SimSum: Simulation of summarizing 
BngRte Endres-N~ggemeyer 
Polytechmc of Hannover 
Department of InformaUon and Commumcalaon 
Hanomagstr 8 
D-30449 Hannover, Germany 
phone +49 511 92 96 606 
fax +49 511 92 96 610 
ben@~ks ~k fh-hannover de 
Abstract 
SimSum (Slmulatxon of Summariz- 
ing) simulates 20 real-world worlang 
steps of expert summarizers It pre- 
sents an empmcally founded cogm- 
Uve model of summarizing that oper- 
aUonahzes the discourse processing 
model developed by van Dijk and 
Kmtsch (1983) The observed strate- 
.gles of expert summarizers have 
g~ven nse to cooperating object-on- 
ented agents commumcatlng through 
dedicated blackboards Each agent ~s 
implemented as a CLOS object with 
an assigned actor at the mulUmedm 
user. interface The interface ~s real- 
ized with Macromed\]a D~rector 
CommumcaUon between CLOS and 
Macromecha Director Is medmted by 
Apple Events 
1 Introduction 
The SlmSum (Slmulatmn of Summarizing) 
system does what its name pronuses It simu- 
lates summanzang of human experts and thus 
produces a computaUonal cogmUve model of 
their processing The model concentrates on 
the specific features of summanzmg It pre- 
supposes "normal" text understan&ng and 
text producUon The slmulaUon serves 
scientific and presentational purposes 
• As usual, the computer model serves to 
explain and check the empirical 
cogmtlve model winch ~s Rs foundation 
• It prepares a cogmtlvely grounded 
approach to automatic summarnzmg, 
something hke agents runmng through 
the net and m response to a user's query, 
bnngmg home a reasonably short 
statement (a summary) of the knowledge 
avmlable 
• To its users of today, SlmSum shows m a 
movle-hke style how expert summanzers 
perform real-world workJng processes, 
thus complementing a textbook about 
summanzmg The advantage of the 
simulation ressembles that of a flight 
simulator As pflotes steer through 
possibly difficult situations ~n the 
physical world, summarizers work their 
way through a flood of mforrnatlon 
Both activities are cogmt~vely de- 
manchng People understand them better 
if they are presented wRh them m 
reahsuc setungs 
S~mulatmn approaches to summarizing are 
few and far between, but one can point to the 
SUSY system (Fum et al, 1982, 1984 and 
1985) as an ancestor of S~mSum SUSY 
aimed at following human performance m a 
hn~ted way, though keeping at a distance 
from real s~mulauon S~mSum represents 
progress wnh respect to SUSY, because ~t ts 
empmcally founded, ~t does a real slmula- 
uon, and ~t ~s implemented Furthermore, 
S~mSum renovates through ~ts mulUmed~a 
user interface 
For practical reasons, the SlmSum simulation 
ts restricted to 20 working steps involving 79 
agents They were chosen from an empmcal 
cogmUve model (a "grounded theory" - 
Glaser & Strauss, 1980, see also Lincoln & 
Guba, 1985, for the way to lmplementatton 
refer to Schrelber et al 1993) of summariz- 
ing wlnch comprises an intellectual toolbox 
of 552 strategies,, knowledge about the pro- 
cess organization and a set of interpreted. 
summarizing steps Its basis are 54 summa- 
nzmg processes of 6 experts from the USA 
and Germany The summarization processes 
were recorded by tinnlang-aloud protocols 
(Ericsson & Simon 1980, 1984) and ana- 
lyzed under the sclentfflc umbrella of the 
discourse comprehension model proposed 
by van DIjk and Kmtsch (1983) The experts 
being professmnals worlang m the context of 
mformatmn systems, three forms of summa- 
89 
nzmg occur abstractmg, mdexmg and classi- 
fying 
A simulation system such as SlmSum is 
bound to empmcal vah&ty, giving a reverse 
engmeenng of a cognmve process Such a 
reconstruction of human cognmve actavmes 
is possible because human experts subdlwde 
long cogmuve efforts hke sumrnanzmg into 
modules, called here working steps In the 
thinking-aloud record they are separated by 
boundary signals such as pauses or mterjec- 
Uons It Is these working steps that are re- 
constructed Put m sequence, they yield the 
model of the process 
Since the sequences m the SlmSum system 
are short, there ~s almost no chance for for 
seriously dealing w~th metacogmtlon (Flavell 
1981) m the system Hence metacognmve 
knowledge ts simply hard-coded m the form 
of working plans etc. 
In the following, SlmSum ts explmned first at 
the macro level of system archRecture and 
system components Then the descnptlon 
narrows down to the xmcro level of process- 
mg After a demonstration of the text repre- 
sentation, two exemplary relevance agents are 
discussed 
2 System overview 
SlmSum currently runs on Macintoshes with 
.System 7.5, a CD-drlve, a 17" momtor and 
some addmonal RAM as Is usual for multt- 
me&a apphcatlons It is Implemented as. an 
obJect-oriented blackboard system m CLOS 
and Macromedla Director (see figs 1 and 2) 
CogmUve strategies are represented by ob- 
ject-oriented agents grouped around thetr re- 
spective blackboards The agents are 
eqmpped wRh specmhzed knowledge, e g an 
indicator phrase lexicon or a basic represen- 
tation of SGML codes They process text 
structure m an SGML-hke coding and text 
meaning m a proposmonal representation 
The fact knowledge referenced m texts Is 
defined m document-specific ontologles On 
the screen the agents appear as ammated 
beasts The CLOS obJects simulate the cog- 
mtlve strategies They send out Apple Events 
to make "their" ammals on the stage perform 
according to the stmulatton 
An access system accommodates user mter- 
act~on..m a movte-hke style the user chooses 
a workzng, sequence and can interrupt at any 
time to get further explanaUons about what 
the cogmtlve agents do, how they are moti- 
vated empmcaUy and how they are Imple- 
mented 
user I 79 
access ~ I a~;e~ts 
i i model : = 
dedicated blackboards 
% 
\ 
\ 
\ 
\ 
\ 
"a, 
text : dictionary r r pr t, ao,  
I" ........... ........ 
• data processing -- explanatmns ~-~ start of process 
Figure 1 System archRecture 
90 
Figure 2 gives a screenshot of the. SlmSum 
multlmedra interface, presenting the rele- 
vance assessment agents at work The agent 
relevant-texthmt (a ladybird) is putting its 
can&date statements on the relevance black- 
board, whale the relevance agents hold and 
relevant-umt are s~tung on the bench, to- 
gether with the suspended control agents ex- 
plore for document explorataon and under- 
start(brig (the bee) and construct for target 
text producuon (the spider) Below, can we 
see the document blackboard with the repre- 
sentaUon of the source text, showing ~ts 
meamng panel, the scheme representation 
stating the document orgamzaUon, and the 
theme representaUon stonng the theme, ~ e 
the top of the macrostructure as far as known 
to the summarizer At the bottom, a comment 
explains what is. currently happemng on the 
screen 
The central system components have been 
derived from observation 
• Cooperating agents 
Experts use recurnng goal-oriented proce- 
dures, corresponding to the strategies 
sketched by van Dijk and Kmtsch (1983) 
These procedures or strategies were opera- 
t~onahzed into mtelhgent agents of the com- 
puterized system Agents consist of a script 
that defines how they deal with tbelr task, 
they have a general cornmumcauon compo- 
nent that allows them to exchange messages 
with other agents and to access global knowl- 
edge sources, they may possess private task- 
oriented knowledge, and they are eqtupped 
with task-onented data wews for input and 
output Control agents (responsible agents 
for a blackboard - see below) are m ad&uon 
assigned a little scheduler They actlvate thetr 
subordinates by d~rect message passing 
Data exchange between agents takes place v~a 
the blackboards The agents keep to the 
commumcat~on rules Strategies / agents co- 
operate in concrete tasks such ass dectdmg 
about relevance or semng up a target sum- 
mary statement Agents may use products of 
other agents, but since they have hmlted 
tasks, they have no soplust~cated commum- 
cataon behavtour such as bargaining or &s- 
cussing 
Hgure 2 A screenshot of the SlmSurn user interface relevance assessment agents 0adyblrds) are 
busy wlule exploraUon (done by bees) and target text producUon (by spiders) are suspended 
91 
s ' Blackboards 
Agents need commumcatlon areas, as a 
medium of cooperation Functionally 
speaking, these are blackboards (Selfndge, 
1959, Carver & Lesser, 1994, Engelmore & 
Morgan, 1988) StmSum blackboards are 
dedicated They are used for reception, stor- 
age of the input text representation, relevance 
assessment, target summary construction and 
so on Central is the document blackboard 
that stores and organizes all knowledge ac- 
qmred from the source document (of fig 2) 
Since m the case of professional summariz- 
ing cogmtlve processing is modular, the 
agents work m task-specific groups using a 
dechcated blackboard For instance, the rele- 
vance assessment agents use the relevance 
blackboard to put the relevance judgement 
together Every blackboard has a control 
specialist It orgamzes the work of the group, 
sums up what they have acineved, executes. 
the group opinion and dehvers the result to 
the next blackboard - 
• Knowledge base 
The SlmSum knowledge base ts a common 
knowledge store compnsmg a text represen- 
tation winch holds all texts in the system and 
an ontology of the concepts which are 
needed to deal with them 
3 Computer-oriented discourse representa. 
tion 
Since summarizing is a text and reformation 
processing task, we have to represent those 
surface text passages and text meamng umts 
m the system winch are really worked upon, 
concentraung on semantic and pragmatic 
structures The representation must support 
pragmatic text handhng and deal with boll- 
sUc text structures as well as wtth local rmcro- 
structures and layout features, because doc- 
ument structure knowledge ts a core item of 
a professional summanzer's competence 
• • The practical coding of the vlslble doc- 
ument arcintecture follows SGML con- 
ventlons SGML tags like "<hl I> 
. <hi It:> "Introducuon" </hl It> "assign a 
layout feature derived from content 
structure In the example m table 1, the 
secUon begmmng Is re&cared by <hi 1> 
Its Utle is included by the tag pmr 
<Ill lt> and </hllt> 
s The passages that are really read m the 
simulated working steps are furthermore 
coded m flrst-order-loglc-hke proposi- 
uonal form (see table 2) Dunng text 
coding we deliberately chose fitting 
predicates and standarchzed presentation 
(e g ordering of arguments, mateinng 
semanucally nearly eqmvalent words m 
one concept) Dommn prechcates are dls- 
tlngmshed from predicates that describe 
an mteracUon between the authors and 
their readers 
<hl 1> <hl It> 1 Introduction </hl It> • 
<bodyl I> <p> Tins study forms part of the project "Atmogenous and geogenous 
components m the heavy metal balance of forest trees" The goal of tins project is, on 
the basis of the distribution within the tree, to trace paths of heavy metal absorpuon 
and the regularities of their mternal redistribution Furthermore, R ts anned to estimate 
absorpUon and rechstnbuuon rates In order to obtmn as clear results as possible, the 
majority of trees analyzed were located m areas with atmogenous or geogenous pol- 
lution In conunuatlon of the prewous studies, winch concentrated on trees m contam- 
inated dead ore areas and Black Forest locations with low atmogenous polluUon, the 
following reports about trees influenced by Ingh atmogenous deposits m the chstnct 
of Stolberg </Ix> </bodyl 1> </hi I> 
Table 1 Text representation, SGML style coding of an introduction 
92 
3 domain_exist (introduction) 
4 domain pollute (heavymetals, forest_trees .... \[geogenous, atmogenous\]) 
5 domain investigate (proJect, 4) 
6 domain_partmipate (study_this, 5) 
7 domain absorb (trees, heavy_metals,, paths) 
8 domaln_redmtnbute (trees, heavy_metals,, internally,, regularity) 
9 dommn_distnbute (trees, heavymetals,, internally) 
10 dommn_trace (project, \[7, 8\], 9 .... aim) 
11 domain_estimate (project, \[7, 8\],,,, ram)) 
Table 2 Text representation, beginning of the introduction (proposmonal coding) 
• To account for discourse level document 
structures, SlmSum uses text-type specific 
superstructures (Kmtsch & van Duk , 
1983) From a practical point of view, su- 
perstructures consist of semantic compo- 
nents which are hnked by discourse rela- 
tions In SlmSum, these are RST relations 
(RST Rhetorical Structure Theory - 
Mann & Thompson, 1987, Hovy, 1993) 
While the SGML and the propositional 
representations are precoded, the dis- 
course level document structures are re- 
constructed during summarizing The 
cognitive agents install the respective RST 
relations Only a few of the most neces- 
sary and most simple RST relations have 
been implemented ELABORATION, 
RESTATEMENT, PURPOSE, CAUSE/ 
RESULT, EXAMPLE 
A small parsimonious ontology has been 
coded for every document, where the used 
concepts are organized m a small and very 
flat hierarchy The ontology is divided into 
two parts according to Penman (1989) The 
upper model is domain independent and 
therefore used for all texts m the system, 
whereas the lower model is dommn specific, 
so that one is modelled for each document 
The agents do some basle lrfferencmg such 
as comparing text units with knowledge base 
entries and installing relanons from a fixed 
set between text umts 
4 Agents 
'The core of the SlmSum simulation are ob- 
ject-onented agents As representatives of the 
empirically found eogn!tive strategies they 
manage the reducUon of a large document to " 
a short summary Agents &ffer m the repre- 
sentations they work upon Some of them are 
sensitive for SGML tags, others need the 
propositional presentation to run their meth- 
ods 
In the &msum system, 39 agents are mod- 
elled m great detml They are revolved m the 
central information reduction task of sum- 
manzmg, e g the relevance agents Reading 
and wnting strategies are realized carefully 
only m so far as they are specific for profes- 
sional summanzaUon, otherwise they remain 
black box agents About half of the agents 
are "real" agents and the rest are "pseudo"- 
agents For instance, the explore agent Is a 
black box agent of understanding It fakes 
text comprehension by assigning input pas- 
sages a precoded propositional representa- 
tion The reorgamze agent is a black box 
agent as •well It Is presumed to impose En- 
ghsh grammar and spelling which is not a 
specific subtask of professional summariz- 
ing Therefore the agent functions more or 
less as a placeholder 
The agents fall into the following functmnal 
classes planmng and control, exploration, 
relevance assessment, target textconstructton, 
quahty enhancement, formulation, and gen- 
eral knowledge processing In addition, there 
are rmnor agents such as readers and writers 
• To make the agents more concrete, we dis- 
cuss m the following two "real" relevance 
agents that happen to be good old acqumn- 
tances of everybody m automatic summanz- 
mg relevant-texthmt (realizing the indicator 
phrase method, see, e g, Palce, 1990 and 
Borko 1968).and relevant-call, which as- 
sesses the importance of an entlty by measur- 
ing its distance from the theme (pnnciple 
used m Jacobs & Rau, 1990, .McKeown, 
1985, Trabasso & Sperry, 1985) More 
about agents is found m Endres-Niggemeyer 
et al (1995) and m Endres-Niggemeyer 
. (I 997) 
93 
Relevance agents work under the control of 
hold, the responsmble agent for the relevance 
blackboard, (cf fig 2) Since the skdled re- 
d.uct~on of document meaning to the most 
relevant items ~s central to professional sum- 
manzatzon, hold ms m charge of the core of 
the whole summarizing process 
• Relevant-texthint 
The relevant-texthmt agent mmplements the 
"mdmcator phrase method" known since the 
early days of automatic abstracting It ex- 
ploits cue phrases by which authors quahfy 
their statements, assunung that the quahfica-. 
tlon apphes to the scope of the indicator 
phrase By rots mere presence, a (posmve) re- 
&cater phrase, expresses the author's empha- 
sis and suggests the relevance of the state- 
ment m its scope In addmon, cue phrases 
often explain what the author announces, e g 
a new fin&ng or the content of the conclu- 
smon, and ~ts role m the" document 
Relevant-texthmt reads the proposltmons on 
the meaning panel of the document black- 
board (see fig 2) To make out relevant 
propos~tlons, mt uses a private dlcuonary, 
where the potential mdmcator predmcates (cf. 
table 3) are hsted Since the &cuonary en- 
tries are annotated with mterpretaUons, the 
agent can draw the attention of other agents 
to these proposmons by passing them parts 
of its pnvate knowledge . " 
Relevant-texth~nt recogmzes the mdmcator 
predmcates by simple pattern matclung as 
contmmng an. indicator phrase, matching rots 
dmctaonary entry with a proposmon such as 
proposmon 5 mn table 2 Consequently, the 
agent annotates proposmon 4 as describing 
the project theme and therefore as Important 
and puts it together wroth others on the rele- 
vance blackboard (see fig 2 and table 4) 
• Relevant-caU 
Relevant-call recoguizes a text meamng totem 
as relevant because it hnks it to the document 
theme (see figure 3) The agent needs the 
themauc structure and, as a candidate for 
linkage to the document topmc, a text 
proposmon The agent checks whether an 
open RST-type hnk of the document theme 
is able to attach the candidate If so, the 
proposmon Is regarded as relevant and added 
to the document theme 
Theme-of-document. 
dommn_lnvestlgate (project, X) 
domam_parUclpate (studyjhls, X) 
dommn_esumate (project, X, ram) 
mteracuon_report (author, X) 
dommn contmue (researchers, Y, X) 
Methods-of-research 
dommn_select (researchers, X) 
dommn_obtmn (researchers, 
results_clear, X, ram) 
Research-background 
dommn_concentrate (research, X, 
past) 
• Table 3 Some proposmons from the indicator phrase dictionary 
4 dommn_pollute (heavy_metals, forest_trees .... \[geogenous, atmogenous\]) theme-of- 
document 
7 dommn_absorb(trees, heavy_metals,, paths) theme-of-document 
8 dommn_redlstnbute (trees, heavy_metals,, internally,, regularity) theme-of-document 
9 domam_thstnbute (trees, heavy_metals,, internally) theme-of-document 
12 domain_select (researchers, \[trees/locatlons_polluted_atmogenous_lugh/, 
trees/locations_polluted geogenous_hlgh/\]) • methods-of-research 
16 dommn_mfluence (deposlts/atmogenous thgh/, trees,, stolberg_&stnct) theme-of~oc- 
ument 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
Table 4 Choice of what relevant-texthmt judges relevant 
94 
extension 3 
(proJect, 
pollute 
(heavy_metals, 
forest_trees, 
\[atmogenous, 
ELABORATION 
extension 2 
(heavymetals, 
forest_trees, 
\[atmogenous, 
olluto 
(heavy_metals' 
foresttrees) 
extension 1 
Figure 3 Relevant-call expands the document theme 
To find the document theme, relevant-call 
accesses the theme panel of the document 
blackboard The agent tries to attach propo- 
smons discovered by other (data-oneuted) 
agents For instance ~t picks up proposmon 4 
recommended by relevant-texthmt because 
it states what the research ~s about ('This m- 
vest~ganon forms part of a project " - cf 
table 4) Relevant-call tries all available 
RST-relataons m order to hnk proposltaon 4 
to the document theme (in extenston 1) It ~s 
easy to see what happens proposmon 4 
rephrases the theme, the concepts "pollute", 
"heavy_metal", and "forest_trees" of the 
theme are repeated The theme and the text 
proposmon unify, but proposmon 4. bnngs 
some addmonal mformatton about the 
(\[geogenous, atmogenous\]) components of 
contmmnataon This corresponds to an elabo- 
rauon of the theme Consequently, the 
proposRon. Is attached by an ELABORATION 
hnk Th6 new hypothesis of a topic structure 
~s given m figure 3 At that moment, two new 
proposltaons have been attached tothe theme, 
so that the theme has three extensions 
5 Conclusion 
Advancing the sc~entfftc frontiers of text 
summanzaUon presupposes more knowledge 
about the way summartzatton works The 
mare frmt of the empmcal mvesUgat~on be- 
hind SlmSum is an tmage of the summanza- 
laon process which Is detmled enough to lay 
the foundattons for a stmulat~on Since the 
resulting summarization model incorporates 
the know-how of human experts, tt has good 
prospects of presenting powerful techmques 
Summarizing by cooperating cognmve 
agents seems to be such a pnnclple 
The researchers have reached their aim to 
show that an observauonally founded Im- 
plementaUon of summarizing processes ~s 
possible However, SlmSum Is a system m- 
the-small It sufftcesto demonstrate how the 
summarization agents work m thetr cogmUve 
environment To meetpracttcal challenges 
such as text summarizing m the WWW, a 
much more comprehensive system must be 
realized Tl~s means m pamcular 
• providing knowledge bases of real-world 
size, be they private ones of agents or 
pubhc resources of the whole system 
• choosing the most useful strategies or 
agents and malang them flexible to deal 
with any legmmate data 
• using text understanders or reformation 
extracUon components as well as genera- 
uon systems provtded by colleagues 
95 
6, Acknowledgements. 
The SlmSurn development has been funded 
.under grant F 916 00 by the German Federal 
Mlmstery of Education and Research 

References 
Borko, H (ed) (1968) Automated language 
processing New York Wdey 

Carver, N, & Lesser, V (1994) Evoluaon of 
blackboard control archRectures Expert 
Systems wzth Apphcanons 7, 1-30 

Endres-Nlggemeyer, B (1997) Summarizing 
text (forthcoming) 

Endres-Ntggemeyer, B, Miner, E, & Slgel, 
A (1995) How to implement a natural- 
lsttc model of abstracting four core 
working steps of an expert abstractor 
Informatwn Processing & Management 
31(5), 631-674 

Engelmore, R, & Morgan, T (Eds) (1988) 
Blackboard systems Wokmgham Addi- 
son Wesley 

FEncsson, K A, & Simon, H A (1980) Ver- 
bal reports as data Psychologwal Review 
87, 215-251. 

Ericsson, K A, & Simon, H A (1984) Pro- 
tocol analysts Verbal reports as data 
Carnbndge MA MIT Press 

Flavell, J H (1981) Cogn!tlve momtonng In 
W P Dtckson (FEd), ChtMren's oral com- 
mumcatwn skdls (pp 35-60) New York 
Aeadenuc Press 

Furn, D, Gmda, G, & Tasso, C (1982) For- 
ward and backward reasomng m auto- 
matlc abstracting In COLING Proceed- 
rags of the 9th lnternatwnal Conference 
on Computattonal Lmgmstws (pp 83- 
88) Prague 

Furn, D, Gmda, G, & Tasso, C (1984) A 
proposmonal language for text represen- 
tation In B G Bara & GG Gmda (Eds), 
Computatwnal models of. natural lan- 
guage processmg (pp 121-150) Am- 
sterdam North-Holland 

Fum, D, Gmda, G, & Tasso, C (1985) 
Evaluating tmportance A step towards 
text summanzatson In IJCAI Proceed- 
rags of the 9th Internatwnal Jomt Con- . 
ference on Artlficzal Intelhgence (pp 
840-844) Los Altos CA Kaufmann 

Glaser, B G, & Strauss, A L (1980) The dzs- 
covery of grounded theory Strategtes for • 
quahtatzve research (llth ed) New 
York Aldme Atherton 

Hovy, E (1993) Automated dtscourse gen- 
eration using discourse structure rela- 
uons Artohczal Intelhgence 63, 341-385 

Jacobs, P S, & Rau, L F (1990) SCISOR 
Extracting mformatton from on-hne 
news Commumcauons of the ACM 33 
(11), 88-97 

Kmtseh, W, & van DIjk, TA (1983) 
Strategzes of chscourse comprehenston 
Orlando FLA Academic Press 

Lincoln, Y S, & Guba, E G (1985) Natural- 
~st~c mqmry Beverly Hills CA Sage 

Mann, WC & Thompson S A (1987) 
Rhetorical Structure Theory A Theory 
of Text OrgamzaUon In L Polany (FEd) 
The Structure of Dtscourse Norwood, 
N J Ablex 

McClelland, J L, & Rumelhart, D E (1981) 
An interactive activation model of con- 
text effects tn letter perception Part 1 
An account of b.~le fln&ngs Psycholog- 
wal Remew 88, 375-407 

MeKeown, K R (1985) Text generatzon 
Uszng dzscourse strategies and focus 
constraints to generate natural language 
text Cambndge Cambridge Umv Press 

Norman, D A (1983) Some observations on 
mental models In D Gentner & A L 
Stevens (Eds), Mental models (pp 7-14) 
Hlllsdale NJ Erlbaum 

Patce, C D (1990) Constructing literature 
abstracts by computer Techniques and 
prospects Information Processing & 
Management 26 (1),171-186 

Penman Project (1989) PENMAN documen- 
tatton the primer, the user grade, the ref- 
erence manual and the Nzgel manual 
Techmcal Report USC/Informaaon Sci- 
ences InstRute, Marina del Rey, Cahfor- 
nla 

Schrmber, G, Wlehnga, B, & Breuker, J 
(1993) KADS A prmctpled approach to 
.: :. ~ knowledge based• system development • 
London Academic Press 

Selfndge, O (1959) Pandemomum A 
paradigm for learmng In Symposmm on 
the mechamzatwn of thought processes 
London HMSO 

Trabasso, T & Sperry, L (1985) Causal re- 
latedness and. importance .of story events - 
Journal of Memory and Language 24, 
595-611 
