Summarising as a lever 
for studying large-scale discourse structure 
Ka.ren Sparck Jones 
Computer Labora.tory, University of Cambridg(' 
New Museums Site, Pembroke Street, Cambridge CB2 3Q(I, IlK 
sparckjones(~uk.a.c.cam.cl 
June 1993 
We are using summarising as a way of studying large-scale discours¢~ struc- 
ture. Much computationally-orieuted work on disconrse structure ha.s been con- 
cerned with dialogue, rather than with 'single-source' text.. Some prop()sals haw. 
been made for singh~-soul'ce text e.g. l~hetorical Structure Theory (Mann ;rod 
Tholnpsoll 1987), I)nt al'~' open to criticism (e.g. Moore and I)olla(:k 1.q92); and 
single- source work has been primarily concerned with generation ((,.g. McK~,- 
own 1985, Maybury 1991). We believc that large-scale discourse stru('tm'\[~ ha.s 
a crucial part to play in SUlnmarising and therefore needs to be captured in the 
source text representation, for use in snmmarising, regardless of its contril)ution 
to source interpretation itself. 
We have been engaged in a systematic examination of Mternative types of 
large-scale text structure, designed to throw light on the kinds of inlbrma.tion 
they make available for the text above the level of individual sentence rel)r~'- 
sentations, and how these call be used in sumlnarising. Thus source, text in- 
terpretation will provide a source representation capturing discourse structur(" 
over sentences, to be exploited in a condensing transformation through which 
the summary representation is formed, in turn leading to the output smmnary 
text. 
This is a deliberately analytical investigation, taking a broa.d view with- 
out preconceptions. We distinguish three types of discourse information with 
structural implications: linguistic, dolnain, and Colmnunicativ(~, and a.r(' s,~eing 
what large-scale text. structures these respectively give. Thus we at'(' inv(~stigat- 
ing representation types categorised as dealing with informatioll either about 
the linguistic properties of the source text (e.g. parallelisln), or about its do- 
main content (e.g. class lnelnbership), or abont its COlmnunicative fimction 
(e.g. counterclaim). We are fill'ther, for any of these types, corlsidering two 
alternative forms of structnl'e that we have labelled 'bottom-up' and 'top-down' 
respectively. Bottom-up structures are individually created using g(,n('ra.I ruh's 
125 
(e.g. by inference from domain facts); top-down structures are obtained by ill- 
stantiating prior proformas (e.g. using domain frames). This is not a processing 
distinction, and the same formal structure (e.g. hierarchical) may result ill ei- 
ther case; there may also be intermediate possibilities of the 'grammar' type. 
These distinctions of information type and representation form are broad ones 
that we are using as heuristics to explore discourse structure. Our aim is a com- 
paratiw~ one, to see what each kind of approach leads to both for representation 
and for summarising. We can then consider how the structures relate to one 
another, whether as dependent, complementary, or reinforcing ones. 
We are as far as possible using 'exemplar' approaches taken from previous 
research in the field, primarily in order to ground our work in what has been 
done so far: we are obliged in the current state of the art to work prirnarily 
through simulation, but we are trying to constrain the resea.rch by folk)wing 
approaches already proposed in the literature and preferably computatioually 
investigated. Thus as an experimental strategy we are taking logical t'oH,ls 
with resolved anaphors as a baseline representation for sentences, a.nd th(.n 
applying exemplar strategies of each type to these to obtain \['ull rel)resentati(,ns 
of the source text.. These full representations capture further relations ~('ross 
the sentences, embodying the large scale source text structure. 
We have obtained alternative discourse structures and summaries for a set. 
of short test texts. Some of the source structures are very simple, others more 
complex, importing significant additional information. So far, we have used the 
source representations in natural ways to obtain summaries: thus a linguistic- 
type source representation leads to a linguistically-motivated summary repre- 
sentation, in a way appropriate to the kind of the linguistic representation. 
As linguistic structures we have so far provided analyses and derived sum- 
maries from the most simple approach, exploiting focus history to pick out key 
discourse entities, to more elaborate ones provided by Rs'r (taking rhetorical 
relations as linguistic). These are bottom-up forms: rhetorical schemata might 
suggest a complementary top-down approach, but we could not readily anal- 
yse our texts as instantiations of these, and we therefore tried an intermediate 
'story (or text) grammar' approach (cf Rumelhart 1975) To obtain domain- 
based structures we have used an extremely simple bottom-u I) al)proach using 
predication participation to identify discourse entities which figure largely in the 
source: we" would like, to try more sophisticated strategies where the bas('lin~' 
representation is enriched using general inference rules. We have applied scripts 
(and frames) as a top-down representation form (cf DeJong 1979; Tait 1983). 
Finally, for communicative structure we have used Grosz and Sidner (1986)'s 
approach to get intentional representations for our test texts. This constitutes 
a bottom-up approach: we have not yet identified an exemplar top-down one. 
The results we haaze obtained have provided stinmlat.ing insights into the 
properties and roles of different types of text structure, and into the respect.iv(, 
contributions they may make to summarising. For summarisiug, all the large- 
scale structures provide good leverage and help to identify source material which 
126 
is intuitively important for use in the condensed summary, through selection 
or generalisation, though the alternative results for the same text may differ 
noticeably and individual results may be only senti-satisfactory. The results 
also illustrate the genuine role, but incomplete contribution, of each type of 
information. 
Our deliherate separation of information types with their application strate- 
gies is thus allowing us to examine each type; to see how large-scale structure 
of any one kind is related to local structure, for instance through focus; and 
to formulate a view of a discourse model as a whole which subsumes distinct 
contributing models with their own necessary functions. Thus for example for 
one text, 'Biographies', there is a linguistic structure showing heavy presenta- 
tional parallelism, a simple sequence of persuasive communicatiw-" intentions, 
and a separate domain object categorisation. There are complex rela.tions Iw- 
tween these, with reinforcing effects on the indication of key cont~'nt. Our 
comparative analyses are thus providing the base (Grosz and Sparck Jones, in 
preparation), for the development of an account of discourse structure, or a dis- 
course model, as a higher-level structure over subsidiary structures each with 
their own character and role. 

References
P. Gladwin, S. Pulman and K. Sparck Jones 'Shallow processing and aut- 
matic summarising: a first study', TR 223, Computer Laboratory, Uniw~rsity of 
Cambridge, 1991. 

K. Sparck Jones 'Discourse modelling for automatic sunamarising', '1'1:~ 290, 
Computer Laboratory, University of Cambridge, 1993, and ill press. 

K. Sparck Jones 'What might be in a smmnary?', Proceedings of the" (;erma.n 
hfformation Retrieval Conference, 1993, in press. 
