Anticipating the Reader's Problems 
and the Automatic Generation of Paraphrases 
Nils Lenke 
Gerhard-Mercator-Universitact-Gtt Duisburg FB3 - Computerlinguistik 
Lotharstr. 65, D-47048 Duisburg 
voice: +49 (0)203-379-2007; c-mail: he2331e@unidui.uni-duisburg.de 
0. ABSTRACT 
The notion of paraphrase is discussed ,'rod compared with 
the similar notion of periphrase, qhe role of paraphrases 
in oral communication is described, and the results of a 
study on the role of paraphrases in texls are given. Fi- 
nally, a system which models the use of p~aplu-ases in 
texts is described. 
1. PARAPtlRASES IN I)IALOGUES 
If you look at ordinary dialogues you will find that com- 
munication failures - i.e. different types of misunderstan- 
dings - happen frequently, cf \[Ringle & Bruce 1982\]. One 
important techaique for the participants of the communi- 
cation to solve these problems is paraphrasing, that is, 
saying it again in other words. Parapla-ases can be offered 
by the hearer ("Is it this what you want to say: ...") or re- 
quested from the speaker by five hearer ("Iluh? 1 don't un- 
derstand."). These kind of paraphrases may be called 
cotmnunicative or pragmatic paraphr,~tses. 
2. OTHER NOTIONS OF "PARAPIIRASE" 
Notions of "paraphrase" exist which differ fiom the one 
presented above. In linguistics, especially Trimslbnnatio- 
nat Grammar, cf. e.g. (Smaby 1971), (Nolml 1970), the 
paraphrase relation is induced by the rules of the limguage 
system. Two formulations count as paraphrases of each 
other if they can be derived from a common deep 
structure, e.g. the active and the passive version of a 
sentence. So, the paraphrase relation is completely 
independent of the situation and communication 
participants. This view has been heavily criticiscd, cf. 
(Ungeheuer 1969). 
In CI,, the generation of a surface form from a 
meaning representation is sometimes called paraphrase 
generation, especially if different surface forms can Ix: 
generated for the same meaning representation. An 
,'unbiguity exists here, because the paraphrase relation can 
be meant to hold (a) between the mcaning representation 
and the NL text derived from it or (b) between two 
alternative formulations which could both be derived from 
the meaning representation. 
1 will shnply call case a) "generation" because that is 
what it means: deriving a text from an underlying mea- 
ning representalion. Case b), exemplified by (Gokhn~m 
1975) and most wolk in the area of Meaning-Text Mo- 
dels, cf. e.g. (lordanskaja, Kittredge & Polgut3re 1991), 
(Mercuk 1981), stresses the possibility of an alternative 
formulation which could be uttered instead of another 
formulation, whereas in seclion 1 we talked of paraphra- 
sing as uttering a formulation in a&lition to imother for- 
mulation. To differentiate between these cases I will not 
call case b) tutraphrase but - in accord:race with classical 
rhetoric - periphrase. 
3. RELATED WORK IN CL 
Quite a lot of work exists on the use of paraphrases in 
connection with Nl.-database frontentks, cf. e.g. 
(McKeown 1979), (Meteer & Shaked 1988). "ll~e formal 
representation g~fined from the user's query is translated 
back to NI, ag~dn and the user is requested to indicate if 
the system understood him correctly. This fits nicely into 
the framework from section 1. 
As indicated atxwe, much of the work presenlcd under 
the title "p,-uaphmse generation" should better be called 
"periphrase generation". Reiter's (1990) system IN  
generates - depending on the user model entry for the 
problematic word "shark"- one of the following alternative 
fonnulations: 
I a) 'll~crc is a shark in the water 
lb) There is a dangerous fish in the water 
Similarly, the system WISBER (Iloracek 1990) generatcs 
one of the following formulations, where the problematic 
word is "Notgroschctf' (nfiuy day fund): 
2a) 1 \[aben Sic cinch Notgroschen? \[Do you have a rainy 
&ly lhnd't\] 
2b) 1 laben Sic ein Sparbuch mit zwci Ncttomonatsein- 
kommcn? \[l)o you have a savings account with two 
montlfs net income?\] 
In the terminology advocated here, the b)-c,-tses ,are peri- 
phrases of the a)-cases. Real tormulations with p~u'a- 
phrases would look so,nethiug like this: 
lc) There is a shark,that is, a dangerous fish, in the water 
2c) l tabcn Sic einen Nolgroschcn, d.h. ein Sparbuch nfit 
zwei Ncltomonatseinkommcn? 
\[1)o you have a rainy day fund, that is, a a savings 
account with two mouth's net income? 
It will be discussed below under which circumstances such 
uttcnmccs could be superior to the ~0- or b)-cases. 
4. ANTICIPATION OF MISUNDER- 
STANDINGS AND TIIEIR AVOII)ANCE 
Turning now to the geaeratiou of written texts it seems to 
bca bit paradox to do this in connection with paraphrases, 
since in scction 1 we showed them to be a phcnomcmm 
of dialogue, i.e. oral communication. But parapla-ases do 
play a role in texts as well, especi,'dly when anticipation 
is considered. This elm ,'already be noted in the cage of 
si×)kcn l~mguage. A well known model of the preduction 
of spoken language is the one of I,evelt (1989). One of its 
mzfiu aspects is the existence of control and revision loops 
which can be used to monitor the planned or re,'dizcd 
utterm~ce and detect errors in it. So, part of the errors c,'m 
319 
already be anticipated in advance by the speaker before the 
hearer even gets to hear the problematic utterm~ce. 
When we now turn to written language again, we also 
find the concept of problem anticipation and revision 
loops. These are of even greater importance here because 
the reader normally has no chance of signalling his pro- 
blems with a text to the author. So, the author has to take 
the role of the reader and anticipate problems he might 
have with the text. Most models of the writing process 
thus include a revision loop, cf. the well-known model of 
Hayes and Flower (1980). In CL, this mechanism is 
known under the name anticipation-feedback loop, cf. 
(Jameson & Wahlster 1982), and in the form of revision- 
based generation systems, cf. (Gabriel 1988), (Vaughan & 
McDonald 1986). 
What are the options for an author if he detects trouble 
sources in his planned text? lte may choose to 
a) add a meta-comment; the addition of meta-conunents 
(Sigurd 87) like "loosely speaking", "to say it 
frankly", "a kind of", etc. is often used to indicate to 
the reader how to interpret a problematic utterance. 
b) add a futaher, alternative formulation (a paraphrase) or 
c) replan the text (formulate a periphrase). 
The rest of the paper will solely deal with b) mid c). What 
was said so far leads to the following hypothesis: 
Writers of texts anticipate reader problems, anti, in some 
cases, include paraphrases to avoid these troubles. 
5. A STUDY ON PARAPIIRASES IN TEXTS 
A study, cf. (Lenke, in preparation) for details, was con- 
ducted in order to find occurrences of paraphrases in texts 
and analyse them with the aim of checking the hypothesis 
mentioned at the end of section 4. 
First, a small corpus of German texts was scanned 
manually for paraphrases; the major results were: 
• Paraphrases of the kind described above can indeed be 
found. Typical examples of such paraphrases are1: 
(3) "... introduces the notion of multiple inheritance 
- that is, the ability of a class to have more than 0he 
direct base class - and presents ..." \[p. 182\] 
(4) "A language is said to support a style of progr,'unming 
if it provides facilities that make it convenient 
(reasonably easy. safe. and efficient) to use that style." 
\[p. 14\] 
• only part (roughly 50%) of the paraphrases ,'u'e 
announced by indicators like "that is", "in other words", 
parentheses or hyphenation. The other paraphrases are 
simply added as an apposition to the pm'aphrased term. 
• the total number of paraphrases differs vastly between 
text types: in narrative texts few and mostly un- 
announced paraphrases occur; in more technical texts, 
1 the following English examples all stem from 
\[Stroustmp 1991\] and were collected just to be English 
examples suitable for the presentation in this paper. 
especially manuals and introductory texts, many para- 
phrases. 
In the second phase of the study, the IJMAS corpus of 
German (1 million running words from 500 texts of diffe- 
rent types) was then scanned automatically for the most 
conunon German paraphrase indicators (a.o. "d.h.", "das 
heiBt", in anderen Worten", "also") Well above 1000 
occurrences of paraphrases were found and analysed. "\[he 
results of the first phase could be confirmed. Other results 
wel~: 
• the syntactic form of the paraphrases is in most cases 
either a complete sentence (in which another complete 
sentence is paraphrased) or an apposition, which be- 
longs to the same syntactic category ,as the word/phrase 
it belongs to. 
• Paraphrases are directed to quite different problem 
sources which were anticipated by the author. Among 
the different types found were the following: 
1. problematic lexical items 
a) unknown words (cf. examples 3 above) 
b) ambiguous words; 
c) words of abstract nature which obtain their concrete 
meaning through the context in which they occur. The 
paraphr,'tses indicate the direction in which this concrete 
meaning should be sem'ched. Cf. example 4 above. 
2. reference problems 
a) ,ambiguous anaphoric references, e.g. pronouns; 
b) anaphoric expressions where the referent is very distant 
(causing memory problems) 
c) missing knowlcdge to understand referring expression. 
3. problems induced by rhetoric figures (metaphors, 
metonymy). 
4. inference problems 
a) problems of ,'k,;pcctualization. (only some aspects of the 
meaning of a word are relevant in a certain contex0. 
b) preblems of logical inferences. (Obvious and relevant 
inferences from an utterance might be too difficult to 
dnaw by tile intended reader). 
Thus, one can conclude that paraphrases ,are indeed used by 
authors to avoid auticipated reader problems. These 
problems can be of all those types that have since long 
been noticed in file ,area of NL understanding. 
6. IMI'LEMENTED MODEL 
The next step in the project was to design and implement 
a model which describes this use of paraphrases in texts. 
It should answer tile following questions: 
• I low can problems of the reader be anticipated? 
• Under which circumstances are paraphr,xses file adequate 
answer to this problems (and not, say, periphrases or 
meta.-co~mnents)? 
• I low can p,araphrases be genenlted? 
Three well known approaches to NL generation are com- 
bined in the model : user modelling, anticipation-feedback 
320 
loops ,'rod revision-b,'t~d text generation, its architecture is 
shown in Fig h 
Fig 1: tile system's architecture 
qlae main feature is the revision/mlticipalion-feedback 
loop, which is highlighted in tile figure. 
The types of problems for which paral)hntses c,'m be 
generated by file system are restricted to problems which 
occur during lexicalization and involve only conceptual 
knowledge (no assertiomtl knowledge) in order to restrict 
complexity. These are (in terms of seclion 5) tile types 
la, lb, 3a, and - with restriclions - 4a, which are (together 
with type lc) by flu" the most flequent types occurring in 
natural texts. "llae other types could princip.'dly be dealt 
with in a similar lhshion. A corpus of about 25 cxmuples, 
,all collected from the s,'une somce, the manual fl)r the 
Apple Macintosh operating system 7, were used as a 
basis. The advantages of this approach is that .'ill 
exmnples are based on a common domain (knowledge 
about Macintosh computers), so that a common lexicon 
and a common knowledge base can be used tot all of 
them. Of course, the techniques and principles used are 
not restricted to this ~t of ex~nnples ,'m(l could be 
transfeffed to other dom~fins. 
6.1 An example 
To demonstrate how the components of the systems work 
together consider example 7, from the co~pus on which 
the system is based: 
(5) Alle Macintosh Modelle sind mit einem Steekplatz 
oder &rt~f/ir Gerlite ausgestattct, tier die SCSI- 
Schnittstelle (Small Computer System Interlace) 
unterst0tzt. \[all Macintosh models fire cqtfil~pcd with 
a slot or ~ for SCSl-devices\] 
qtLe content pl,'mner of the system is only implemented as 
an oracle, that is, it is preset to produce tile concepts to be 
formnlated ,-rod to answer certain questions by tile form 
pl,'mner as if it were a full-fledged content planner in a 
complete NL system. In the concrete ex,'unple, it would 
first inform the other components that the linguistic con- 
text of the target item consists of the concept Macintosh 
(the only concept that precedes slot in the lthmned 
sentence) ,and would then request the form phmner to 
verb,'dize tile concept slot. 
The form planner would then look up tile first 
possible linguistic items for tim concept slot in the 
lexicon. The lexicon not only iucoq~rates inlommtion 
about the linguistic items but also about their 
connections to items of tile concept-base, qlmse 
connections take the form of ZOOM-schemata, as known 
from the WISBER system, cf. (lloracek 1~)0). Briefly, 
Z(X)Ms are links between concepts or slmdl sub- 
slrticltlres Of tile concept-netwo,k on the one h,-md and 
linguistic items (words) on the other hand. 
In our example, the first choice to verbalize slot would be 
'AnschluB'. This proposal is then put torward to the re- 
vision component which tries to anticipate reader trouhle. 
To do this, it uses a simple user model, which employs 
the well known stereotype approach (Wahlster & Kobsa 
1989). All concepts, lcxical entries and ZOOMs belong to 
one of the three categories common vocabulary, computer 
jargon and Macintosh specific jargon, qlle static part of 
the user lm~lcl then simply consists of three variables 
which indicate if the intended reader is expected to be 
fiuniliar with tile respective jargon. 
This user models dilfem from other approaches 
because it allows tile special wdue "?" which indicates 
incomplete (you never know ~dl about the readers) or 
inconsistent (a text can be meant simultaneously for 
novices and experts) knowledge. From this static part of 
the u~r model a delault value can be c~dculated which c,'m 
be ovcLTidden through learning (see below). "lkl be a bit 
more exact, two wdues are calculated in a kind of "worst- 
cw~e-amdysis" due to the "?" values in tile user model. 
In our example two Z(X)M-schemata exist for slot: 
slot <-> 'Anschlug' 
slot <-> 'Steckphttz' 
'Anschhff~' (and tile ZOOM connecting it with slot) is 
marked Macintosh, the alternative lexical entry 'Steck- 
platz' is marked comttlon. So, if the user model indicated 
that Macintosh w~cabulary was yes, the revision compo- 
nent would judge tile wording 'Anschlug' ok and the 
realization COmlmnent would output 
"Alle Nacintosh Modelle sind mit EINEM 
ANSCIILUSS for Ger~te ausgestattet, der 
die SCSI-.Schnittstelle unterstQtzt." 
B ut now consider a user model which indicates timt the 
knowledge of computer aml Macintosh jargon is known 
to be no. Of course, the revision component would indi- 
cate that the term 'Anschlug' cmmot be used. A possible 
solution would be to generate a periphrase, i.e. replacing 
'Anschluss' by 'Stcckltialz' which would be the next 
choice of the fonu planner. This would then be accepted 
by the revision component. In some cases, however, tiffs 
wouhl be less than perfect: (a) if the concept has repea- 
tedly to be verbalized in tile course of the text, (b) if them 
are stylistic reasons to use the first choice term (here: 
'Anschluss'), (c) if there are pedagogical reasons to use the 
first choice. 
(a) consider a case in which the periphmse is a longish de- 
Iinilion. It would be a Ixlre to replace a short term by 
this dcllnition 15 times around the text. So you do it 
once and simltiy use the now learned term in the rest 
of tim text. 
(b) Cerlain texls can loose their "feel" if slripped of e.g. 
tile expert wycabulary of a ce~lain area. 
(c) Manutds and inmxluctory texts are oftcn mc,'mt to 
teach tile vocabulary in addition to the concepts. In 
this ease it would be nonsense to replace the to-be- 
taught vocabulary by e~tsier "terms". 
321 
All these conditions can only be determined by the 
content planner (demonstrating the need for an interaction 
between form planner and content planner); in the system, 
the form planner asks the content planner, which works 
as an oracle, i.e. gives the correct answers (by forwarding 
the questions to the human operator). If one of the 
conditions holds, it would be unwise to formulate a 
periphrase. The next choice of the form planner would 
then be to ask the content planner to completely replan 
this part of the text, namely to include a new sentence 
def'ming the problematic term. The system output looks 
like this: 
ANSCHLUSS BEDEUTET STECKPLATZ. Alle 
Macintosh Modelle sind mit EINEM AN- 
SCHLUSS f~r Ger&te ausgestattet, der 
die SCSI-Schnittstelle unterst~tzt. 
Even this solution doesn't work in some cases and that is 
where paraphrases come into play. If stylistic variation is 
necessary or if the problematic term is embedded in the 
definition of still another term it is the right place to use 
a paraphrase: 
Alle Macintosh Modelle sind mit EINEM 
ANSCHLUSS D.H. EINEM STECKPLATZ f~r 
Ger&te ausgestattet, der die SCSI- 
Schnittstelle unterst~tzt. 
6.2 A second example 
Just another path may lead to the generation of para- 
phrases for an unknown term, ,as the next example will 
show: 
(6) "Mit der MR~- dem Ger~it zum Zeigen und Klicken - 
werden die meisten Macintosh Funktionen aktivicrt." 
\[With the mouse - the device for pointing and clicking 
- most Macintosh fimctions am activated\] 
Fig. 2 shows the part of the conceptual network under- 
lying this example: 
Fig. 2: the concept mouse 
t. 
Instr for 
The term 'Maus' is classified computer jargon and may 
not be known to the user. The replacement of the term by 
a definition (no synonym is available) yields the danger of 
encouraging false conversational implicatures by the 
reader, cf. (Reiter 90). Consider a user model where 
computer and Mac jargon are indicated as "?". A worst 
case analysis by the revision component would show that 
the use of 'Mans' is inappropriate because some novices 
wouldn't know the term, but that the pcriphrase 'Gerfit 
zum Zeigen (und Klicken)' is inappropriate either, because 
some experts will know the tenn 'Mans' and conclude 
from its absence that some other pointing device, but not 
the mouse, was meant. 2 So, a paraphra~ would again be 
the best solution. The system thus generates: 
Mit DER MAUS D.H. DEM GERAET DES 
COMPUTERSYSTEMS ZUM ZEIGEN werden die 
meisten Macintosh Funktionen akti- 
viert. 
11ere the paraphrase is a definition of the form per genus 
proximum et differentia specifica which results from part 
of the systems' concept net shown in figure 2. The sy- 
stem is capable of generating two other forms of definiti- 
ons (paraphrases), definition by antonymy and by enume- 
ration. 
6.3 Detection and resolution of arnhiguity 
Up to now, only the problem type of unknown words has 
been discussed. Due to lack of space only one more pro- 
blem type which leads to the generation of paraphr,'tses 
can be discussed, n;unely the problem of ambiguous 
words. This problem type has since long been discussed 
in the area of NL understanding. Techniques for its solu- 
tion include the use of spreading-activation mechanisms 
workiug on conceptual networks, cf. (llirst 1987). This 
can now be used for the purpose of problem anticipation. 
We just try to disambiguate terms and interest ourself in 
the c,xses in which it fails: these are candidates for para- 
phrase generation. Cf. the following ex,'unple from the 
corpus: 
(7) "l)as aktive Fenstcr stcht im Vordergrund\[,\] also 
yor allen anderen ee~l'flleten Fenstcrn." 
\[llm active window stands iu the foreground, that is, 
in front of all other open windows\] 
1 lere, for beginners two readings of 'ira Vordergrund' ,-Ire 
possible: a literati (this is file correct reading) ,and a meta- 
phorical (in the sense of "important, to be regarded") 
which are equally propable. The revision component 
comes to this conclusiou by conducling a worst case ana- 
lysis using the concept net, an activation-spreading aigo- 
rithm and the user model. Only Ihose concepts and links 
that are known to at reader may forwmd energy, so in tile 
case of "?" values in the user model, both alternatives 
have to be tested (hence the term "worst case anaysis"). If 
comparable quantities of the activalion energy induced 
into the net by the liuguistic context find their way to 
both (or more) readings (concepts) of the ,'unbiguous 
terms it is concluded (and then indicated to tile lbrm 
planner by the revision compo,mnt) that the ambiguity 
might not be resolved by the reader. Then, a paraphrase 
could eventually (in a process similar to that described 
above) be generated, defining the correct reading. See 
(Lenke, in preparation) for derails of the spreading-acti- 
wttion mechanism used. 
2 cf. Reiter's (1990) "dangerous fish" vs. "shark" 
example. 
322 
6.4 Two more features of the system 
These can only be discussed brietly. See (I.enkc, in 
preparation) 1or details. 
• Paraphrases of the aspoctualization type (see above, 
section 5) can ,also be generated, llere, only one of the 
defining elements of a concept, either the superclass 
(genus plx)xhnum) or one of the roles (differentiae) is 
verbalized. At the moment, this kind of paraphrase is 
only generated when requested by the content planner; in 
the future, it will be necessary to model the anticipation 
of inference processes based on relcvmlcc by the reader 
to correctly predict the need for such paraphrases. An 
exmnple from the corpus, the mtderlying concept net 
and the equiwdeut produced by the system ,are shown 
below. 
(8) Durch das Klicken werden die Objekte akti- 
viert, d.h., sic werden nun sd~warz (¢~!¢~ 
anderen Farbe) dargestellt und somit hervor~ehobcn. 
\[Caused by the clicking the objects are activated, that 
is, printed in black (or another colour) and so high- 
lightedl 
h.~.. ~-k~--~_ ~______~ b~qtt% ) 
Fig. 3: the concept actiwtte 
Durch das Klicken werden die Objekte 
AKTIVIERT D.H. SCHWARZ DARGESTELLT . 
• if a p,-u'aphrase for ,an unknown term has been generated, 
it can be concluded tlmt the reader now knows this temt 
qlais is modelled by an active component of the user 
model which overrides the default values computed by 
the static component decribed above. So, only for the 
tkst (or first ,and second) appe~ucnce of a term a para- 
phrase is generated. "ll~cre:dter the term is simply used. 
This nicely mimics the obserwitions made in naturally 
occurring texts. 
7. IMPLEMENTATION DETAII,S 
The system is implemented in an object-oriented pro- 
gramming language and runs on Macintosh computers. It 
cont~dns a conceptual network similar to KL-ONE, con- 
sisting of approx. 130 concepts ,'rod 65 roles. Its lexicon 
consists of 70 ZOOM schemata and 50 lexieal entries. 
8. FUTURE WORK 
Some possibilites for future work have ~dready been indi- 
cated in the text, most notably the embedding of the pro- 
cedures descrilx-.d into a full-fledged NL-system. The 
approach described could al~ be transferred to oilier kinds 
of possible reader problems as enumerated in section 5. 
Since these are the problem areas of NL-undefstanding, 
algorithms exist which try to solve the understanding 
problems posed by these language features. These could 
be used to predict f~filure (as w,'ts demonstrated above for 
the actiw~tion-spreading mechanisms). 
9. RI~FERENCI~S 
Gabriel, R.P. (1988). Deliberate Writing. In: McDonald 
& Bole F.ds. Natural Language Generation Systems. 
pp. 1-46. 
Goldman, N.M. (1975) Conceptual Generation. Ill: R.C. 
Sclmnk 1.21. Conceptual Information Processing. 
Hayes, J. R. & L. S. Flower (1980). Identifying the 
Organization of Writing lhocesses, lu 1 ,W. Gregg & 
E.R. Steinberg Eds. Cognitive t'rocesses in Writing. 
pp. 3 - 30. 
tlirst, (;. (1987) Semantic interpretation and the 
Resolution of Ambiguity. 
lloracek, II. (1990). The Architecture of a Generation 
Component in a Complete NL Dialogue System. In 
1)ale, Mellish, Zoek Eds. Current Research in 
Natural Language Generation. pp. 193-227. 
Iordanskaja, L., R. Kittredge & AI. Polgut3re (lt~)l). 
l.exical Selection and l'araphm~ in a Meaning-Text 
(;eneration Model. In Paris, Swmlout, Mann Fxls. 
Natural Language Generation in AI and CL. 
Jameson, A.& W.WMdster (1982). User Modeling in 
Anaphora Generation: FAlipsis and l)efinite 
1)escription In: Proc. of ECAI 82. pp. 222-227, 
McKcown, K. R. (1979) "Paraphrasing Using Given and 
New Infommtion in a Question-Answer System", in 
Proc. of the 7th Confereuce of the ACL, La Jolla, 
1979, pp. 67 - 72. 
Lcuke, N. (in prep.). Paraphrascn - L6sungen f/Jr 
antizipie~ae I.,cseq~robleme Ix~i tier automatischcn 
Textgenerierung. l)issertation, Univ. of Duisburg. 
Levelt, W. (1989). Speaking. l:rom Intention to 
Articulation. 
Mel'cnk, 1. A.. Meaning-Text M~×lcls (1982): A Recent 
Trend iu Soviet Linguistics. Ann. Rev. 
Anthropology 10:27-62. 
Mcteer, M. & V. Shaked (1988). Slrategies h>r Effective 
l~araphrasing. Proc. of Coling '88. pp. 431 -436. 
Nohm, R. (1970). Foundations for an Adequate Criterion 
of l~araphntse. 
Rotter, E. (1990). Generating Descriptions that Exploit a 
User's I)omain Knowledge. Iu l)~de, Mellish, Zock 
Eds. Current Research in Natttral Language 
Generation. pp. 257 - 285 
Riuglc, M. II. & B. C. Bruce (1982). Conversation 
Failure In Lehuert; Ringle Eds. Strategies for Natural 
l_xmguage Processing. pp. 203- 221. 
Sigurd, B. (1987). Mctacommcnts in Text Generation. In 
G. Kempen 1.21. Natural lxmguage Generation. pl t. 
453-461. 
S maby, R. M. (1971). Paraphrase Grammars. 
Stroustrup, B. (1991) The C++ Progr,'unming Language. 
2d. 1'21. 
\[lngcheucr, G. (1969) i)araphmse und syntaklisehe 
Ticfcustruktur. Folia Ling. 3. pp. 178-227. 
Vauglmn, M. & I). Mcl)onald (1986). A Model of 
Revision in Natural Language Geuemtion. Proc. of 
the 241h Annual Meeting of the ACL. pp. 90-96. 
W:dflster, W. & A. Kobsa (1989) User Models in Dialog 
Systems. In A. Kobsa & W. WahlsterFxls. User 
Models in Dialog Systems. pp. 4 - 34. 
323 
