An Algorithm for High-Level Organization 
of Multi-Paragraph Texts 
Robert Granville 
rgranvilOBBN. COM 
Our experience with MACH-III \[Kurlaud et al 1992\] showed us that there is more to multiparagraph 
text than stringing together isolated well-formed paragraphs. The underlying structure of tile entire 
text, depicting interparagraph relationships and emphases, must also be determined for suceessfill 
generation. Fortunately, RST \[Mann-Thoml)son 1987\] is capable of representing interl)aragraph 
structure ms well ms intraparagraph structure, tlowever, RST does not specify how to build large 
structures representing multiparagraph text. This paper presents an algorithm to construct such 
nmltiparagraph structures representing a critique of a student's performance in troubleshooting the 
HAWK radar, as determined by MACH-III. This critique is based on the functional hierarchy tree 
(FH tree), which is the heaxt of the expert system component of MACH-III \[Kurland et al 1989\]. 
Each student action is judged based on the structure of the FH tree, and where the student is 
currently located in that tree. The generated critique thus needs to describe each action and how 
that action helps or hinders proper navigation through the FH tree. 
Functional hierarchy is a new paradigm for organizing expert system knowledge bases, based on the 
procedural abstraction principles of Liskov and Gutag \[Liskov-Gutag 1986\]. Functional hierarchy 
differs greatly from production rules (the customary basis for an expert system) in that functional 
hierarchy rules define the actions a system can take, rather than the conditions under which actions 
may take place. The concept of "action" is expanded to include all actions the system takes, 
including control decisions, rather than just changes to the database, thereby eliminating the need 
for a separate control structure. These rules are arranged in a hierarchy, where the action of a rule 
is defined as a combination of other actions. 
Our algorithm does not need an elaborate component to build a plan resulting in an RST for two 
reasons. The first is that the desired output text is a non-interactive monologue of written text, 
rather than text that models interactive dialogue, as in the text of Moore's system \[Moore 1989\]. 
Therefore, we don't need information for revising text in reaction to a listener's misunderstanding. 
The second is that the database from which we are generating consists of the MACH-III functional 
hierarchy trees. These FH trees are structured purposely to explicitly reflect the very organization 
we need to explain. Because of these FH trees, we don't have to build plans to determine text 
structure, and the job of organizing the text, that is, building RST structures is greatly simplilied. 
As stated above, the purpose of our generated text is describe student actions in the course of a 
radar troubleshooting session, and how these actions relate to the organization in the FH trees. In 
themselves, these actions are completely independent. The way RST organizes events that have no 
19 
relationship between them other than tile ordcr in which they occurred is with the SEQUENCE 
relationship. 
This leaves tile problem of tile higher organization of tile text, that is, where to break paragraphs in 
the lIST. It was argued in \[Granville 1990\] that the structural organization of a text is as important 
to the message to be conveyed as its factual content. Therefore, any artificial metric for paragraphs, 
such as limiting them to a specific number of sentences, must be unsatisfactory, flowever, the 
problem is not so daunting when we consider the purpose of a paragraph, which is to describe 
one idea or topic. The obvious topic category for our text consists of events in the course of 
troubleshooting that require explanation. The two events in the MACII-III domain that require 
explanation are mistakes committed by the student, which are mistakes in navigating through the 
FII trees, and milestones telling the student when branches in the FtI trees are entered or completed. 
The mistake or milestone being explained is obviously the topic of the explanation, and therefore 
deserves its own paragraph. Mistakes and milestones that have already been explained don't require 
full explanations again, and therefore don't merit paragraph treatment. 
This does not completely solve the problem, however. Depending on the FH tree and the student 
actions, we may have a set of actions requiring no detailed explanation but is still too large to describe 
in a single paragraph. This problem can be solved if we rememl~er that the ultimate goal of our text 
is to explain how actions should reflect, the FI1 tree organization of the troubleshooting knowledge. 
By causing paragraph hreaks when new FII tree I.~ranehes are enl.ered, our text explicitly reflects 
the FH tree organization. Another problem is that we may want to avoid a para.graph break after 
a detailed explanation, if the next text item is closely related, such as a single student action that 
is an error needing an explanation which also happens to start a new FII branch. Therefore, if two 
consecutive items are closely related, such as due to arising from the same action, or a repetition 
of an action, they will appear in the same paragraph, even though the first item would cause a 
paragraph break under normal circumstances. 
A brief example will help make these ideas concrete. Assume a MACH-III student is presented 
with an LO fault, indicating that the local oscillator (the LO) is either malfunctioning or not being 
monitored correctly. Our student performs four actions, each with a mistake. He first tests the W2 
cable, a component of the IIAWK radar that can't cause an LO fault even if the cable itself is faulty. 
In MACH-III, this is known as an OUT OF 'FREE mistake, because since the W2 can't cause an 
1,O fault, it doesn't appear in the subtree of the functional hierarchy that deals with LO faults. 
Our student next replaces the All card, a component that could cause the fault. Ilowever, he fails 
to follow this action with a test of the system (known as a BITE test) to see if the A 11 was in fact 
the ca.use of the problem. MACII-III ca.lls this an EXPECTEI) BITE TEST mistake. 
The student's third action is to test the W8 cable. Like the W2, the W8 cannot possibly cause an 
LO fault, so this is again an OUT OF TREE mistake. 
Finally, the student tests the P12J3 plug. This component can cause an LO fault, so it isn't an 
OUT OF TREE problem. However, testing the All card is part of determining whether the local 
oscillator is being monitored correctly, while testing the P12J3 is part of determining whether the 
local oscillator itself is malfunctioning. As stated above, one of the goals of MACtI-III is to teach 
students the organization exhibited in the functional hierarchy tree. Therefore, jumping around 
from topic to topic, as our student is doing here, is flagged as a SKIPPING fault. 
When the first action is encountered, the OUT OF TP~EE mistake hasn't been explained yet, so the 
20 
action of testing the W2 and the OUT OF TREE problem warrant their own paragraph detailing 
why OUT OF TREE actions are mistakes. Similarly, replacing the All card requires a separate 
paragraph to explain the EXPECTED BITE TEST problem. The third action, testing the W8, 
does not force a paragraph break on its own, since it exhibits the OUT OF TREE mistake, which 
has already been explained. Instead, it merely mentions the problem, and the paragraph continues. 
The final action of testing the P12J3 does exhibit a new problem, though, namely SKIPPING. It is 
added to the third paragraph in a sequence relationship, and this paragraph is closed off because of 
the required explanation. The resulting RST looks like the following: 
TEXT 
J 
I PARAGRAPH I SEQUENCE I 
.................. i 
EXPLAIN EXPLAIN --- 
OUT-OF-TREE EXPECTED-BITE-TEST I 
W2 All 
{ SEQUENCE { 
MENTION EXPLAIN 
OUT-OF-TREE SKIPPING 
W8 PIRJ3 

References
\[Granville 1990\] Granville, Robert A., "The Role of Underlying Structure ill 'l~×t Generation," ill 
Proceedings of the Fifth International Workshop on Natural Language Generation, 1990 

\[Kurland et al 1989\] Kurland, Laura C., Robert Granville, and Dawn MacLaughlin, "HAWK MACH- 
Ill Explanations of the Receiver "lh'oubleshootiag Tree," Techuical Report, BBN Systcms and Tech- 
uologies, Cambridge, Massachusetts, 1989 

\[Kurland et al 1992\] Kurland, Laura C., Robert Alan Granville, and Dawn M. MacLaughliu, "De- 
sign, Development and Implementation of an Intelligent Tutoring System (ITS) for Training Radar 
Mechanics to Troubleshoot," ia "lutelligeut Instruction by Computer," edited by Marshall J. Farr 
and Joseph Psotka, Taylor &; Francis, Washington, DC, 1992 

\[Liskov-Guttag 1986\] Liskov, Barbara, and John Guttag, Abstraction and Specification in Program 
Development, MIT Press, Cambridge, Massachusetts, 1986 

\[Mann-Thompson 1987\] Mmm, William C., and Sandra A. Thompson, "Rhetorical Structure Theory: 
A Theory of Text Organization," ISI/RS-87-190, Information Sciences Institute, Marina del Rey, 
California, 1987 

\[Moore 1989\] Moore, Johanna Doris, "A Reactive Approach to Explanation in Expert and Advice- 
Giving Systems," Ph.D. thesis, University of California, Los Angeles, 1989 
