NARRATIVE SUPPORT FOR TECHNICAL DOCUMENTS
Formalising Rhetorical Structure Theory
Nishadi De Silva, Peter Henderson
School of Electronics & Computer Science, University of Southampton, Southampton, UK
Keywords: Document narratives, Technical documents for BPR, Rhetorical Structure Theory, XML
Abstract: Business Process Re-engineering (BPR) is an area tha
t requires a lot of technical documents and an
important feature of a well-written document is a coherent narrative. Even though computer software has
helped authors in many other aspects of writing, support for document narratives is almost non-existent.
Therefore, we introduce CANS (Computer-Aided Narrative Support), a tool that uses Rhetorical Structure
Theory to enhance the narrative of a document. From this narrative, the tool generates questions to prompt
the author for the content of the document. CANS also allows the author to explore alternative narratives for
a document. A catalogue of predefined narrative structures for popular types of documents is provided too.
Our tool is still in its rudimentary stages but sufficiently complete to be demonstrated.
1 INTRODUCTION
Written communication is an integral part in many
fields of work and study. BPR is an area in particular
that requires a lot of technical documents.
A fundamental aspect of a document is the
‘sto
ry’ it conveys to the reader. This is referred to as
a document’s narrative. A coherent, well-structured
narrative will convey the information better and be
more convincing. With the advancement of
technology, software support for the writing process
has been manifold. However, computer support for
document narratives is almost non-existent.
There are many theories for the structure of a
narrative.
We have studied Rhetorical Structure
Theory (RST) (Mann & Thompson, 1988) to build a
tool which will help authors construct a document
with a more coherent, convincing narrative.
CANS (Computer-Aided Narrative Support)
al
lows the author to build, modify and create
instances of a narrative for a document. The tool
uses this narrative to generate a set of questions that
prompts the author for the document’s content. More
importantly, CANS also enables the author to
explore alternative narrative structures for an
important technical document.
This paper describes RST, introduces our tool
and
discusses further enhancements. We have also
looked at other tools that aid the writing process and
a brief overview of these are given in section 4.
2 OVERVIEW OF NARRATIVE
THEORIES
Studies into narratives have existed for over a
century. Many narratologists have identified
structures that are optimal for specific genres of
writing (e.g. Propp 1928). For instance, as early as
1863, the German journalist and writer, Gustav
Freytag, introduced a five-part pyramidal structure
which he believed to be the most successful format
for a play (Freytag 1863). Even formats for technical
documents have often been defined (e.g. Paradis and
Zimmerman, 2002). However, defined formats alone
do not complete a document. It is important to
construct a coherent narrative too.
Many researchers have studied the coherence of
narratives in
general and with each theory, new
notations and understandings of narratives have
emerged (Lehnert 1981, Grosz & Sidner 1986,
Grosz et al. 1995). Among them is the Rhetorical
Structure Theory (RST) (Mann & Thompson, 1988).
RST is one of the most popular discourse theories of
the last decade (Marcu 2000) and is explained in
more detail below.
2.1 Rhetorical Structure Theory
This theory uses relationships between segments of
text to explicate the coherence of a narrative.
105
De Silva N. and Henderson P. (2005).
NARRATIVE SUPPORT FOR TECHNICAL DOCUMENTS - Formalising Rhetorical Str ucture Theory.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 105-110
DOI: 10.5220/0002533501050110
Copyright
c
 SciTePress

1
2
3
4
5
In RST, a text segment assumes one of two roles
in a relationship: the nucleus (N) or satellite (S).
Nuclei express what is more essential to the
understanding of the narrative than the satellites.
The size of a text segment is arbitrary but each
should have independent functional integrity.
Relations hold between non-overlapping text
segments and are of two kinds: hypotactic and
paratactic. Hypotactic relations connect one nucleus
and one satellite. Paratactic relations hold between
text segments of equal importance; that is, multiple
nuclei. There are 23 relations defined in Mann &
Thompson’s original paper. Two of them are
illustrated below. In these diagrams, the arrow
always points towards the nucleus in a hypotactic
relationship.
N N
1-3
SequenceSequence
N
Sequence
Figure 1: A paratactic relationship
N S
Motivation
Figure 2: A hypotactic relationship
Text coherence arises due to an overall effect
associated with each relation. For instance, in a
MOTIVATION relation, the satellite presents some
information that increases the reader’s desire to
perform the action presented in the nucleus.
Generally, a relation is not expected to dictate
the order of the text spans. However, after analysing
many texts, Mann & Thompson (1988) identified
patterns for the order of the nucleus and satellite for
some relations (reproduced below).
Table 1: Order of
text spans for some relations (Mann & Thompson, 1988)
Satellite before Nucleus
Antithesis Conditional
Background Justify
Concessive Solutionhood
Nucleus before Satellite
Elaboration Purpose
Enablement Restatement
Evidence
In order to illustrate how we apply RST to our
work and to explain the theory further, we produce
the narrative below for a very simple story.
[There is an initial condition.]
1
[Then a problem arises]
2
[that disrupts this condition.]
3
[A solution is sought.
One of the solutions fixes the problem]
4
[and restores the
initial condition.]
5
For a coherent narrative, RST is expected to
produce a tree of relations. It is possible to have
several valid RST trees for a narrative. One possible
RST tree for the narrative above is given below. A
more traditional tree diagram also appears on the
right with the RST relations superimposed in red.
<hypRelation id="A" type="Volitional-
result">
<satellite id="3" />
<nucleus id="2" />
</hypRelation>
<hypRelation id="B" type="Background">
<satellite id="1" />
<nucleus id="2" />
</hypRelation>
1-3
Then a problem
arises
There is an initial
condition.
Background
that disrupts this
condition.
Volitional-result
A
solution is sought.
One of the solutions
fixes the problem
4-5
Solutionhood
and restores the
initial condition.
Motivation
Figure 3: A possible RST tree for the narrative for a simple story (left). A more traditional tree view (right).
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
106

It is possible to narrate the same story in several
different ways. An alternative narrative is given
below (produced by visiting the nucleus first in
every relationship of the tree in Figure 3).
Fido’s owner took him to the vet.
The vet recommended a flea treatment which got rid of Fido’s
fleas.
Then Fido stopped scratching and was happy again!
Last week Fido got fleas and started scratching.
Fido is usually a happy dog but the scratching made Fido
unhappy.
3 CANS (Computer-Aided Narrative
Support)
We use RST to help the author enhance the
document narrative. After the narrative is created,
CANS generates a sequence of questions that
prompts the user for the document’s content. An
author can also investigate alternative narratives that
better suit the document. These features are
elaborated in the following sections. Our tool is still
rudimentary and is very much a work in progress.
CANS is implemented using JSP (Hall, M. &
Brown, L., 2004) and XSLT (Kay, M., 2002).
Central to this tool is an XML database. The user
interface is in HTML.
3.1 Creating the narrative structure
The writing process begins by constructing a
narrative for the document and producing a RST tree
for it. This can be done by typing the narrative,
breaking it into segments and defining the relations
between these segments. By defining these relations,
the existence of each text segment is justified and it
is easy to identify segments that are unnecessary or
out of place.
This functionality is successfully provided by the
free software tool, RSTTool (O’Donnell, 2000).
RSTTool has also been used to produce the
diagrams in this paper. We might consider
incorporating this tool as part of our work. RSTTool,
however, produces .rs3 files which are also in an
XML format, but different to URML. We are
currently working on an XSL stylesheet that can
transform this format to URML.
To demonstrate how our tool can be used by a
technical author, we present the narrative below. It
was created to represent the typical ‘story’ of a
Research Proposal. The italicized phrases are
expected to expand to a section in the Research
Proposal and are used in section 3.3 to discuss
alternative narratives. We have drawn a RST tree for
this narrative and a collapsed version of it is
illustrated in Figure 4.
The narrative structures thus created are stored using URML (see section 2.1) in the XML database.
[We want you to fund us]
1
[because we will achieve these objectives/results.]
2
[We believe these
results are important to you]
3
[because of benefits-to-beneficiaries]
4
[and to the whole world]
5
[because there exists an unsolved-problem.]
6
[We know this is unsolved]
7
[because we have
studied the background.]
8
[We will solve this problem]
9
[by this method.]
10
[We know this is the
best method]
11
[because we have studied alternative-methods.]
12
[To achieve this, we will need
total-time]
13
[and these resources]
14
[because justification-of-resources.]
15
[The research will be
carried out by these researchers]
16
[and they are the most qualified to do this because
justification-of-researchers.]
17
[The research will be conducted at these locations]
18
[because
justification-of-locations.]
19
NARRATIVE SUPPORT FOR TECHNICAL DOCUMENTS: Formalising Rhetorical Structure Theory
107

We want you to fund
us
2-19
Motivation
because we will
achieve these
objectives/ result
s
3-19
Evidence
9-12
10-12
Evidence
by this method 11-12
Elaboration
16-19
ElaborationSolutionhood
13-15
Condition
We will need
total-time
Sequence
14-15
Sequence
We will solve this
problem
3-8
Figure 4: A RST tree for the Research Proposal narrative (collapsed version)
3.2 Generating the questions from the
narrative structure
During the second stage of the writing process, the
user can select a narrative from a list, along with a
mode of traversing the RST tree for this narrative
(explained in section 3.3).
At the moment, the questions are relatively
simple; there is a question generated for every
segment in the narrative. We hope to improve this in
the future. Preceding the question is a history of its
relations to other segments, so that the author can
realise how the content in the answer integrates with
the rest of the document. For instance:
(Motivation:: We want you to fund us)
What are the OBJECTIVES/RESULTS?
The user can type the answers in HTML text
areas and save the content in the XML database.
Later on, other narrative structures can be applied to
this same content to transform it to different
documents.
3.3 Exploring alternative narratives
The narrative of a technical document often needs to
be altered to suit the reader.
For example, the narrative of a proposal pitched to
an audience of investors needs an explanation of
how the technical plan achieves something that
others cannot. The story should convince the
investors that the customers will be willing to pay
for it.
Such a proposal should contain a clear definition of
costs and time requirements, along with evidence to
show that the research team is capable of using the
investors’ money wisely. In contrast, a proposal read
by other researchers in the field, should enhance the
understanding of the unsolved problem and the
chosen method of solution (Paradis & Zimmerman,
2002).
Narrative 1 Narrative 2
Objectives/Results Objectives/Results
Benefits-to-beneficiaries Methods
Background Alternative-methods
Unsolved-problem Total-time
Total-time Resources
Justification-of-resources Justification-of-resources
Resources Researchers
Methods Justification-of-researchers
Alternative-methods Locations
Justification-of-researchers Justification-of-locations
Researchers Benefits-to-beneficiaries
Justification-of-locations Unsolved-problem
Locations Background
Figure 5: Outline of narratives from traversal method 1 (left) and traversal method 2 (right)
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
108

Alternative narratives are produced by traversing
the RST tree in different ways. For now, there are
two traversal methods, each producing a different
sequence of questions for the user. The first method
visits the nucleus and satellite in an order dictated by
the name of the relationship (see Table 1). The
second method always processes the nucleus before
the satellite for every relationship. To make the
traversal easier, the RST tree in Figure 4 was
converted to a binary tree. Figure 5 shows the
outlines of the narratives produced by each method
using just the italicized phrases in the Research
Proposal narrative.
More traversal methods will be investigated.
3.4 Viewing the narrative structure
While typing the answers to the questions, the user
has the option to view the current narrative structure
in either a tree format or as a textual narrative.
3.5 Predefined narrative structures
There is a list of predefined narrative structures for
popular types of documents provided by the tool.
This list is expected to grow as more research is
done into document narratives. For now, we hope to
remain within the domain of technical writing.
4 RELATED WORK
In this section we briefly describe a few existing
tools that help authors with writing and list some of
their features so as to differentiate them from our
work.
a) New Novelist software
This software (purchased from
www.amazon.co.uk) helps a novice write a
novel in 12 steps. The user is asked to choose
the genre of the novel, define characters, add
attributes to these characters and fill in
templates for the content. Each genre has a
fixed sequence of sections that fits most novels
in that genre, along with the optimum number
of pages for each section.
b) ActiveDocs Document creation
Active Docs provides templates for the
automatic creation of documents such as Sales
Proposals and Lease Agreements by prompting
the user for essential information. It has an
HTML interface and supports many popular
document formats (ActiveDocs, Document
Automation Solutions).
c) WiCKEd
This is a prototype tool to assist document
authoring in the Semantic Web context
(Woukeu, et al. 2004). As an example, they
present the process of writing a research
proposal. While the user types in the provided
text editor, the tool continuously analyses this
text to recognise known words. These words are
then used to find relevant information for the
proposal on the intranet.
d) Several tools exist that detect RST relations in a
given text (Mahmud 2004) and few others make
use of RST to enhance the quality of the
produced text. For instance, Rizzo et al. (2002)
describe a tool that uses RST to produce
rhetorically-structured digital puppet
presentations.
e) ArtEquAkt
This tool (
Kim S., et al. 2002) uses knowledge
acquisition and analysis techniques to extract
information from web pages on a given subject
domain and creates a knowledge base overlaid
with an ontology. The ontology can then be
used to construct stories by using story
templates.
5 CONCLUSIONS AND FUTURE
WORK
CANS is still in need of many improvements to its
user interface and functionality. Several specific
improvements are discussed in this section.
A prominent feature of this tool is the ability to
explore different narratives for a document.
However, as illustrated by the two simple stories
about Fido in section 2.1, a change in the narrative
structure requires a change in the words of the
sentences. We hope to improve our tool, in a way
less pedantic than Natural Language Processing, to
mimic this alteration of words so that the alternative
narratives remain coherent.
Other traversal methods of a RST tree will be
researched along with ways of producing different
RST trees for the same narrative. We can get some
useful ideas from Marcu (2000) about exploring all
valid RST trees for a given text. A further
enhancement would be to allow the combination of
NARRATIVE SUPPORT FOR TECHNICAL DOCUMENTS: Formalising Rhetorical Structure Theory
109

RST trees so that several narratives could be merged
into one document.
Currently the XML database is maintained using
the Java API for XML processing. We have studied
Xindice as an alternative (Apache Xindice, 2001)
and hope to start using it soon. We are also
considering other XML formats that can be used to
store the narrative structures instead of URML.
We will also implement the ability to define new
relations, apart from those specified by RST.
Most deliverables in a technical environment are
in the form of various kinds of factual genres. The
challenge in our work is to understand narrative
forms and then to transform them into professionally
acceptable technical documents. We believe this tool
is useful because it encourages an organisation of
thought and structure which is considered essential
for good writing. Our studies show that this feature
is absent in most other writing tools. In particular,
we hope that the ability to explore alternative,
coherent narratives for a document will be helpful
for technical authors in BPR.
REFERENCES
ActiveDocs, Document Automation Solutions. (n.d.)
Retrieved June 8, 2004, from
http://www.activedocs.com/
Apache Xindice. 2001. Retrieved November 11, 2004
from http://xml.apache.org/xindice/
Freytag, G., 1863. Freytag’s technique of the drama.
Benjamin Blom. New York and London, translated
from the 6
th
German edition by Ellias J. MacEwan in
1968.
Grosz, B. & Sidner, C., 1986. Attention, intentions, and
the structure of discourse. Computational Linguistics,
12, 3, 175-204.
Grosz, B., Joshi, A. & Weinstein, S., 1995. Centering: A
Framework for Modelling the Local Coherence of
Discourse. Computational Linguistics. 21, 2, 203-225.
Hall, M. & Brown, L., 2004. Core Servlets and
JavaServer Pages. Prentice Hall. USA. 2
nd
edition.
Kay, M. 2002. XSLT. Wrox Press. Canada. 2
nd
edition.
Kim, S., et al., 2002. Artequakt: Generating Tailored
Biographies from Automatically Annotated Fragments
from the Web. In Proceedings of Workshop on
Semantic Authoring, Annotation & Knowledge
Markup (SAAKM’02), pp: 1-6, Lyon, France.
Lehnert, W., 1981. Plot Units: A Narrative Summarization
Strategy. In Strategies for Natural Language
Processing, 375-412, edited by Lehnert & M. Ringle
in 1982. New Jersey: Lawrence Erlbaum Associates.
Mahmud, R., 2004. Revealing Discourse Relations
Structure: an Approach for a Dynamic Computer
Aided Writing. Computers and Writing conference
2004. Hawaii.
Mann, W. & Thompson, S., 1988. Rhetorical Structure
Theory: Toward a functional theory of text
organisation. Text, 8:3:243-281
Marcu, D. 2000. The Theory and Practice of Discourse
Parsing and Summarization. The MIT Press.
O’Donnell, M., 2000. RSTTool 2.4 – A markup tool for
Rhetorical Structure Theory. In Proceedings of
International Natural Language Generation
Conference (INLG’2000), 253-256, Mitzpe Ramon,
Israel.
Paradis, J. and Zimmerman, M., 2002. The MIT Guide to
Science and Engineering Communication. The MIT
Press. 2
nd
Edition.
Propp, V., 1928. Morphology of the Folktale, (pp:25-65),
University of Texas Press. Austin, 2
nd
edition.
Reitter, D. & Stede, M. 2003. Step by step: underspecified
markup in incremental rhetorical analysis. In
Proceedings of the 4th International Workshop on
Linguistically Interpreted Corpora (LINC-03),
Budapest.
Rizzo, P., et al. 2002. An Agent That Helps Children to
Author Rhetorically-Structured Digital Puppet
Presentations. In Proceedings of the 6th International
Conference on Intelligent Tutoring Systems, pp:903-
912.
Woukeu, A., Carr, L. and Hall, W. 2004. WiCKEd: A
Tool for Writing in the Context of Knowledge. In
Proceedings of Hypertext 2004 - Fifteenth ACM
Conference on Hypertext and Hypermedia (in press),
University of California, Santa Cruz, USA.
ICEIS 2005 - SOFTWARE AGENTS AND INTERNET COMPUTING
110