COSY-MATS: An Intelligent and Scalable Summarisation Shell 
Maria Aretoulaki 
Dept of Pattern Recognlt~on (Computer ~Sc~ence 5) 
Unlvers~ty of Erlangen-Nuremberg 
Martensstrasse 3 
D - 91058 Erlangen, Germany 
aret oula@~nformat ~k. un~-erlangen de 
Abstract 
In tins paper, an architecture is presented 
for robust and portable summansatlon, 
COSY-MATS COSY-MATS Can avmd the su- 
perfimahty and domain-dependence of IE 
approaches by means of lngh-level (prag- 
matic and rhetorical) content selectmn fea- 
tures It can also obviate the text type- 
dependence and cumbersome computation• 
revolved m NLU-based snmmansatl0n sys- 
tems, because surface criteria are add~t~on- 
ally used m the content selectmn process, 
as are ~dent~fied mappings between those 
and the htgh-level features In ths way, 
COSY-MATS should retain ~ts generic and 
scalable character, wlnle also pernuttmg 
mtelhgent apphcatmn-spec~Sc processing 
1 Motivations behind the Design of 
COSY-MATS 
The goal of the research reported here has been 
to develop a fleable, easdy-portable and scalable, 
but also efficient and robust, NLP system that au- 
tomatically generates summaries of real-world un- 
restricted texts To tins effect, an archttecture 
was designed for a hybrid COnnectmmst-SYmbohc 
MAchine for Text Summansatlon (henceforth, 
COSY-1VIATS) (Axetoulah, 1996) 
A major concern m des~gmng COSY-MATS has 
been to identify content selectmn features that are 
generic and apphcatmn-mdependent (Section 2) 
The features should be apphcable to any text, Ir- 
respective of domain or text type Tins is so that 
COSY-MATS 1s readdy portable to dflferent operation 
ehv~onments vath a nnmmum amount of cnstonnsa~ 
tton The isolation of such features would provzde a 
permanent infrastructure for both content selection 
and analysis The front-end text analysis modules 
can be developed so that they are geared towards the 
s,mmansatlon task, rather than text understand- 
mg m general, wlnch is computatlonally-mteuslve 
Thus, these modules need only perform an analysis 
that 1s suiBc~ent for the evaluation of the selected 
content selectlon features The estabhshment of um- 
versal unportance determlnatlon criteria means that 
the permanent set of analysis, interpretation, con- 
tent selection and generation processors can be ex- 
tended with apphcatlon-specfflc modules dunng the 
porting of COSY-MATS Tlns m also what renders 
COSY-MATS a type of summartsatson shell Slgmfi- 
cantly, the computatlons of the supplementary mod- 
ules wdl already be accommodated for m the stan- 
dard flow of processing of the system by wrtue of 
these features (Section 3) 
Admittedly, the ldentxficatlon of content selectlon 
features of general apphcablhty is a very d,mcnlt 
task Tins is demonstrated m the lnmtatlons of the 
two mare trends m current surnmansatlon research 
(cf (Aretoulah, 1996)) There are In/ormatzon Ez- 
trachon (Is) enwronments, wlnch perform a super- 
iiclal and partial analysis of the input text based on 
the progressmn of keywords and apphcatmn-specfllc 
phrasal patterns thereto, e g (BT, 1994, Jacobs 
and Rau, 1990, L,lhn, 1958, MUC-5, 1993, Palce, 
1981, Patce, 1990, Salton et al, 1994) The prob- 
lem with IE systems ~s that, although they can 
be used very efllclently on any type of text, they 
are domain-dependent and hkely to produce mac- 
curate output Ths ~s due to their excessive re- 
hance on speclahsed content words There are also 
systems winch are based on Natural Language Un- 
derstandsng (NLU) methods revolving deeper pro- 
cessmg Apart from syntactic and lexlco-semantlc 
analysis, the lnerarclncal rhetorical orgamsatlon of 
the source text can also be taken into account, as 
can certain aspects of the context of the dmcourse, 
e g (Ganghano et al, 1993, Lehnert, 1981, lVIltkov 
et al, 1994) Such more soplnstlcated types of sys- 
tem, however, are prohlbltlvely slow as a result of 
the extenmve processing revolved They are also 
very fragde, because the hlgh-level knowledge em- 
• ployed is usually hand-coded and hence arbitrary 
and incomplete Even when this knowledge has 
been acqurred automatically, e g (Maybury, 1993, 
Soderland and Lehnert, 1994), it is apphcatlon- 
dependent Consequently, despite their occasional 
74 
I 
i 
i 
I 
l 
I 
I 
! 
I 
domain-independence (e g (En&es-Nlggemeyer and 
Neugebauer, 1995, Ono et al, 1994, Rau et al, 1993, 
Sharp, 1991)), NLU approaches are ---on the whole-- 
specmhsed m a particular text-type 
For the demgn of COSY-MATS, a ~o~$c and 
umfymg approach has been adopted that revolves 
both extrahngmst:c, NLU-type, analysts and selec- 
tive statmt~cs-based lmgmst:c processing reminiscent 
of IZ, m co-ordination S~mdarly to NLU, analysts 
m COSY-MATS is sufl~cmntly deep for the semantic, 
rhetorical and contextual aspects of the input text 
to be considered m content selection In contrast 
to what the case ts with such systems, however, the 
computation of these d~verse aspects of the text ts ef- 
ficmnt Thin ts because objective cues on the surface 
of the text are also explmted m COSY-MATS, echoing 
. the I~ approach Nevertheless, unhke rE, these cues 
are function words'and g~nenc content words winch 
point towards the ingh-level functmus of the respec- 
tive textual umts m the context of the dmcourse, 
while at the same tnne being domain-independent 
Thus, apart from ldent~fymg umversal content selec- 
tion criteria that should render COSY-MATS portable 
and scalable, the research reported here has also at- 
tempted to establtsh mappings between the concrete 
and the more abstract criteria m the devmed feature 
scheme, so that the system ts also mtelhgent and 
pruct~cnl, ~ e so that the evaluation of these abstract 
criteria ts fully automated (Section 2) 
2 Intelligent Content Selection 
Criteria 
In order to identlfy generic content eelectlon features 
that can be used by COSY-MATS m any apphcatlon 
context, an extensive corpus analyms was camed out 
on a variety of real-world texts Three mare types 
of text have been analysed newspaper articles, scz- 
ent~fic papers and (s~entsfic) author abstracts The 
subcorpus of newspaper artxcles (160) m extremely 
dwerse m both ~ts content • and form The topics 
range from business news and legal reports to so- 
cial commentary, me&cal msues and pohtlcs Slma- 
larly, the other two subcorpora constst of 170 articles 
and abstracts, respectxvely, that pertain to scmnttfic 
fields such as computer science, the natural scmnces, 
as well as pinlosophy and hngmstxcs In addltmn, 
the texts are of varying length from half a page m 
the case of the abstracts and most news agency re- 
ports, to four or more pages, when smentlfic papers 
and newspaper special reports are revolved Conse- 
quently, apart from covering a range of subject do- 
mama, the corpus used m designing the content se- 
lection processes m COSY-MATS a\]8o represents more 
than two text types 
The corpus was analysed both on the surface and 
on more abstract levels Given the chvermty of the 
types of text and the writing styles exh~inted m the 
corpus, regulantms regarding the rhetorical develop- 
75 
ment of the texts and the central mformat4onal umts 
thereto could not be easdy estabhshed 0nly m the 
case of the smenttfic papers and thetr abstracts could 
any statements be made on the logical progresmon of 
the presentation of the content, from the purpose of 
the research, to the methodology, the experimental 
set-up and the evaluation of any results (d (Gop- 
ink, 1972, Jordan, 1993, Lucas et ai, 1993, Mmzell 
et al, 1971)) The newspaper articles were mainly 
studmd m terms of groups of ad3acent sentences and 
the rhetoncal relatlonsinp between them (d (Ono 
et al, 1994)) No generahsatlons could be made re- 
garding their top-level orgamsatlon 
A number of theories of pragmatlcs, dmcourse 
analyms and text development have prowded use- 
ful concepts for tins study of the corpus at a Ingher 
level 
• a) theories winch are preoccupmd with the 
comrnun:catsng agents, their goals, plans and 
behefs, such as Speech Act Theory (Austin, 
1962, Searle, 1969), Rhetorical Structure The- 
ory (RAT) (Mann and Thompson, 1987), or AI 
research on scripts (Lehnert, 1981, Schank and 
Abelson, 1977) and behef ascnptmn (Wdks and 
Balhm, 1987) 
• • b) theories on the tracking of the dsscourse h~- 
tory by means of identlfymg the focused items 
thereto, e g (Grosz, 1986, Hobbs, 1978, Relch- 
man, 1985, S1dner, 1983, Webber, 1983) 
• c) theories of cohesson and coherence and how 
these are m~mfested on the surface of the text, 
e g SysteInlc-~uuctlonal Lmgutstlcs (Halhday 
and Hasan, 1976) and the Problem-Solutmn m- 
formation metastructure (Hoey, 1994, Jordan, 
1984) (cf (Pmce, 1981)) 
The &verslty of the subject matter covered m 
the corpus has meant that specmhsed keywords 
were ignored m its analysm Instead the emphasts 
was placed on functlon words and.regular general- 
language content words winch are assooated wlth 
the mstantlatmn of the semantlc, rhetorical and 
pragmatlc functlous cous~dered Such lemcal xtems 
can be employed as markers, not only of the de- 
velopment of the dmcourse but also of the focused 
and central points thereto In thin process, the var- 
1ous cohesion and coherence theorettlcal frameworks 
were very mfluentlal, as were the computatlonal ap- 
proaches to focus pre&ctmn and identtficatmn 
As a result of thin corpus analysm at the sur-. 
face and more abstract levels, 87 features have been 
identified as relevant to content selectmn and im- 
portance determmatlon across domains and, largely, 
text types (Aretoulakt, 1996) Three descnptlve 
levels are used for thezr classflicatlon the prag- 
matzc, the :ntermed:ary and the surface, m decreas- 
ing order of abstractlon The three levels reflect, 
m a sense, the three maul trends m dtscourse the- 
ory identflled Thus, the 24 pragmatxc features 
! 
 omm - ' 
mcatmg agents Pragmatxc features such as Plan 
and Goat for instance, are remmmcent of AI work 
on scripts (Sch~nk and Abelson, 1977), Elabora- II 
twn and FEzplanatwn can be parallelled to P,.STrela- " 
ttons (Mann and Thompson, 1987) 
• ,, 
developed COSY-MATS (of ectlon 3) To tins ef- 
fect, a number of mterlevel mappings were identl- | 
fled both between the pragmatxc and the lower lev- . 
els, and between the mtermechary and the surface 
Clusters of Pragmatic Features 
Figure 1 
The intermediary .features (Fig 2) represent 
rhetorical semantxc criteria often employed m the 
processing of focus reformation and m anaphor chs- 
amblguatlon For example, Topscalzsahon, Focus 
Change, Cardznahty and Elhps~ have all been used 
m computatmnal contexts such as (Hobbs, 1978, Re- 
xchman, 1985, Sldner, 1983, Webber, 1983) 
Finally, the surface features (Fig 3) comclde 
mostly with exphmt cues m the text wlnch de- 
note cohesive and coherence relatlous among sen- 
tences (d (Li~hn, 1958, Pmce, 1981)) The 
Functzon Word and the Common Content Word 
Pools, for instance, conmst of lemcal 1terns with 
a semantic/rhetorical load exteuslvely dmcussed 
m a Systennc-Fauctxonal (Coulthard, 1994) and 
Problem-Solutxon context (Jordan, 1984, Jordan, 
!995) Consequently, by using features such as these 
m COSY-MATS, all three levels of language --from 
the low-level surface to the Ingh-level pragmatic-- 
can be CoUectlvely consxdered m order to 'hohstl- 
cally' determine the unportance of m&wdual propo- 
S~tlons m a text 
Apart from tins grouping of the features into dif- 
ferent levels, the surface and the mterme&ary fea- 
tures proposed m tins scheme have also been used 
to objectify the abstract pragmahc features Tins 
was m order to faclhtate the automatic evaluation of 
the latter during the actual operatxon of the fully- 
levels These mappings were compiled m a manual 
which was used by 5 subjects m encoding texts from 
tins corpus (Aretoulakl, i996) The encoded texts 
were then employed for the empzncal testing of a 
prototype of the content selection module, reported 
m Section 4 Example mappings are given below 
• The pragmatic feature Repehtzon m correlated 
to the surface features Personal and Possessz~e 
Pronouns and Demonstr~tz~es (Sldner, 1983) 
It is also associated with the mtermechary Focus 
Change (Sldner, 1983, Webber, 1983).and El- 
lspszs (Hovy, 1987) Tins m because the central 
topxcs m a text are often resumed by means of 
.anaphora, both m the same sentence and later 
on m other nnportant.sentences 
• The presence of unpersonal phrases m the Pas- 
ss~e on the surface level m extensively used to 
express a Generahsahon on the pragmatic level 
The latter denotes a central text umt by deft- 
ration (Gopmk~ 1972, Lehnert, 1981, van Dyk 
and Kmtsch, 1978) 
• The surface Negatzon m correlated to the mter- 
me&ary Contrast (Jordan, 1984) 
• Modals such as "should" are also exteusxvely 
used on the surface of discourse, when propos- 
ing, evaluating, or making tentative claims m 
general Thus, tins feature m also related to 
the pragmatic Behef/Doubt, Volttzon/Fear and 
Plan (cf (Fakumoto and TsUjnl 1994)) 
Ewdence for the usefulness of the mterlevel map- 
pings proposed m the context of the COSY-MATS con- 
! 
76 
mm 
Clarets of Surfa~ Featmw 
~gure 3 
tent selectson feature scheme was provided by vaL 
sdatmn tests regarding the tunfornnty of the few 
ture evaluation practices among the human en- 
coders (Aretoulakl, 1996) The encoding of an iden- 
tical part of the corpus by means of all the pragmatic 
features showed that there was a total of 79 6% 
agreement among the encoders on the evaluation of 
the pragmatic features, using the above-mentioned 
manual Consequently, the identified surface .and 
other less subjective features can be fully exploited 
later on for the automation of the encoding of the ab- 
stract pragmatic features The vahdatwn tests also 
mdtcated that there was 96% agreement on which 
of the corpus sentences were m~portant and wlnch 
nmmportant for the corresponding texts 
3 A Scalable Architecture for 
Intelligent Sumn~risation 
Having identified 'umversal' content selection fea- 
tures, as well as some of the ways these interact vath 
each other, the following arc\]ntecture w~ designed 
for a full-scale zmplementation of thecosY-MATS 
s-mmansatlon shell (Fig 4) (Aretoulala, 1996) Ev- 
ery sentence m the text to be sttmmansed s is first 
processed by a cluster of standard symbohc analy- 
sets, morphological, syntactic, semantsc and prag- 
matic The resnlt of tlus processing ~s the e~valua - 
tion o£ a set of basic hngtustic and extrahngtustic 
Xwlnch is assumed to be integral and coherent, rather 
than a random co/lectson of prop0smons, 
t0uc 
0 
S 
Y 
M 
A 
T 
S. 
W 
I~BOUC 
PdPPU~'t'tQtt ~ 
Figure 4 The Arclutecture of COSY-MATS 
features that prowde the input for a Cascade of low 
and lngher-level Artdic~al Neural Networks (ANNS), 
each responslble for specific subtusks The low-level 
ANNs map hngmstxc features (surface and mtermech- 
ary) into extrahngtustzc features (mtermedlary and 
pragmatic) The pragmatzc features provide the m- 
put to the lnghest-level content selection ANN that 
ultunately determines the relative degree of nnpor- 
tauce of each sentence This latter ANN zs also the 
only component of COSY-MATS that has been im- 
plemented to date Finally, the sentences selected 
as unportant during the content selectlon phase vnll 
be used as the basis for generating either a compre- 
henssve summary or a more concise abstract (Are- 
toulakl, 1996) This processing wdl take place m 
another duster of symbohc processors, almost sym- 
metric to that used for text analysm and mterpre- 
tatzon It is here that the plan-rag and the actual 
synthesm of the summary/abstract wdl be reahsed 
However, it is unportant to note that the output hat 
of the best-sconng sentences produced by the con- 
tent selection ANN can a/so be used to pro~nde a draft 
summary, z e a concatenation of already-e~tmg 
sentences instead of an original text (cf (Kuplec et 
77 
al, 1995)) Tins m also the only type of generatlon 
that m currently preoccupying tlns research (cf Sec- 
tmn 4 1) 
D~plte the dominance of the generic modules 
thereto, COSY-MATS does provide for the incorpo- 
ration of apphcatlon-spectfic mformatlon F~rst of 
all, the architecture m lnghly modular, so that new 
specaahsed processors can be --m prmclple--- rumply 
plugged m The smaphcaty of the interface between 
the various modules means that new modules that 
are either symbohc or connectmmst can equally well 
be accommodated For example, m adchtmn to the 
extstmg lower-level ANNS, other ANNS can be easily 
incorporated winch have been trained to recogmse 
specfllc keywords and structural phrases that dflfer- 
entmte one domain or text type from the other m 
expressing the same rhetorical and pragmatic func- 
tions Hence, COSY-MATS can function as a shell for 
the btuldmgof specmhsed summ~rtsers 
As regards the front-end symbohc analysers, the 
processing that will take place thereto wall be dic- 
tated by the type of data that needs to be computed 
m the ANNs The latter computatmn, m turn, wdl 
be based on .the ldentzfied generic and apphcatmn- 
specflic mappings across the three levels of descnp- 
taon the pragmatic, the mterme&ary and the sur- 
face (Sectmn 2) In ad&tmn, it ts the nnplemen- 
tat~on of the content selection ANN that will deter- 
mine the eventual type and number of pragmatic 
features reqmred for the whole process of summan- 
eatmn (Sectmn 4) As a result, a partial analysts 
and interpretation of the input text only need to be 
performed m COSY-MATS The common problem m 
NLU-based systems of combmatozaal explosmn and 
mefllcment computatmn m the search for a solution 
will thus be largely avoided At the same tnne, thts 
pragmatmm m the analysts and interpretation pro- 
cesses does not decrease the amount of deep process- 
tug (semant!c, chscourse and pragmatic) camed out 
m the system High-level processing ts sahent m the 
pragmatic featuresldenttfied These are~ nonethe- 
less, 'grounded' by means of the generic lower-level 
features, as well as other surface and semantic char- 
actenstlcs of texts pertaining to the specL~C apph- 
cation of interest 
In surnm.~lT, the proposed arclntecture ts both 
modular and hybrid The complex task of content 
selectmn ts systematically decomposed into much 
more manageable computations In ad&tlon, the. 
strong points of both symbohc and connectlon- 
tst processing are combined m a compiementary 
Way (cf (Axetoulah, 1996)) The symbohc anal- 
ysers can. work vnth structured data of arbitrary 
length laden w~th variables They also have power- 
ful symbol-matching faczhtms (as ts appropriate for 
lower-level text analysas) In contrast, the ANNS are 
able to deal wtth fuzzy and inexact proceasmg (as 
ts revolved m nnportance determination and rater- 
level feature mappings) (McClelland and Rumelhartl 
1986, Rumelhart and McClelland, 1986) 
4 Empirical Evidence 
As the first and most cructal step m unplement- 
mg COSY-MATS, a prototype of its content selectmn 
ANN was developed Tins ts a standard feed-forward 
back-propagation network (Rumelhatt et al, 1986) 
Tins ANN receives m&wdual text sentences from the 
text to be snmmansed, hand-coded 2 by means of the 
identified pragmatic features, and assagns to them 
degrees of maportance It has been a major assump- 
tion behind tins work that it m feature combmatzons 
rather than individual features that charactense sen- 
tence importance (Sectmns 1 & 2) An ANN learns 
such interactions naturally, wlnch m why the con- 
nectlomst paradigm was.adopted for the content se- 
lection task 
The training corpus conststed of 1,8801 sentences 
m total, taken from the real-world text collection 
described m Sectmn 2 1,100 of them are sentences 
largely out of thetr context, wtule the remmnmg 780 
sentences make up 29 full texts In contrast to the 
dwersaty of the former subcorpus, each of the lat- 
ter texts ts approxLmately 23 sentences long and was 
fully encoded The encoding was camed out by 5 m- 
chvlduals on the basas of the above-mentmned man- 
ual wlnch exemphfies the correlations between the 
surface and the more abstract features m the pro- 
posed scheme The manual was used m order to 
standarchse the encoding process as much as pos- 
sable, as well as to vahdate the proposed ways m 
wlnch the evaluatmn of the abstract pragmatic fea- 
tures can be objectified and fully automated later on 
m the completed system 
Experiments to date (cf (Aretoulah, 1996)) have 
demonstrated the superiority of the pragmattc fea- 
tures over input to the ANN from aLcross the three 
levels of abstraction (58 1% vs 56 1% success on av- 
erage, where 'success' coincides with agreement vnth 
the judgement made by the human encoder regard- 
mg the level of nnpo/'tance of the corresponding sen- 
tence) The snnultaneous use of control experiments 
wtth nomy data S has ensured the vah&ty of these 
results (50 1% success) In addttlon, the testmg on 
whole texts has prowded comparable results to those 
acqmred with molated sentences, namely 56 8% suc- 
cess on average, thts suggests that the pragmatic 
features are sufficiently abstract to capture tuerar- 
ch~cal and structural aspects of the corresponchng 
dmcourses 
The dlversaty of the corpus m terms of subject 
matter, text type and length provides sutBcaent ew- 
dence for the appropriateness of the pragmatic fea- 
2given that the remaining components of COSY-MATS 
have not been tmplemented as yet, 
8These used characterLstlcs of the text that should 
be n~elevant to the content selection task, such as 'The 
second word m the sentence ends m a vowel' 
78 
tures for the Ingh-level representatlon of texts from 
any domain or text type Moreover, the portabfl- 
ity Of these pragmatlc content selectlon features has 
also been partly proved wlth experiments on whole 
texts (AretoulaJa, 1996) These re&cared that only 
a small amount of retraining ~s reqmred for the ANN 
to deal wlth new text types, winch mvolves a hm- 
ited number of representatlve texts Thus, what is 
pre&cted to dL~er between text types is the relatlve 
influence of each of the identflled features m the final 
wmghtmg of the corresponding sentence 
4.1 Generating Draft Sllmmarles 
The 'draft' s11,nmanes that result after concatenat- 
ing the sentences of the input text that were selected 
by the ANN as Important are, on the whole, adequate 
for current awareness purposes (See (Aretoula\]a, to 
appear) for a detailed evuluatmn of tins and other 
draft output) The ANN recelves a single --coherent 
and largely cohemve--- text each tIme, rather than a 
collection of unrelated texts Sentence selection was 
based on the 24 pragmatic features used for their en- 
coding and the statmt~cal correlatlous among them, 
as mchcated m the tratmng corpus Most Impor- 
tantly1 by faltering out the sentences for winch the 
AnN &d not have a clear dectslon, I e by adapting 
the corresponchng threshold on-/me, content selec- 
tmn can be more fiue-grarned and the output sum- 
manes more brief An example draft summary for a 
newspaper article after the apphcatlon of tiLtS type 
of fdtermg ~s shown below In tins case, there was 
8~ 6~ agreement between the ANN decision and the 
corresponding human judgement regarding the im- 
portance of m&v~dual sentences m thin article 4 
(I) Moscow e&tors fee\] the old-fashmned grip 
of the state (Headline) 
(~) Intense party pressure for the &enuseal of 
a prominent hberal e&tor and a new campmgn 
to d~sere&t the ra&cal pohtw~sn Bona Yeltsm 
- both apparently with the badang of Presi- 
dent Gorbachev - have rinsed fears among re- 
formers of a conservative swing by the Soviet 
leaderslup (5) On Monday evemug, he was 
summoned to the Central Comm,ttee to be 
told m so many words by Va&m Medvedev, 
the Pohtburo member m charge of ideology, 
that he should leave has post (6) The move 
follows a harsh talk dehvered last week by Mr 
Gorbachev to a group of semor Soviet e&tors, 
m which he gave several a dressing down (12) 
Some joumalmts are talking of a protest strike. 
(13) 'The press ~s qmte stmp ly now facing bans 
on what ~t can write about, we're going back 
4The 5 subjects were free as to the number of sen- 
tences they could p\]ek out from any text as unportant 
Importance, m turn, was defined as the relative m&s- 
pensab~hty from the final S,,rnm~ry of the proposmons 
expressed m the corresponding sentence Thts was deter- 
mined on the basra of the whole text the sentence belongs. 
to 
i 
to the situation of years ago,' one complained 
yesterday (16) The motion, which could pre- 
figure a head-on clash between the party and a 
steadily more assertive parhament, attacks the 
Central Comrmttee ldeol0gy department for fits 
'unacosptable attempts' to cow a newspaper 
(22) Ba~n$ for Mr Yeltsm zs not umversal 
(23) But the fact that the parhamentary ex- 
changes were broadcast on prime tune televl- 
szon leaves no doubt that a campmgn m un- 
der way to smear a man whose huge following 
makes hun Mr Gorbachev's only real rival 
Despite the coincidental coheslveness thereto, tins 
draft output comprises the majority of the seman- 
tically substantial sentences m the input text The 
concatenation of sentences from the original Is un- 
doubtedly a much simpler task than the generatmn 
of an extended summary or a concise abstract Novel 
text synthes~s m the fully-developed COSY-~/ATS wall 
also benefit from the proposed mappings between 
the surface and the more abstract content selectmn 
features Since the corresponding modules, however, 
have not been implemented yet, the processes re- 
volved wall not be exemphfied here 
5 Conclusion: COSY-MATS is not a 
Utopia 
All experimental results to date indicate that con- 
tent selection m the completed COSY=MATS environ- 
ment can be robust and efficient, even m the absence 
of any custonnsatlon to the spemfic apphcatlon (do- 
mare or text type) or the user reqmrements Tins m 
due to the adoption of the connectlon~t paradzgm 
for fins fuzzy task and the proven generic nature of 
the pragmatic and lower-level features used thereto 
In the context of further tmplementmg tins sum- 
mansatlon shell, current work mcludes the testing 
of ulternatlve.learnmg algolr/thms for the prototype 
content selectlon ANN m order to Improve ~ts success 
rate In addlt\]on, the more ngourous specflicatlon 
of the mappings between the surface cues and the. 
mtermechary and pragmatlc features is attempted 
for the subsequent development of speclahsed pro- 
cessors that compute them Thus, the encoding 
of the pragmatlc features will be fully automated 
and It will also be posslble to measure the pre- 
cme effect that tlns wfl/ have on the trmnlng of 
the whole cascade of ANNa, glven the current praco 
tlce of hand-coding Moreover, the impact on the 
content selectlon ANN Of incorporating apphcatlono 
dependent mformatlon m the system will also be 
stu&ed (cf (Aretoulakl, 1996)) What is nnportant 
Is that research to date has proved that the reahsa- 
tlon of the COSY-MATSmtelhgent and scalable sum- 
mansatzon shell m by no means a utopla 
79 
6 Acknowledgements 
The research reported m tins paper was earned out 
as part of a Ph D programme at the Centre for Com- 
putatzonal Lmgmst~cs at U M I S T I am indebted 
to my supexvasor, Prof Jun-zcha Tsujn, for hts m- 
valuable feedback and encouragement during that 
tame I am also grateful to CANON Europe Ltd for 
granting me with a 2-year Research Studentsinp, 
vnthout winch tins research would not have ever 
been possible Finally1 I would hke to thank the 
two anonymous rewewers for constructive comments 
winch have increased the degree of elanty of tins pa- 
per 

References 
M A~etoulak~ COSY-MATS A Hybrzd Connec- 
tsontst - Syrabohc Approach To The Pragmatic 
Analysts Of Tezts For Their Automatic Summarr- 
satzon PhD Thesis, Dept of Language Enga- 
neermg, U M I S T, Manchester, U K, March, 
1996 

M Azetoulaka A Hybrsd Connectwntst-Symbohc 
Approd~h To Pragrnatscs and Tezt Summartsa- 
twn Umverslty College London' (UCL), London, 
U K, To Appear 

J L Austin How To Do Thzngs Wsth Words Ox- 
ford Umverszty Press, Oxford, U K, 1962 
Science and Technology Feature Short Cuts The 
Economts~ pages 97-98, 1994 December 17th 

M Coulthard, edator Advances m Wrztten Tezt 
Analysts Routledgel London, U K, 1994 

B Endres-Nlggemeyer and E Neugebauer Pro- 
fessxonal Summansmg No Cogmtlve Szmulatmn 
wtthout Observatzon In Proceedings of the 4th 
Internatsonal Colloqmum on Cogmtsve Sczence 
(ICCS-g5), Donostza, San Sebastz~u, Spain, 1995 

J Fukumoto and J Tsujn Breaking Down Pdaetor- 
zcal ttelatlons for the Purpose of Analysing Dxs- 
course Structures In Proceedings of the tSth 
Interuatsonal Conference on Computatsonal Lm- 
gutstzcs (COLING-9~), volume 2, pages 1177- 
1183, Kyoto, Japan, August 1994 Assocaatzon for 
Computatmnal Lmgmstzcs 

R Ganghano, It G Morgan, and M H Snnth 
The LOLITA System as a Contents Scanmng Tool 
In Proceedings of the 13th Internatzonal Confer- 
. enee on Artsfic~al Intelhgence, Ezpert Systems and 
Natural Language Processing, A~guon, ~anee, 
May 1993 

M Gopmk Lmgutstsc Structures m Scientific Tezts 
Mouton, The Hague, the Netherlands, 1972 

B J Grosz The Itepreeentatmn and Use of Fo- 
cus m a System for Understanding Dml0gs In 
B J Grosz, K Sparck Jones, and B L Webber, 
edztors, Readzngs In Natural Langzuzge Process- 
mg Morgan Ksllfi'n~m~ Cahforma, 1986 Fzrst 
appeared m Proceedings of the 5th 13CAI, Cam- 
bridge, Massachusetts, 67-76, los Altos Wzlham 
Kaufmann, 1977 

M A K Halhday and R Hasan Cohesion m En- 
gltsh Longman, London, U K, 1976 

J R Hobbs Resolving Pronoun References Lin- 
gua, 44 311-338, 1978 

M Hoey Szgnallmg m Dxscourse A Panctlonal 
Analysm of a common Discourse pattern m writ- 
ten and spoken Enghsh In M Coulthard, e&torl 
Advances m Wrvtten Tezt Analysts, pages 26--45 
Routledge, london, U K, 1994 

E Hovy Generating Natural Language Under 
Pragmatzc Constraints Journal of Pragmatws, 
11 689--719, 1987 

P S Jacobs and L F Ran scIsoR Extracting In- 
fonnatzon from On-Line News Cornmumcatsons 
of the ACM, 33(11), pages 88--97, 1990 

M P Jordan The Rhetorsc of Everyday Enghsh 
Tezts George Allen & Unwm, London, U K, 
1984 

M P Jordan Openings m Very Formal Techmcal 
Texts Technostyle, 11(1) 1-28, 1993 

M P Jordan The Power of Negatzon Pragmat- 
zccs, Discourse Patterns and Clausal Semantzcs In 
Annual Meetzng of the Canadmn Assocmtzon of 
Teachers of Techmcal Wrdmg, Canachan Learned 
Conferences, Montreal, Canada, 1995 

J Kuplee, J Pedersen, and F Chen A Trainable 
Document Summarizer In Proceedings of the 18th 
ACM-8IGIR Conference, pages 68-731 1995 
W G Lehnert Plot Umts and Narrative Summ&- 
nzatlon Cogmtwe Sczence, 4, 1981, pages 293- 
331 

N Lucas, K Nmhma, T Alaba, and KG Suresh 
Dmcourse Analyszs of Scientific Textbooks m 
Japanese A Tool for Producing Automatic Sum- 
manes Techmcal Report 93TR-0004, Dept of 
Computer Science, Tokyo Instztute of Technology, 
Tokyo, Japan, March 1993 

H P Luhn The Automatze Creatmn of L~terature 
Abstracts ram Journal of Itesearch & Develop- 
ment, 2 (2), April, 1958, pages 159--165 

R E Mamell, J F Smith, and T E R Stager Ab- 
stracting Sczentzfie and Techmcal £s~rature An 
Introductory Grade and Tezt for Scsentis~, Ab- 
stractors, and Management Wfley-Intersclence 
John Wdey & Sons, Inc, New York, 1971 

W C Mann and S A Thompson Pdaetoncal Struc- 
ture Theory A Theory of Text Orgamzatmn 
Techmcal Report, ISZ Repnnt Series IS1/RS-87- 
190, usc Informatzon Sciences Instztute, Manna 
Del Rey, Ca, June 1987 

M T Maybury Automated Event Summ~za- 
tlon Techmques In Dagstuhl-Semmar-Reeport 79 
Sumrnartsmg Tezt for Intelhgent Commumeat~on, 
B Endres-Nlggemeyer, J Hobbs, and K Sparck- 
Jones, editors, Dagstuhl, Wadern, Germany, Dec 
13-17, 1993 

J L McClelland, D E Rumelhart, and the PDP 
Research Group (Eds) Parallel Dtstrsbuted Pro- 
ceasing Ezploratsons sn the Mtcrostrueture of 
Cognstson Volume 2 Psychologseal and Bsolog- 
seal Models The MIT Press, Cambridge, Mas- 
sachusetts, 1986 

R Mltkov, D Le Roux, and J-P Desclbs 
Knowledge-Based Automattc Abstracting Exper- 
nnents m the Sublanguage of Elementary Geome- 
try In C Martm-Vlde, editor, Current Issues m 
Mathematgcal gmgu~tzcs, pages 415-421 North- 
Holland, the Netherlands, 1994 

Proceedings of the Fifth Message Understanding 
Conference (MUC-5), Aug ~5-~7, San Mateo, CA, 
Baltnnore,'Maryland, 1993 Morgan Ka, fmann 

K Ono, K Smmta, andS Mnke Abstract Genera- 
tion based on Rhetorical Structure Extraction In 
Proccedmgs of the 15th lnternatwnal Conference 
on Computatsonal Lmgusstzes (COLING-94), vol- 
ume 1, pages 344-348, Kyoto, Japan, August 
1994 Association for Computatmnal Lmgtustics 
Also appeared as cmp-lg/9411023 

C D Pmce Automatic Generatmn of Literature 
Abstracts - An Approach Based on the Identzfics- 
tmn of Self-In&citing Phrases In R NOddy, 
S E Robertson, C J van RIjsbergen, and 
P W Wflhums, editors, Information Retrieval 
Research, pages 172-191 Butterworths, London, 
U K, 1981 

C D Pmce Constructing Literature Abstracts by 
Computer Techmques and Prospects In In/or- 
matwn Processing ~ Management, 26(1), pages 
171-186, 1990 - 

L F Ran, R Brandow, and K M~tze Domain- 
Independent S-mmarmat~on of News In Dagstuhl 
Seminar Report 79 Summartszng Tezt for In- 
telhgent Communzeatwn, B Endres-Niggemeyer, 
J Hobbs, and K Sparck-Jones, editors, Dagstuhl, 
Wadern, Germany, Dec 13-17, 1993 

R Re~chman Getting Computers to Talk Like You 
and Me lhscourse Contezt, Focus, and ~ernan- 
tics (An ATN Model) The MIT Press, Cambridge, 
Massachusetts, 1985 

D" E Rumelhaxt, G E Hmton, and R J Wflllam.~ 
Ch 8 Learumg Internal Representations by Error 
Propagation In D E Rumelhart, J L McClel- 
land, and the PDP Research Group, edttors, Par- 
alld Distributed Processing Ezplorat~ons m the 
M~crostrac~ure of Cogmt~on Volume 1 Founda- 
hous, pages 318--362 The MIT Press, Cambndge, 
Massachusetts, 1986 

D E Rumelhart, J L McClelland, and the PDP 
Research Group (Eds) Parallel Dtstrsbuted Pro- 
... ce ssm9 Ezplorahons m the Msc~ros~ructu~ of 
Cogmt~on Volume 1 Foundatson~ The MIT 
Press, Cambridge, Massachusetts, 1986 

G Salton, J Allan, C Bucldey, and A Smghal Au- 
tomatic Analysm, Theme Generatmn, and Sum- 
marLzatlon of Machine-Readable Texts 8csence, 
264 1421-1426, 1994 3 June 

R C Schank and R P Abelson 8cenpts, Plans, 
Goals, and. Understanding An Inquiry into Hu- 
man Knowledge Structures Erlbaum, Hdl~dale, 
N J, 1977 

J R Searle Speech Acts An Essay m the Phdos- 
ophyof Language Cambridge Umverslty Press, 
Cambridge, 1969 

B Sharp INFORMEX - An Informatmn Extractor 
System In Proceedings of the Internahonal Con- 
ference on Current Issues m Computahonal Lzn- 
gu~t~cs, pages 361-371, Penang, Malaysia, June 
1991 

C L Sldner Focusing m the Comprehensmn of Def- 
• mite Anaphora In M Brady and R Berwlck, ed- 
itors, Computa~onal Models of Discourse, pages 
267 = 330 The MIT Press, Cambridge, Mas- 
sachusetts, 1983 

S Soderland and W Lehnert Corpus-Driven 
Knowledge Acqmsmon for Discourse Analysm In 
Proceedings of the l~h Natsonal Conference on 
Artsf~al Intellsgence (AAAI-94), Vol 1, pages 
827-832, Seattle, Washington, Jul 31 - Aug 4, 
1994 

T A van Dijk and W Kmtsch Cognitive Psychol- 
ogy and Discourse Recalling and Summarmmg 
Stones In W U Dressier, editor, Current Trends 
m Te.zthngmst~cs, volume 2 of Research m Tezt 
Theory, pages 61-80 de Gruyter, Berhn, 1978 

B L Webber So, What Can We Talk About Now ? 
In M Brady and R Berwlck, editors, Computa- 
tional Models of Dtsconrse, pages 331-371 The 
MIT Press, Cambridge, Massachusetts, 1983 

Y Walks and A Balhm Multiple Agents and the 
Heuristic Asccnptzon of Behef In Proceedings Of 
the lOth International Joint Conference on Arts- 
final Intelligence (IJCAI-87), pages 119-124, Mi- 
lan, Italy, August 1987 
