility in Text Planning. 
Ph.D. thesis, University of Massachusetts. 
Milosavljevic, Maria and Robert Dale. 1996. 
Strategies for comparison in encyclopedia 
descriptions. In Proceedings of the Eighth 
International Natural Language Generation 
Workshop, pages 161-170, Herstmonceux, 
Sussex, UK, June. 
Paris, C4cile, Keith Vander Linden, Markus 
Fischer, Anthony Hartley, Lyn Pemberton, 
Richard Power, and Donia Scott. 1995. A 
support tool for writing multilingual 
instructions. In Proceedings of the Fourteenth 
International Joint Conference on Artificial 
Intelligence, pages 1398-1404. 
Paris, C6cile L., William R. Swartout, and 
William C. Mann, editors. 1991. Natural 
Language Generation in Artificial Intelligence 
and Computational Linguistics. Kluwer 
Academic Publishers, Boston. 
Penman group. 1989. Documentation of the 
Penman Sentence Generation System. 
USC Information Sciences Institute, 
Marina del Rey, CA. 
Reiter, Ehud and Robert Dale. 1997. 
Building applied natural language 
generation systems. Natural Language 
Engineering, 3:57-87. 
Robin, Jacques and Kathleen McKeown. 
1996. Empirically designing and 
evaluating a new revision-based model 
for summary generation. Artificial 
Intelligence, 85:135--179. 
Roth, Steven F., John Kolojejchick, Joe 
Mattis, and Jade Goldstein. 1994. 
Interactive graphic design using 
automatic presentation knowledge. In 
Proceedings of CHI'94: Human Factors in 
Computing Systems, pages 193-200. 
Scott, Donia and Richard Power, editors. 
1994. Characteristics of administrative 
forms in English, German and Italian. 
Deliverable EV-1, GIST project LRE 
062-09. http: / / ecate.itc.it:1024 / projects / 
gist / gist-bibliography.html. 
Shieber, Stuart, Gertjan van Noord, 
Fernando Pereira, and Robert Moore. 
1990. Semantic head-driven generation. 
Computational Linguistics, 16(1):30-42. 
Stevenson, Rosemary Rosalind Crawley, 
and David Kleinman. 1994. Thematic 
roles, focus and the representation of 
events. Language and Cognitive Processes, 
9:519-548. 
Stevenson, Rosemary and Agnieszka 
Urbanowicz. 1995. Structural focusing, 
thematic role focusing and the 
comprehension of pronouns. In 
Proceedings of the Seventeenth Annual 
Conference of the Cognive Science Society, 
pages 328-332. 
Thompson, Henry S. 1977. Strategy and 
tactics: A model for language production. 
In W. A. Beach, S. E. Fox, and 
S. Philosoph, editors, Papers from the 13th 
Regional Meeting of the Chicago Linguistics 
Society, pages 651-668, Chicago, IL. 
Yeh, Ching-Long and Chris Mellish. 1997. 
An empirical study of the generation of 
anaphora in Chinese. Computational 
Linguistics, 23(1):169-190. 
Young, R. Michael. 1997. Generating 
Descriptions of Complex Activities. Ph.D. 
thesis, Intelligent Systems Program, 
University of Pittsburgh. 
Zock, Michael and G6rard Sabah, editors. 
1988. Advances in Natural Language 
Generation: An Interdisciplinary Perspective. 
Ablex Publishing Corp., Norwood, NJ. 

