Perception, Concepts and Language: 7~oA~9 and IPaGe 
Matthias Rehm and Karl Ulrich Goecke 
Faculty of Linguistics and Literature 
University of Bielef('ld 
{rehm, goecke} @coli.uifi-bielefeld.de 
Abstract 
A two-level natural  generation system tbr 
situation and action descriptions (SADs) of a sinm- 
lated assembly robot is presented, in the first step, 
multimodal inibrmation is used to obtain a concep- 
tual representation (CR) of an event. The second 
step is the lmrallel, incremental surface realizatiou of 
utterances from the robot's t)erspecl;ive bascxl on the 
CR. Theoretical issues addressed are semantics of 
SAI)s and distribution of lexical and syntactic pro- 
cessing, leading to a natural type of in(:rementality 
in NLG. 
1 Introduction 
Following (Reiter, 1.994), in the ma,jority of systems 
 generation starts with content determina- 
tion. This step is fllrther subdivided in deep con- 
tent dcl.cr'm, inatio,n where a decision takes i)lat-e what 
infornmtion shouhl 1)e (:ommmficated to the lmareJ' 
mM 7q,,eto'rical planning where this ilfl'ormation is or- 
galfized in a rhetorically coher(mt rammer. However, 
the definition of deet) content deternfination allows 
two difl'ering interpretations: One perspective is that 
there already exists a set of representations of dif- 
ferent contents in a specific representation tbrmat. 
The task is to select one of these representations to 
be commmficated. This process is tyl)icalty imple- 
mented using discourse structure information and 
other criteria (e.g. Grice's maximes, (Grice, 1975)). 
Tim or, her t)erspective is more extensive. If, ini- 
tially, no set; of rcpresentations is available, deep 
content deternfination could also comprise the cre- 
ation of this set. In many technical apt)lications, 
this viewpoint does not really make sense. In the 
most widesl)read case -  generation out of 
~ data base - the systein is the set of representa- 
tions. A possibly more complex case -  gen- 
eration by agents 1 - seems to be no problem either: 
Take tile world model and tile lflan representing the 
agent's environment and fl~ture actions, respectively 
and then select fi'om these representations the con- 
tent to be verbalized. Approaches of this kind are 
1For a discussion of this notion see e.g. (Franklin mid 
(\] raessc.r, t996). 
indeed al)plied in various scene description or robotic 
systems, e.g. (Now~k, 1987), (Lfingle et al., 1995). 
Itowever, there are agent architectures not enter- 
tahfing a coherent world inodel and not having an 
explicit plan at ('.very level of action available. Such 
agent architectures have emerged within behavior- 
oriented robotics over the past fifteen years (e.g. 
(Brooks, 1986)). Percet)tion and control 1)rocesses 
are distributed over many inodules and interacl; 012 
a local rather than a global basis. Thus, reactive 
l~ehavior is modeled. 
Considering the problem of  generation , 
it becomes more natural to adopt the second t)er- 
spective about what deep content determination 
should mean ill this context. In this paper, we 
present an aI)proach to natural hmguage generation 
where in a first step, a set of possible utterances 
of an agent is constructed 1) 3, the system "R.oAD 
l~)r the lack of a better term, we call l;his pro- 
eess concCl)tualizatio,n. The intermediate struclurc, 
is based on the Conceptual Semantics paradigm (e.g. 
(Jackendoff, 1.990)). In a second stel) , what is clas- 
sically thought of as  generation is accom- 
plished with the system IPaGe (Increnmntal Parallel 
Generator). Here we propose a nmssively parallel 
approach that leads to a natural kind of either quali- 
tative and quantitative incrementality (c.f. (Flakier, 
1997)). 
The l)at)er is structured as follows: In section 2, 
some theoretical claims are clarified and the imple- 
mentation of ~,oJl'D is illustrated with an example 
conceptualization. The corresponding CR is also 
used in section 3 where IPaGe is described. An out- 
look on directions for further research ill section 4 
concludes this paper. 
2 Conceptualization 
The example domain of our work is the artificial 
agent CoT~,el (c.f. (Milde et al., 1997)). It is a sin> 
ulated assembly robot, al)le to manipulate wooden 
toy parts and instructable by a human user. Due to 
its behavior-l)ased control architecture, not every in- 
structed action is carried out as expected. Thus, an 
explanation of ongoing actions by the robot is desir- 
1091 
able. We shall deal with a specific kind of utterances, 
namely situation/action descriptions (SADs). 
SADs are descriptions of the enviromnent and 
the actions of an agent from its own perspective. 
In the chosen domain, possible SADs are: Ich 
bewege mich zu dem blauen Wiirfel (I am mov- 
ing towards the bhm cube) or Ich lege den Wiirfel 
auf eine Leiste (I am placing a cube on a connec- 
tion bar). 
A theoretical and an application-oriented issue are 
addressed with the investigation of SADs: First, we 
claim that internal and sensoric parameters of the 
agent play a decisive role in deternfining the seman- 
tics of SADs, especially of action verbs. Second, by 
generating SADs, we enable an interactor to under- 
stand what the agent is doing. This is particularly 
interesting if there is no direct visual contact be- 
tween the agent and the interlocutor. 
Sensor data (visual, telemetric, haptic) and inter- 
nal states (activation of behavior generating modules 
(BM), joint values) of Co7¢A set've as selection crite- 
ria (SC) of interpretative schemata (ISM) contain- 
ing also conceptual units (Goeeke and Milde, 1998). 
These units constitute the interface to tile surface 
generation process. If the SC are present in the cur- 
rent sensoric pattern of the robot, an ISM may be- 
come active and the corresponding concept informa- 
tion is passed to surface generation (c.f. section 3). 
Some SC are temporally extended: A sensor value 
has to be in some interval for a certain amount of 
time to function as a trigger for an ISM. 
Additional to sensor data and internal values, ISM 
themselves can be part of the SC of other ISM. The 
latter ones are then said to have a higher degree 
of complexity then the former ones. ISM may be 
subsumed by other ISM or may bear a part/whole- 
relation to other ISM. An ISM is subsumed by an- 
other one if its SC constitute a proper subset of 
the SC of the subsuming ISM, including time re- 
strictions. For example, the ISM MOVE, detecting a 
movement of the robot that is not specified with re- 
spect to direction or velocity, is subsumed by MOVE- 
TO-OBJECT which, in addition, recognizes a goal of 
the movement, namely some object. Part/whole- 
relations of ISM exist if a "higher" ISM combines 
others to identify more complex actions. In this case, 
the "lower" ISM not necessarily has to be active over 
the whole time interval that the higher one covers. 
Thus, ISM form a hierarchy (see fig. 1). ISM on 
the lowest level (St.;E, MOVE, BUMI', ...) are basic in 
the sense that they only contain sensor data as SC. 
Complex ISM in levels 1 to 4 integrate sensor data 
as well as other ISM. 
When an ISM becomes active, the corresponding 
conceptual representation (CR) is a possible candi- 
date to be passed to the surfime generation compo- 
nent. As it is possible that several ISM are active 
.-d; ~5:'f~'~7"S~Z~{~it:?gfiJ:;'~5g'~}f<'Y'~T~2"-' i''::-¢{:' i';}~ ; "-' !"' " .::'" ~'~ i::: :'~"" ", 
..- . . 
Figure 1: Screen depiction of the ISM-hierarchy. Ac- 
tive ISM are coloured; edges mark subsumption or 
part/whole-relations. 
at the same time, a selection has to take place. At 
present, the only criterium relevant for selection is 
the position of the corresponding ISM within the hi- 
erarchy. Thus, only the CR of the highest active 
ISM is going to be verbalized at the given time. 
CRs contain information about objects, their 
properties attd about events tile agent is involved 
in. They follow the Conceptual Semantics by ,Jack- 
endoff (Jackendoff, 1990). In the next section, an 
example shows the representations used in goAD 
in more detail. 
2.1 Example 
In the following, the conceptualization of tile 
SAD Ich drehe mich zu dem blauen Wiirfel (I 
am turning to the bhm cube) is going to illustrate 
the processing mechanisms described in the previous 
section. 
Suppose a situation where visual and movement 
information is provided by the sensors of tile robot. 
Among other things, the ISM SEIC checks for the ex- 
istance of values in the interface for either object 
type (type), object color (color) or object position 
within the visual field (x-g und y-g). A possible 
configuration is 
type: c color: b 
x-g: 102 y-g: 99 
width: 83 height: 200 
The SC of SEE is a disjunction of several con- 
ditions. If any of these conditions is met, tile 
ISM becomes active: x-g > 0 \[ y-g > 0 \[ width 
1092 
> 0 \[ height > 0. The ('Oml)arison between the 
attribute-value pairs in the interface and the SC of 
Sl,',I," shows that solne relevant paranmters are indee(1 
present. Thus, SEE t)ecomes active. 
The understmcified CR of SEE is EVENT: see, 
AGENT: ±, OBJECT: 0BJ, (COLOR: COL). EVENT 
and AGENT arc instantiated with default vahms. Tile 
associated transition rules sl)ecit\[y the remaining 
conceptual parameters: 
type OBJ ---+ OBJ 
color COL --+ COL 
Consequently, the comt)lete CI1, of sI.:g is EVENT: 
see, AGENT: i, OBJECT: c, COLOR: b 2 
On the basis of the sensor data down: -8 and 
velocity: 0.441355, denoting a downward move- 
inent with a certain velocity, the ISM MOVF becomes 
active at the same time. 
The basic ISM SEE and MOVE serve as SC for the 
COlnplex ISM TUI{N-TO which identifies the turning 
of the robot towards an object;. Furthermore, TURN- 
q'O has some additional SC. The complete set ix as 
follows: 
MOVE ~ SEE 
((cont 100) x-g) (,:,,,it lOO) y-g)) fo,.s 
cycles 
Thus, if MOVE an(1 SEE are active tbr five cycles 
and, in addition, the object in the visual tMd is mov- 
ing to the center of vision, 'I'\[JI/.N-'I'O is activated. 
MOVE and SEE make available their CRs to TUI/N- 
TO. By this nmans, 1)reviously unst)e(:ificd t)arame- 
ters in 't'UII,N-TO's Cl~, c&n })e spell(~d out via the. 
transition rule 
see(OBa, COL) ~ OBJECT, C()L()1\]. 
resulting in the flflly instantiated CR EVENT: turn, 
AGENT: ±, PATH: to, OBJECT: c, COLOR: b. 
If TUI/N-TO turns out to be the highest active ISM 
at; a given time, it is selected for surface generation. 
3 Generating the utterance 
After conceptualizing the agent's current action at 
the level of ISMs (content determination) it has to 
be decided how the corresl)onding CR can be artic- 
ulated in a natural  utterance. A fimda- 
mental division in lexieal and syntactical processing 
along with an incremental and parallel processing 
behaviour are the crucial features of tile proposed 
architecture for surface realization. Such a process- 
ing behaviour ix facilitated by the use of CRs as in- 
terface between "R.oAD and iPaGe. 
2c stands for the concel~tual entity 'cube', b for 'l)hm'. 
" ~;oipho\]D~i,i, \[ "t ~ l 
SYNTACTICAL 
IPIIOCF.S SING 
inMnltliMioll 
Figure 2: The architecture of the generation system. 
Conceptual, lcxical, and syntactical information is 
stored centrally on the blackboard. All processes 
can obtain this information if needed. 
3.1 Blackboard architecture 
The blackl)oard is the central storing device of tile 
system. All processes write their results on tile 
blackboard and obtain their input structures there. 
In such an architecture parallel processing can be 
achieved in a convenient fashion. 
3.2 Parallel processing 
Parallel processing can be found on various levels of 
abstraction. The flmdmnental division runs 1)etween 
lexical and syntactical 1)rocessing (see fig. 2). When 
a (1)art of a) CR originating from tile ISM is written 
on the blackboard, processing of this structure starts 
sinmltaneously in both components. Using a concep- 
tual lexicon as described by Jackendoff a transfor- 
nlation process constitutes the first si:e l) on the lexi- 
eal side (Jackendoff, 1990). On the syntactical side, 
a 1)locess based on the type-phrase-corresl)ondence, 
also described by Jackendoff, starts processing in 
this component. The type-t)hrase-correst)ondence 
constrains the choice of possible phrases to realize 
a structure with a given conceptual type in an ut- 
terance. 
Different processes inside these two components 
work sinmltaneously on different CRs. On the lexieal 
side, the three subprocesses transformation, choos- 
ing of lexeme and morphological processing can be 
identified. On the syntactical side the real)ping of 
conceptual types on possible phrase structure rules 
and the instantiation of these rules take place. 
3.3 Incremental processing 
In natural  generation, incremental process- 
ing is a central issue. According to Finkler's defini- 
tion two fundamental types can be identified: quan- 
titative and qualitative incremental processing (\]?in- 
kier, 1997). 
Quantitative incremental processing nmans that 
the inlmt to a process can be divided into differ- 
1093 
\]EVENT turn (\[AGENT\] 52' \[PATH\] 
\]AGENT i152 
29 )\] 7 
\[PATH to (\[OB.I~CT\])\] 29 
I 
OB JECT cube-~ 
COiOR\] J 29 
\[COLOR blue\] 29 
Figure 3: A complex CR with its subparts which 
can serve as inlmt increments to the generation pro- 
tess. Parts of the same structure are coindexed. The 
relevallt features fbr generation fronl CRs are: tilt 
type, the head, and tile number and types of argu- 
lnents. Thus, \[PATH to(\[0BJECT\])\] is a possible 
increment. 
ent parts and that processing of these parts is possi- 
ble. Qualitative incremental processing ou the other 
hand denotes the possibility to ()btain a first, subop- 
timal result and to elaborate it wtmn new ilfforma- 
lion is available. 
IPaGe realizes both kinds of incremental process- 
ing. Parts of CRs correspond to possil)le phrase 
structures that constitute the utterance, i.e. type- 
ptlrase-corresl)ondence. Thus quantitative incre- 
inental processing can be achieved ill a natural way. 
An arbitrarily complex part of a CR can serve as an 
increment. 
Qualitative increnlental processing is accom- 
plished on tile level of instantiating phrase structure 
rules. A pllrase Call always be realized by several 
rules of differing colnplexity. D)r exalnple, a noun 
phrase call be realized ill one of tile following wws: 
as a noun, as a determiner and a noun, as an ad- 
jective and a noun, etc. All rules for a given CR 
m:e started as independent processes trying to ii1- 
stantiate ttmir right hand sides. The result of a suc- 
cessful process is written on the blackboard and all 
processes with equal or less complexity are stopped. 
Processes of higher complexity can try fllrther to in- 
stantiate their phrase strnctnre rules. Ill case Olm of 
these processes succeeds, the former result is over- 
written by the more colnplex one. Depending on 
the utterance tilne of tile given phrase a more or 
less complex result is achieved. 
3.4 Example 
The exmnple introduced in section 2.1 is continued 
here. Outlmt of the ISM and thus input to tile 
generation process is tlm CR dei)icted in figln'e 3. 
This structure will be realized in an utterance like 
Ich drehe reich zu dem Wiirfel (I anl turning to 
tlm cube)bzw. Zu dem blauen Quader drehe ich 
reich (To the blue cuboid I am turning). Tlm gener- 
ation process is exemplified by the processing of the 
PATH-structure. 
3.4.1 Transformation 
Every CR carries enough information to initiate dif- 
ferent processes sinlultaneously. At the moinent all 
input structure is supplied, it triggers processing in- 
side the lexical and tile syntactical comt)onent: tile 
transformation and the mapping process. 
As CRs describe meaning by a structural mech- 
alfism, the same head can have different ineanings 
ill different structural constellations. ~iS'anst'orina- 
lion - a disanlbiguation process - is imt)lemented as 
a lookup process in a conceptual lexicon. The en- 
tries in tiffs lexicon are sorted by different keys. The 
first key is the type of the CR. Int)ut increments 
to tile lexical processing component are typed CI/,s. 
Thus a type specific distribution of processing seelns 
natural here. 
By the PATH-stnmture tim PATH-specific trans- 
formation process is triggered. The lookup pro- 
cess will yield a so-called intermediate structure 
already with some syntactic intbrmation such as 
category infornlation: \[PRED "co, CAT prep, ARG 
29\]29. \[OBJECT cube\] will yield: \[PRED block, 
CAT n\] 29. 
3.4.2 Mapping 
PATH- all(\] 0BJECT-structure initiate tlm mal)- 
t)ing process simultaneously to the transfiwma- 
lion i)rocess. During mal)t)ing the tyl)e-1)hrase- 
correspondence as described by (Jackendotl', 1990) 
is used. A given tyl)e can be expressed in an utter- 
ante only by a restricted set of possible phrases. A 
nml)t)ing of the types of the int)ut structures to the 
relevant 1)hrase structure rides takes place. Tllese 
rules are then started as independent threads. 
In Gerlnan, structures with type PATH are nearly 
always realized as prepositioiml pllrases. The PATH- 
specific mapl)ing process will tlms start PP-rlfles, 
e.g. PP29 -+ PREP NP or PP29 -+ PREP ADV. Tile 
0BJECT-structure will trigger the 0BJECT-sl)ecific 
mat)ping ttlrea(1 which will start NP-rules, e.g. NP2:) 
-+ N, NP29 -4 DET N, orNP29 ~ DET ADJ N. 
3.4.3 Instantiation 
All threads started during the mapping process con- 
stitute the instantiation module. These threads co> 
respond to phrase structure rules and try to substi- 
tute the right hand side non-ternfinals with inflected 
words. If one rule is fillfilled, the result will be writ- 
ten on the blackboard. All rules for the same con- 
stituent which m:e more colnplex will continue pro- 
cessing. If such a more complex rifle is flflfille(1, it; 
will overwrite the result on the blackboard by a inore 
1094 
eomt)lex, i.e. more elaborate one. Which one, will 
t)e uttered del)ends (m the time the outtmt t)rocess 
reaches the eorrest)onding 1)hrase and on the time 
eontraint given 1)y ml 'aEqcncy tlarameter. A low ur- 
gency corresponds to more 1)recessing time whereas 
a high urgency denotes the need for a fast result. 
Let's have a closer look on dm NP29-rules. If 
Wiirfel is SUl)t)lied 1)y tile h;xical t)rocessing corn- 
t/orient, all N-(:onsl;ituents will be sul/sdtuted 1)y 
this word: (i) NP29 -+ Wiirfel, (ii) NP29 -+ DET 
Wiirfel, (iii) NP29 -+ DET ADJ WiirfeL Rule (i) is 
complete, and will be written on the 1)ladd)oard. If 
the OUtlmt 1)roeess reaches the noun t)hrase at that 
moment of time Wiirfel (cube) will lie used as noun 
phrase. Otherwise the more complex threads ca.n 
try to flfltill their rules. Next, e.g., the del;erminer 
reaches the t)laekl)oard, lhfle (ii) will t)e (:Oml)lel;e 
mM will overwrite the, former result;: NP2:) -+ dem 
Wiirfel. Now, the ul;l;ere(1 i1(31_111 phrase will be dem 
Wiirfel (tile cut)el. 
3.4.4 Choosing Lexemes 
The infle(:l;ed words a.re 1)rovided t)y the lexical pro- 
(:essing (:onq)onent. After th(; disambiguat;ion of l:he 
inf)ul; stru(:i;m'es during the trmmfl)rmation f)rocess 
1)ossibh; lexenw, s :{ have to lm chosen that will con- 
st;il;ute the ul;tera, n(:e. This pl'()(:e,qs is ()ll(',(~ re(ire a 
h/okul) in a lexi(:on. The tirst key to this h/okul) is 
the synl;acdc ca l;egory which is given in the interm(> 
diate sl;r11(;l;l11"e~s. ThllS ~l cat(~gory-sl)e(:iti(: distribu- 
tion of 1)ro(:esses is realized here. 
h,'bmy coneel)tS can be paral)hras(!d 1)y a mun- 
l)er of words, l%)r (~'Xmnl)le , l;lm (:(m(:el)t of cube 
('~t.ll t)o, e, xl)ressed ill German t)y words like Wiirfel, 
Quader, Block, Klotz, Qua&rat, eL(:. ((:ulm, 
(:,lboid, 1)h)ek, 7?7, S(luare'l). One of these has I;() 1)e 
chosen. This is mainly done rmMomly but the pro- 
cess takes illl;O :,lC(:()llll|; l)r(?f(',rell(;e values l;hal; gll;ll- 
a,lll;ee l;lmt 111ore UllllSllal words wilt not be chosen as 
often as fl'equ(',nt words. 
\[PRED block, CAT n\]2:) will trigger the N- 
Slm(:ifi(: choosing process. It is 11tos\[; likely that the 
result of this 1)ro(:ess will 1)e \[STEM Wiirfel, CAT 
n\]2:). \[PRED zu, CAT prep, ARG 29\]2:) will yiehl 
\[STEM zu, CAT prep, SUBCAT np'):)\]2!). A noun 
t)hrase is sut)(:a.l;egorize.d, as tim tyt/e of the argu- 
ment is OBJECT. 
a.4.5 Morphological Processing 
The chosen lexemes have to 1)e inth',(:l;ed. Agmn, this 
morl)hological processing c(ntsists at the moment of 
a k)okuI) l)ro(:ess. All 1)ossible woM forlns are listed 
along with the needed agreenm, n(; infornmdon. 
3L(}X(!III(~S }~l'(} llll(|(!rs{iO()d \]l(}l'(! ;Is the {~(}ll(!\]';ll |'Ol'lll ot" D. 
word, i.e. llO|, inflected. This 118~145(} is not to 1)e confused with 
the notions of lemma and lea:tree as introduced by Imvelt 
(1 ,evelt, 1989). 
4Nol; corro, ct \])Ill; found ill ~tCtllal l~-ttlgtHl~ge data. 
Thus, \[STEM Wiirfel, CAT n\]29 is changed lille 
\[SURF Wiirfel, CAT n\]29 mM \[STEM zu, CAT 
prep\]~9 into \[SURF zu, CAT prep\]20 which are 
used during the instmltiadon process. 
4 Conclusion 
The systems deserit)ed have been imi)lemented ill 
.\]av~l mM |;esl;ed in tile at)ove-nmntioned context. 
Currently, the coneeptuMization is a pure t)ol;l;om- 
up lneclmnism. No deliberative information like a 
llartner model or intentions is taken into aecollll{;. A 
discourse model could improve the selection mecha- 
llisln for CI{s significmMy. In tPAGE, it coukl also 
be used for t;he treatment of ellipses or anal)hera. 

References 

11. A. Brooks. 1986. A rotmst layered control system 
for a mobil(,' robot. IEEE Journal of R, obotics and 
Automation, 2:14-23. 

W. Finkler. 1997. Automatische Selbstkorrektur 
bei d('r inkrcmcntcllen G('nerier'ung gcsprochcner 
Sprache untcr Rcalzeitbcdingungen. lnfix. 

S. lh'anklin and A. Gr~msser. 1996. Is It; an Agent, or 
,Jllsl; ~l Progranf?: A Ta,xOllOltly for Autonomous 
Agents. In J. \]). Miiller, M. 3. Wt/olridge, and 
N. R..\](innings, editors, Intelligent Agcnl, s IlL 
Agent Th, cories, Arch, ile(:l, ures, and Langv, age,s., 
t)ages 21-35, Berlin. S1)ringer. 

K. U. Goecke and J.-T. Milde. 1998. Situations- 
und Aktionsl)eschreil/ungen dureh einen teilau- 
lOll()llleU Montagerot)oter. Iu Comput('.rs, Lin- 
g'u, istic,% and l)h, onetics b('.t, wc(m, Lan, g'tt, age and 
,gp('.cch,., vohmm 1, pages 331 335, Frankfmt a. M. 
\]~eter l ,ang. 

II. P. Grice. 1975. l,ogi(" mid conversation. In 
P. Cole and .I.L. Morgan, editors, Syntaz and Se- 
ma'n,t, ics, volume 3, pages 41- 58. A(:a(lelni(: l>re.qs, 
New York. 

11. Jackendotf. 1990. Semantic Str'u(:tu~vs. MIT 
l)ress, Cambridge, MA. 

T. bSngle, T.C. Liil, h, E. Sl;opl) , G. I\]erzog, and 
G. Kamstrut). 1995. KANTRA A Natural Lan- 
gua.ge 17nterfa(:e for Intelligent Robots. lk'richt 
114, UniversitSt des Sambmdes, 3. SFB 314. 

W. J. M. Levelt. 1989. 6'peaking: Frvm Intention to 
Articulation. MIT Press, Cambridge, MA. 

.J.-T. Mihle, S. Strit)l)gen, ;rod K. Peters. 1997. 
Situated communication with robots. Ill 1st Int. 
Workshop on Ituman-Computcr ConvcTwation. 

n.-z. Nowlk. 11987. Tcztgencricrung aus vis'aclh:n 
Datcn : 13csch, reibungcn yon Stra,~'c'nszcncn. 
Springer, lh~'rlin. 

E. Reiter. 1994. Has a Consensus NL Generation 
Architecture Allpea.red ,mM is it Psycholinguisti- 
tally l'lausible? Ill Prec. of th, c 7th Int. Workshop 
on Natural Language General, ion, t/ages 163 170. 
