'" e'" " ' ' " 1)I'1 ' ' , I h. I cxl, pla, nnmg (,ott~pot~<.nl of l,ho. I.,IIX)(I Sysl;¢.tn 
J. Kreyfl nnd H.-J. Novak 
IBM Deutschla,nd (;mt)ll, Wissc.scha, fl, liches Zenl, rulll 
last, lent fiir Wissenslm..~derl.e Sysl,(,me 
Postfa.c\]l 80 08 80, 7000 St.uttga.rt. 80, Wesl.-(:lerma.ny 
Phone: (-t-49) 711-6695-447/540, e-mail: N()VAK a.t. I)S0blI, OCI.I31'I?NET 
Abstract 
In thi+ lmper we detcrlhe the conRtruction and impl,-me~v 
ration of PIT (Pre,enting Information by "l'extplamfing), tt 
~ubsy,tem of the LILOO. textunder.tandlng Jy~tem. Pl'F i~ 
uted for plmming rtnswertt of pnretgraph \]eltgth to qne~tlo,A of 
the kind What do lieu kttolJ~ about X q. We concentrated m~ t~ 
*imple. envy to implement mechanltm thtd can be further ex- 
tended. Experiencet with thit planning cottll~Onent , e~pecirtlly 
concerning the integrntion of new plrmt anti further exiellslonN 
are dilcllAle<l I 
1. Introduction 
As PIT is wholly integrated into the IAI,OG sys- 
tem, first some general remarks about LII,OG. 
In the LILOG project (Linguistic and logic meth- 
ods for the automatic understanding of German) we 
aim primarily at constructing a text understanding 
system. F'or the analysis part we use an llPSG-lmsed 
(Pollard and Sag 5'7) syntax and semantics that is 
further developed for German. For the representa- 
tion of world knowledge and the knowledge extracted 
from the texts we have devised the representa.tiou 
language Lt, rt, oa ( Plelat and yon. Luck 8g). Lr, tt, oa 
is an order sorted first-order predicate logic t.hat al- 
lows to define and farther describe sorts by using a 
KI,-ONE like sort descript.ion langnage, i.e. sorts can 
be described hy supersort and subsort relation~ as 
well as by roles (relations) and featnres (f,,,,ctions). 
The sorts themselves can be either primitive (at,m~s) 
or complex e.g. defined as miles, intersection, c,r 
complement of other sorts with constraints on role:i 
and features. The sorts form the conceptual entities 
of the system (they build au ontology) mid they are 
organized as a lattice. The sentantics of a word in 
the lexicon is given by a pointer into th in sort ln~.tice. 
In order to find out what the text-mlderstaudi,g 
system has really understood we can ask questions 
about the texts. In the first prot<)type (llollinger 
el al. 89) we could only ask. yes/no and constituent 
questions. In the present scenario the system in to 
understand and combine the information of several 
lyIoth author~ are indebted to F;,duard tlovy who devlted 
the planning ten,portent while ~taying a. a guest ~<:ienti~t in 
the I,ILOG project. The refinements, th~ i,.plementrttiotl, nn<l 
the experlencet reported here have been mri<le I+y the txtithorL 
The vlewI expret~e¢l in thi~ paper are our ~ole re.q.mtihillt.v. 
paragraph length texts about places of interest ill the 
city of l)iisseldorf and we also want to be able to ask 
questio,s of the kind What do ~ou know about the 
L,tmbrrtus cathedral? Questions of this type neces- 
sitate a. textplanniug component that decides first, 
which entities and second, in which order they should 
be verbalized such I.\]mtl a colwrellt descriptive para- 
graph is generated. 
There have heen several approaches to the gener- 
ation of co\]~ereut texts t.hat can be coarsely divided 
into two kinds: the schema based approach and the 
plan based approach. 
The first in desc,'ibed in detail in McKron, n 85. 
Schemata are representational structures for stereo- 
typical paragraphs that describe objects. A variant 
of this approach, somewhere between the schema and 
the plan based approach is described in Novak 86 
and Novak 87. }fete the structure of the whole text is 
l>a~ed on a schema whereas the sit, ale paragraphs are 
c,,~strueuted using domain restrictions and a tech- 
nique called anticipated visua.li~,ation. The aim i~ to 
describe the movement of an object such that the 
hearer can visualiT, e it aM H. has been seen by the 
system. 
The plan based approach has been put forth, 
mnoug others, by Merha.n 76, ~ohen 7,q, Avpclt 8,5, 
and \[f,,vTI 85. Mann. and 'l'h,,tTp.~on. 88 propose a 
,;et of al~ont 20 relations sufl~cient to represent the 
relntim~s that hohl within t,.×ts oceurri,g in F,n- 
glish. These relations, called R.ST (Rhetorical Struc- 
ture Theory), have been operat.ionali~ed and used an 
plans (lh,,v 5'8) in a top-down hierarchical expan- 
sion planner. The planner takes n.s input one or more 
COmlnllnicative goals along with a sel. of clause-.si~ed 
input;s to be generated as 1111 Pmglish pa:ragraph. It 
nssemldes the input entities into a t;ree that embodies 
the paragraph structure. Nonterminal nodes in the 
tree are RS'I! relations and terminal elements contain 
the iapnts, 
In our approach the same kiud of planner as de- 
scribed in llov~ 5'5' is used to find the entities in the 
knowledge hase that should be generated. 
In the following we first describe the overall a.rcbi- 
lecture oF the planner a.nd then its implementation. 
I 431 
2. Architecture 
Our textplanner basically decides what to say 
and gives as output a linear list of the conceptual 
entities that should be verbalized as answer to gen- 
eral questions of the kind What do you know about 
X? or What can you tell me about X? 
The planner takes as input a conummicative goal, 
e.g. describe(~), and needs access to all knowledge 
sources of the system, namely to the user model, 
the ontology, the background knowledge and the tex- 
tknowledge. As the knowledge of the system is rep- 
resented in LLILOG we use the inference engine for 
lookup and inferences. The user model currently 
only contains the facts that are already known to the 
hearer. The ontology is given by the sort hierarchy 
of the system, the background kmowledge contains 
world knowledge in the form of facts and inference 
rules and finally the textknowledge results from the 
analysis of seven short paragraphs describing places 
of interest in the city of Dfisseldorf. 
The output is a list of the entities and their at- 
tributes that should to be verbalized in. this order. 
This list is passed on to the generator that deter- 
urines sentence boundaries and decides on the syn- 
tactic realization of the entities. The result of the 
generator is a formal description of the output sen- 
tence. This description is then takeu by the formu- 
lator that constructs a correctly inflected Germtu~ 
sentence. The formulator is. a system similar to SU- 
TRA (Busemann 88) or MUMBLE-86 (McDonald 
and Mercer 88). 
2.1 Implementation 
In general, our implementation of the phmner is 
along the same lines a.s described in Hovy 88 except 
that we incorporate not only RST relations but also 
domaJnspecifie relations like ACCESS (how does one 
get to an object) and OPEN (when are the opening 
hours of an object). Moore and Swartout 89 and 
Moore and Paris 89 use the snme planning algorithm 
aud they have added plans like e.g. PERSUADE to 
the RST plans. This enables them to answer follow- 
up questions in advisory dialogues or in the explana- 
tion facility of an expert system. Most of the ques- 
tions they cml answer are Why questions except two 
What questions: What is a concept? and What is the 
difference between two concepts? The general idea 
of their approach too is to gather the information 
that should be communicated but using their plaits 
we could not answer the kind of general question we 
have in mind. 
Like RST plans our plans consist of a nucleus 
and a satellite each associated with requirements and 
growth points. The nuclei contain the information 
that has to be verbalized obligatorily which is either 
done by recursively invoking other subplans or by an 
explicit verbalization cmmnand say(aQ. All plans are 
recursively expanded until they lead to a verbaliza- 
tion command. In contrast to nuclei sateBites, using 
the same notation, contain the same kind ofinfornm- 
tion that can be optionally verbalized. The growth 
points allow for the inclusion of further infornmtion 
into the list of entities that is finally passed on to 
the generator. They again contain plans. Finally, 
the requirements for nucleus and satellite contain in- 
quiries to the inference engine about e.g. the validity 
of certain subsort relations and about beliefs of the 
hearer. An exmnple of a plau, inleresting_~eature , 
is given below (the planner is implemented in PRO- 
LOG so the atoms with capital letters are variables): 
plan(intoxosting.featuxe(0bject), 
nucleus: \[say(0bjeot)\], 
satellite: \[say(Featuxo)\], 
nuoleuszequizement: 
axtd(\[subsozt(Objoct,object)\]), 
satellitezequizement: \[\], 
nucleus and..satelliterequirement: 
and( \[ 
attxibute(0bjeet, zemarkability: 
not(bel(heazer, attribute(Object, 
remarkability: Feature)))\]), 
nucleus gzowthpoirtt: 
\[interesting. feature(Feature)i, 
satellite.growth_point: \[\]) 
Fontal:e), 
Among the 12 plans used by PIT are domain de- 
pendent ones as. well as domain independent ones. 
The latter are formalizations of RST relations that 
lead to small text structures. The domain depen- 
dent Plans lead to larger structures, e.g. whole para- 
graphs. Each plan, even if it can be seen as domain 
independent, contains a domain specific part, namely 
the requirements for nucleus and satellite which are 
inquiries to the inference engine that have to heed the 
names of entities, roles, and features of the knowl- 
edge representation. 
The planning algorithm uses four data structures: 
the plans, a tree, a stack, and a usedllst. The text 
structure tree is binary. The root contains the com- 
muuicative goal that initiates the phmning process. 
The nodes represent the executed plaits. Each node 
has successive nucleus and satellite edges.whose cor- 
responding nodes may be either empty or contain an 
explicit verbalization command or further plans. 
The stack is used as an agenda. Its elements are 
tuples consisting of the plan to be executed next and 
a pointer to that leaf of the tree where the subtree 
stemming from the execution of the plan should be 
added. 
The used hst is a bookkeeping device representing 
which plan has been'used for which entity. 
The plmming algorithm consists of three phases: 
first, the text structure tree is built by a top-down hi- 
erarchical plmmer (Sacerdoti 75) using reeursive de- 
scent. Second, the verbalization eonunmlds are col- 
lected by traversing the tree depth-first, left-to-right. 
Third, the entities to be verbalized are expanded by 
their attributes contained in the knowledge base and 
432 2 
~re passed on to the generator in a suit, able form. 
At the ;~tart of the planning process, i.e. when 
'~he communicative goal comes in, the tree, the stack, 
~nd the used-list are empty. If the plan library offers 
a:n appropriate plan to achieve the goal it is tested 
whether this phm has already been executed for the 
entity in question. If so, the execution is aborted, 
otherwise the plan is put on the used list. 
Next the requirements of the plan are checked, 
first, the ones common to both nucleus and satellite 
and then the nucleus requirements. If they cannot 
be met, execution of the plan aborted, otherwise the 
requirements of the satellite are checked. If they can- 
rot be met the corresponding plans of the satellite 
and the satelfite growth points are skipped. Are all 
requirements met, the new plans together with their 
pointers to that leaf of the tree where the subtrees 
~;hould be added are pushed onto the stack in the fol- 
lowing order: satellite growth points, nucleus growth 
points, satellite and nucleus. 
The second plan that is to be executed is popped 
fcom the stack and dealt with as described above 
with the addition that the agenda has to be updated 
when the tree has been expanded. The pointers of 
all plans to that leaf of the tree where a subtree has 
been added have to be changed in order to point to 
the nucleus of the new subtree. 
Planning stops when the agenda is empty. 
3. Shortcomings and possible exten- 
:,;ions 
The origlnal plans like the one shown above are 
based oil an extensive analysis of seven paragraphs 
describing places of interest in Diisseldorf. tlenee, 
they capture the typical structure of such descrip- 
tions and act as more flexible schemas that can be 
adapted to a user's needs by incorporating more com- 
municative goals. Nevertheless, problems arise when 
new plans are added or when old ones are changed. 
\]It proved to be difficult to say in advance which text 
structure will be the outconle of the planl~ing pro- 
tess. Through the top-down expansion of the text 
~tructure tree new plans may be inserted into the tree 
t'A places w\]lere they do not have the desired effect ¢,1 
the text structure. E.g. the plan \]cat,ires(X) may be 
the nucleus of the initiating plan deseriplion(X) and 
~dso satellite of a more fine grained plan. As those 
plans that have been pushed last onto the stack are 
executed first and no plan is executed twice the fea- 
tures may be verbalized at the wrong place in the 
text. 
Generally speaking, these problems point to the 
need to strictly separate the planning of the proposl- 
tlonal and the rhetorical. Although our hierarchical 
planner can be used successfully to plan the con- 
tent of the descriptive paragraphs we feel that a non- 
linear planning algorithm 1night be better snited for 
the planning of the propositional content followed 
by a hierarchical planner for the rhetorical struc- 
ture. Another problem is the domain dependence 
of the propositional planner which always snacks 
in through the requirements placed on nncleus and 
satellite. The :requirements are stated in terms of the 
knowledge representation langamge. The only partial 
solution to this problem is to use general terms in 
the planner and a separate mapping of these general 
terms onto the knowledge representation language. 
Our further research is directed in this direction. 

References 

Appelt 85 Appelt, D.E.: Planning English Sentences. 
Cambridge University Press, 1985. 

Bollinger et al. 80 Bollinger, T., Hedtstfick, U., Rollinger, 
C.-R.: Reasoning \]or Test Understanding - Kno~vledge 
Processing in the 1 '~ LILOG-Prototype. In: Met~ing, 
D. (Ed.): GWAI-89, 13 .1' German Workshop on Artifi- 
cial Intelligence. Sprhlger, 1989, 203-212. 

I'lllsemann 88 Busemann t S.: Satrface Trantt\]ovmations 
during the Generation o/ Written German Sentences. 
In: Bole, L. (tit.), Natural Language Generation Sys- 
tems. Springer, Berlin, 1988, 98-165. 

Cohen 78 Cohen, P.: On Knowing What to Say: Plan- 
ning Speech Acts. Techn. Report No. 118, University 
of Toronto, 1978. 

ltovy ~q~ tIovy, E.H.: \[ntegratinq Tenet Planning and Pro- 
daction in Generation. IJCAI-85, 8't8-851. 

lIovy 88 Hovy, ~.It.: Planning Coherent Multisentential 
Te:et. Proc. of the 26 ~h Annual Meeting of the AC.L, 
1988, 163-169. 

Mann and Thompson 88 Mann, W.C.,Thompson, S.A.: 
l~hetorical Structure Theory: Toward a ~mctional the- 
ory o\] te~t organization. In: Text 8 (3)~ 1988, 2'13-281. 

McDonald and Meteer 88 McDotaald, D.D., Mercer, 
M.W.: ~rom Water to Wine: Generating natural 
Language Test front Today'* Application Programs. 
Proc. of the 2 ''4 ACL Conference on Applied Natural 
Language Proce~shag, 1988, 41-48. 

MeKeown 85 McKeown, K.R.: Te~t Generation: Using 
Discoatrs Stralegle~ and Focu~ (7on~traints to Generate 
Natu~'al Langua9e Tezt. Cambridge, Cambridge Uni- 
versity Press, 1985. 

Moehan 70 Meehan, F.: The Metanovel: Writing Sio- 
,'ies Ay Camp,def. Ph.D. dissertation, Yale University, 
1976. 

Moore and Paris 80 Moore, J.D., Paris, C.L.: Planning 
7'e,'rt Jar Advisory Diologue~. ACL-89, 203-211. 

Moore and Swartout 80 Moors, J.D., Swartout, W.R.: 
A Reactive Approach to E:rplanation. IJCAI-89, De- 
trolt. 

Novak 8(t Novak~ It.-J.: G'enerating a Coherent Tecet 1)e- 
Fcribing a Traffic Sceite. COLING-86, 570-575. 

Novak 87 Novak, It.-J.: Teztgenerierung aus 
visuellen Daten: Beschreibungen van Strassenszenen. 
Berlin/Heidelberg/New York, Springer, 1987. 

Pletat and van Luck 80 Pletat, U., van Luck, K.: 
Knowledge Representation in LfLOG. In: Blash,s, K.- 
H., Hedt~tlick, U., Rolllngei, C.-R. (ed*.): Sorts and 
Types in Artificial Intelligence. Springer, 1989, (to ap- 
pear). 

Pollard and Sag 87 
Pollard, C., Sag, I.A.: An In/ormatlon-lJased Synta~ 
and Semantics. VoL I, Fundamentals. CSLI Lecture 
Nora's 13, Stanford, 1987. 

Saeerdotl 75 Sacerdoti, E.D.: A structure \]or plans and 
behaviour. North-Holland Publishing Company, Ams- 
t erdam, 1975. 
