Integrating Text Plans for Conciseness and Coherence* 
Terrence Harvey and Sandra Carberry 
Department of Computer Science 
University of Delaware 
Newark, DE 19716 
{harvey, carberry}@cis.udel.edu 
Abstract 
Our experience with a critiquing system shows 
that when the system detects problems with the 
user's performance, multiple critiques are often 
produced. Analysis of a corpus of actual cri- 
tiques revealed that even though each individ- 
ual critique is concise and coherent, the set of 
critiques as a whole may exhibit several prob- 
lems that detract from conciseness and coher- 
ence, and consequently assimilation. Thus a 
text planner was needed that could integrate the 
text plans for individual communicative goals to 
produce an overall text plan representing a con- 
cise, coherent message. 
This paper presents our general rule-based 
system for accomplishing this task. The sys- 
tem takes as input a set of individual text plans 
represented as RST-style trees, and produces 
a smaller set of more complex trees represent- 
ing integrated messages that still achieve the 
multiple communicative goals of the individual 
text plans. Domain-independent rules are used 
to capture strategies across domains, while the 
facility for addition of domain-dependent rules 
enables the system to be tuned to the require- 
ments of a particular domain. The system has 
been tested on a corpus of critiques in the do- 
main of trauma care. 
1 Overview 
Many natural language systems have been de- 
veloped to generate coherent text plans (Moore 
and Paris, 1993; Hovy, 1991; Wanner and Hovy, 
1996; Zukerman and McConachy, 1995). How- 
ever, none has the ability to take a set of inde- 
pendently generated yet inter-related text plans 
and produce integrated plans that realize all of 
the communicative goals in a concise and coher- 
ent manner. 
RTPI (Rule-based Text Plan Integrator) was 
designed to perform this task. The need for 
coherence requires that the system be able to 
* This work was supported by the National Library of Medicine under grant R01-LM-05764-01. We thank Bon- 
nie Webber and John Clarke for their suggestions and advice during the course of this research. 
identify and resolve conflict across multiple, in- 
dependent text plans, and exploit relations be- 
tween communicative goals. Conciseness re- 
quires the ability to aggregate and subsume 
communicative goals. Although our work was 
motivated by the need to produce coherent, in- 
tegrated messages from the individual critiques 
produced by a decision support system for emer- 
gency center trauma care, this same task will 
arise in future systems as they make use of in- 
dependent modules that need to communicate 
with a user. Thus the system should have sim- 
ple, domain-independent rules, but should also 
be flexible enough to allow the addition of rules 
specific to the domain at hand. 
This paper describes RTPI and our initial 
implementation that works with the kinds of 
text plans representative of a critiquing system. 
While our examples are taken from the domain 
of trauma care, the domain-independent rules 
make the system applicable to other domains of 
critiquing and instruction as well. The moti- 
vation behind RTPI is presented in Section 2, 
and Section 3 contrasts it with other work. 
Then we describe the system's parameters that 
allow flexible response in multiple environments 
(Section 4). The heart of the system is RTPTs 
domain-independent rule base (Section 5) for 
integrating text plans. The implemented algo- 
rithm and the results of its application are pre- 
sented last. 
2 Motivation 
TraumAID (Webber et al., 1992) is a decision 
support system for addressing the initial defini- 
tive management of multiple trauma. Trauma- 
TIQ (Gertner and Webber, 1996) is a module 
that infers a physician's plan for managing pa- 
tient care, compares it to TraumAID's plan, and 
critiques significant differences between them. 
TraumaTIQ recognizes four classes of differ- 
ences: errors of omission, errors of commission, 
scheduling errors, and procedure choice errors. 
Experimentation with TraumaTIQ showed that 
when the physician's plan is deficient, several 
problems are generally detected, and thus mul- 
tiple critiques are independently produced. 
512 
We analyzed 5361 individual critiques com- 
prising 753 critique sets produced by Trauma- 
TIQ on actual cases of trauma care. A critique 
set represents the critiques that are produced at 
a particular point in a case. While each critique 
was coherent and concise in isolation, we found 
several problems within critique sets: some cri- 
tiques detracted from others in the critique set; 
some would make more sense if they took ex- 
plicit account of other critiques appearing ear- 
lier in the set; and there was informational over- 
lap among critiques. 
Our analysis revealed 22 common patterns 
of inter-related critiques, each pattern covering 
some subset of a critique set. While we initially 
developed a domain-dependent system, Trau- 
maGEN, that operated directly on the logical 
form of the critiques produced by TraumaTIQ, 
we noted that many of the patterns were more 
generally applicable, and that the problems we 
were addressing would also arise in other sophis- 
ticated systems that distribute their processing 
across multiple independent modules, each of 
which may need to communicate with the user. 
While such systems could be designed to try 
to prevent problems of this kind from arising, 
the result would be less modular, more complex, 
and more difficult to extend. 
Thus we developed RTPI, a system for con- 
structing a set of integrated RST-style text 
plans from a set of individual text plans. RTPI 
contains a set of domain-independent rules, 
along with adjustable parameters that deter- 
mine when and how rules are invoked. In ad- 
dition, RTPI allows the addition of domain- 
dependent rules, so the system can account for 
interactions and strategies particular to a do- 
main. 
3 Other Work 
The idea of domain-independent text planning 
rules is not new. Appelt (1985) used "inter- 
actions typical of linguistic actions" to design 
critics for action subsumption in KAMP. REVI- 
SOR (Callaway and Lester, 1997) used domain- 
independent operators for revision of a text plan 
for explanation. Because our rules operate on 
full RST-style text plans that include commu- 
nicative goals, the rules can be designed to in- 
tegrate the text plans in ways that still satisfy 
those goals. 
The Sentence Planner (Wanner and Hovy, 
1996) uses rules to refine a single initial tree rep- 
resentation. In contrast, RTPI operates on sets 
of complete, independent text plan trees. And 
while REVISOR handles clause aggregation, 
and Sentence Planner removes redundancies by 
aggregating neighboring expressions, neither of 
them addresses the aggregation of communica- 
tive goals (often requiring reorganization), the 
TraumaTIQ critiques: 
Caution: check for medication allergies.and 
do a laparotomy immediately to treat the 
intra-abdominal injury. 
Consider checking for medication allergies 
now to treat a possible GI tract injury. 
Please remember to check .for medication al- 
lergies before you give antibiotics. 
Message from RTPI integrated plan: 
Caution: check for medication allergies to 
treat the intra-abdominal injury and a possi- 
ble GI tract injury, and do it before giving an- 
tibiotics. Then do a laparotomy to complete 
treating the intra-abdominal injury. 
Figure 1: Result of communicative goal aggre- 
gation. 
revision and integration of text plans to remove 
conflict, or the exploiting of relations between 
communicative goals as done by RTPI. Simi- 
larly, WISHFUL (Zukerman and McConachy, 
1995) includes an optimization phase during 
which it chooses the optimal way to achieve a 
set of related communicative goals. However, 
the system can choose to eliminate propositions 
and does not have to deal with potential conflict 
within the information to be conveyed. 
4 System Parameters 
Although RTPI's rules are intended to be 
domain-independent, environmental factors 
such as the purpose of the messages and the 
social role of the system affect how individual 
text plans should be integrated. For example, 
if the system's purpose is to provide directions 
for performing a task, then an ordered set of 
actions will be acceptable; in contrast, if the 
system's purpose is decision support, with the 
user retaining responsibility for the selected 
actions, then a better organization will be one 
in which actions are grouped in terms of the 
objectives they achieve (see Section 5.1.1). 
Similarly, in some environments it might be 
reasonable to resolve conflict by omitting 
communicative goals that conflict with the sys- 
tem's action recommendations, while in other 
environments such omission is undesirable (see 
Section 5.1.2). 
RTPI has a set of system parameters that 
capture these environmental factors. These pa- 
rameters affect what rules are applied, and in 
some cases how they are applied. They allow 
characteristics of the output text plans to be 
tailored to broad classes of domains, giving the 
system the flexibility to be effective over a wide 
range of problems. 
513 
5 The Rule-Base 
RTPTs input consists of a set of text plans, 
each of which has a top-level communicative 
goal. Rhetorical Structure Theory (Mann and 
Thompson, 1987) posits that a coherent text 
plan consists of segments related to one an- 
other by rhetorical relations such as MOTIVA- 
TION or BACKGROUND. Each text plan pre- 
sented to RTPI is a tree structure in which 
individual nodes are related by RST-style re- 
lations. The top-level communicative goal for 
each text plan is expressed as an intended effect 
on the user's mental state (Moore, 1995), such 
as (GOAL USER (DO ACTION27)), The kinds of goals 
that RTPI handles are typical of critiquing sys- 
tems, systems that provide instructions for per- 
forming a task, etc. These goals may consist of 
getting the user to perform actions, refrain from 
performing actions, use an alternate method to 
achieve a goal, or recognize the temporal con- 
straints on actions. 
Rules are defined in terms of tree specifica- 
tions and operators, and are stylistically simi- 
lar to the kinds of rules proposed in (Wanner 
and Hovy, 1996). When all the tree specifica- 
tions are matched, the score function of the rule 
is evaluated. The score function is a heuristic 
specific to each rule, and is used to determine 
which rule instantiation has the best potential 
text realization. Scores for aggregation rules, 
for example, measure the opportunity to re- 
duce repetition through aggregation, subsump- 
tion, or pronominal reference, and penalize for 
paragraph complexity. 
Once a rule instantiation is chosen, the sys- 
tem performs any substitutions, pruning, and 
moving of branches specified by the rule's op- 
erators. The rules currently in use operate on 
text plan trees in a pairwise fashion, and re- 
cursively add more text plans to larger, already 
integrated plans. 
5.1 Classes of Rules 
RTPI has three classes of rules, all of which 
produce an integrated text plan from separate 
text plans. The classes of rules correlate with 
the three categories of problems that we identi- 
fied from our analysis of TraumaTIQ's critiques, 
namely, the need to: 1) aggregate communica- 
tive goals to achieve more succinct text plans; 
2) resolve conflict among text plans; and 3) ex- 
ploit the relationships between communicative 
goals to enhance coherence. 
5.1.1 Aggregatlon 
Our analysis of TraumaTIQ's output showed 
that one prevalent problem was informational 
overlap, i.e. the same actions and objectives 
often appeared as part of several different in- 
put text plans, and thus the resulting messages 
(N .  
'A3,A4\])) 
(Recommend User {AI,A2,A3}) (Pefsuad~U (Do U (AI,A2,A3.A4})) 
Jo A I,A2, and A3 |N 
(Mi)llvaticm \[AI.A2.A3.A4} (G2}) 
(Bel User ( Pan-of ( A I ,A2.A3,A4 ) { G2 ) )1 
(Inform User (Pan-of {AI,A2,A3,A4} {G21)) 
as l~ext of G$. ~ 
^31)) 
(Recommend User {AO}) (Persuaded U (~o U {A0,A2.A3})) 
De AO,A2, a*~ AJ 
(Molivafio~( A0,A2.A3 \] \[GI 1) 
(Bel User (Pan-~ {A0,A2.A3} \[GI \])) 
\[ 
\[Inftwrn User (Pan-of {A0,A2,A3} (GI })) 
as part of Gl. 
Figure 2: Input to RTPI (see Figure 3). 
appear repetitious. Aggregation of the commu- 
nicative goals associated with these actions and 
objectives allows RTPI to make the message 
more concise. 
Aggregation of overlapping communicative 
goals is not usually straightforward, however, 
and often requires substantial reorganizing of 
the trees. Our approach was to draw on the or- 
dered, multi-nuclear SEQUENCE relation of RST. 
We posited that separate plans with overlapping 
communicative goals could often be reorganized 
as a sequence of communicative goals in a sin- 
gle plan. The recommended actions can be dis- 
tributed over the sequentially related goals as 
long as the new plan captures the relationships 
between the actions and their motivations given 
in the original plans. 
For example, one complex class of aggrega- 
tion is the integration of text plans that have 
overlapping actions or objectives, but also con- 
tain actions and objectives that do not overlap. 
When those that overlap can be placed together 
as part of a valid sequence, a multi-part message 
can be generated. RTPI produces an integrated 
text plan comprised of sequentially related seg- 
ments, with the middle segment conveying the 
shared actions and their collected motivations. 
The other segments convey the actions that 
temporally precede or follow the shared actions, 
and are also presented with their motivations. 
For example (Fig. 5), suppose that one text 
plan has the goal of getting the user to perform 
actions A0, A2, and A3 to achieve G1, while a 
second text plan has a goal of getting the user 
to perform A1,A2, A3, and A4 to achieve G2. 
Figure 3 presents the text plan resulting from 
the application of this rule. Realization of this 
text plan in English produces the message: 
Do AO as part of G1, and A1 as part of G2. 
Next do A2 and A3 to address both of these 
goals. Then do A4 to complete G2. 
514 
(G~Jal U (DO U |h0, AI, A2, A3. A4))) 
( C'~);bl U (IA) U {A2, A3|))" (Goal U (Do U (A41)) (¢~u(Dou c^0~,,) seo ~ sEo 
(ooal ULIMU (A0})) SEQ do m ~ ,0 
(inform U~r (~..of {A2~3} (G),°2),) (Infr, nn Or,~r (End {A4) {G2})) 
(Bel Us~'r ( P,m-~ {A0} {GI })) (kl Ur, er (P~N6 (A l ) (O2|)) ~ dm~Idrr, baA off,m*,# ge~/,v. ~ comeuu G2. 
(Inform User (Pro't-of {AO} (GI })) (lnfocm Usct (F~l-of {AI ) {G2})) 
st Fart of °l ez lmr# o\] G2. 
Figure 3: Result of a complex aggregation rule (see Figure 2). 
This kind of aggregation is especially appropri- 
ate in a domain (such as trauma care) where 
the clause re-ordering normally applied to en- 
able aggregation (e.g. Sentence Planner) is re- 
stricted by the partial ordering of sequenced in- 
structions. 
RTPI can also handle aggregation when ac- 
tions or objectives are shared between differ- 
ent kinds of communicative goals. The bot- 
tom part of Figure 1 is the text realized from 
a text plan that was produced by the appli- 
cation of two rules to three initial text plans: 
one rule that applies to trees of the same form, 
and one that applies to two distinct forms. The 
first rule aggregates the communicative goal 
(GOAL USER (DO USER check_med_allergies)) that exists 
in two of the text plans. The second rule looks 
for overlap between the communicative goal of 
getting the user to do an action and the goal of 
having the user recognize a temporal constraint 
on actions. The application of these two rules 
to the text plans of the three initial messages 
shown in the top part of Figure 1 creates the 
integrated text plan shown in Figure 4 whose 
English realization appears in the bottom part 
of Figure 1. 
RTPI's parameter settings capture aspects of 
the environment in which the messages will be 
generated that will affect the kind of aggrega- 
tion that is most appropriate. The settings for 
aggregation determine whether RTPI empha- 
sizes actions or objectives. In the latter case 
(appropriate in the trauma decision-support en- 
vironment), an arbitrary limit of three is placed 
on the number of sequentially related segments 
in a multi-part message, though each segment 
can still address multiple goals. This allows the 
reorganization of communicative goals to enable 
aggregation while maintaining focus on objec- 
tives. 
5.1.2 Resolving Conflict 
The ability to recognize and resolve conflict is 
required in a text planner because both the ap- 
pearance and resolution of conflict can be the 
result of text structure. RTPI identifies and re- 
solves a class of domain-independent conflict, 
with the resolution strategies dependent upon 
the social relationship between the user and the 
system. In addition, the system allows the user 
to add rules for domain-specific classes of con- 
flict. 
One class of conflict that can best be resolved 
at the text planning level results from implicit 
messages in text. Resolving conflict of this kind 
within independent modules of a critiquing sys- 
tem would require sharing extensive knowledge, 
thereby violating modularity concepts and mak- 
ing the planning process much more complex. 
For example, suppose that the user has con- 
veyed an intention to achieve a particular objec- 
tive by performing act Au. One system module 
might post the communicative goal of getting 
the user to recognize that act Ap must precede 
Au, while a different module posts the goal of 
getting the user to achieve the objective by ex- 
ecuting As instead of Au. While each of these 
communicative goals might be well-motivated 
and coherent in isolation, together they are in- 
coherent, since the first presumes that Au will 
be executed, while the second recommends re- 
tracting the intention to perform Au. A text 
planner with access to both of these top-level 
communicative goals and their text plans can 
recognize this implicit conflict and revise and 
integrate the text plans to resolve it. 
There are many ways to unambiguously re- 
solve this class of implicit conflict. Strategy se- 
lection depends on the social relationship be- 
tween the system and the user, as captured by 
three of RTPTs parameter settings. This re- 
lationship is defined by the relative levels of 
knowledge, expertise, and responsibility of the 
system and user. Three strategies used by our 
system, and their motivations, are: 
I. Discard communicative goals that implicitly 
conflict with a system recommendation. In 
the above example, this would result in a 
text plan that only recommends doing As 
515 
(Goals U {(Do U {A0}),(Know U (In-Order {A0} {At })),(DO U {A2})}) 
(Goals U {(Do U {A0}),(Know U...)}) SEQ (Goal U (Do U {A2})) 
do A2 
(Recommend U \[A0}) (Persuaded U (Do U \[A0})) (Inform U (In-Order{A0} {AI }) (Persuaded U (In-Order {A0} {A1 })) 
DoAO t~ doitbeforeAi IN 
(Motivation 0} {GI,G2}) (Evidence (In-Order {AO} {AI }) RI)) JN 
(Bel User (Pan-of {A0} {G I,G21)) (Bel User (Reason (In-OrderlA0} {AI }) RI)) IN IN 
(Inform User (Pan-of {A0} {GI,G2})) (Inform User (Reason (In-Order{A0} {A1 }) RI)) as part of Gi and G2 (because RI). 
IN 
(Motivation {A2} {G1 }) 
(Bel User (Pan-o I {A2} {GI })) 
(Inform User (Part-of { A2 } { G I } )) 
to complete G2. 
Figure 4: Result of two rules applied to input shown in Fig. 5. First, a rule that applies to trees 
with top level goals of the form (GOAL USER (DO ...))uses two trees from Fig. 5 to make a tree with 
the two subtrees labelled (1) and (2). Next, a rule that places scheduling trees ( (GOAL U (KNOW 
U (IN-ORDER ...))) ) with related goals inserts a third subtree (3), in this case the entire scheduling 
tree. A domain specific realizer traverses the tree and inserts cue words and conjunctions based on 
relations. 
instead of An. This strategy would be ap- 
propriate if the system is an expert in the 
domain, has full knowledge of the current 
situation, and is the sole arbiter of correct 
performance. 
II. Integrate the text plan that implicitly con- 
flicts with the system recommendation as 
a concession that the user may choose not 
to accept the recommendation. This strat- 
egy is appropriate if the system is an ex- 
pert in the domain, but the user has better 
knowledge of the current situation and/or 
retains responsibility for selecting the best 
plan of action. Decision support is such 
an environment. The top half of Figure 6 
presents two TraumaTIQ critiques that ex- 
hibit implicit conflict, while the bottom 
part presents the English realization of the 
integrated text plan, which uses a CONCES- 
SION relation to achieve coherence. 
III. Present the system recommendation as an 
alternative to the user plan. This may 
be appropriate if the parameters indicate 
the user has more complete knowledge and 
more expertise. 
(Goal UJ.D~ U {h0,h2})) 
(Recommend U (A0,A2}) (Persuaded U (Do U {A0,A21)) 
Do AO and A2 IN 
(Motivation \[A0,A2} {GI }) 
(Bel User (Pan-of { A0,A2 } { G I })) IN 
(Inform User (Pan-of {A0,A2} {GI 1)) as part of Gl. 
(Goal U (Do U {A0})) 
( Recommend U {A0}) (Persuaded U (Do U {A0})) 
Do AO I N 
(Motivation {A2} {GI }) {N 
(Bel User (Pan-of {A0} {Ol })) IN 
(Inform User (Pan-of {A01 {G! })) 
as ~ of G2. 
(Inform U (ln-OrderlA01{Al }) (Persuaded U (In-Order \[A0}{AI D) DoAObeforeAl 
\[N 
(Evidence (In-Order (A0 | { A I \]) R l )) IN 
(Bel User (Reason (In-Order{ A0 } { A ! }) R l)) IN 
(In form User (Reason (In-Order{ A0 } { A I } ) R I)) 
(because RI). 
Figure 5: Input to RTPI (see Figure 4). 
5.1.3 Exploiting Related Goals 
Occasionally two text plans may exhibit no con- 
flict, yet the relationships between their com- 
municative goals can be exploited to produce 
more coherent text. For example, consider the 
following two individual critiques produced by 
TraumaTIQ: 
Caution: do a peritoneal lavage immediately 
as part of ruling out abdominal bleeding. 
Do not reassess the patient in 6 to 24 hours 
until after doing a peritoneal lavage. The out- 
come of the latter may affect the need to do 
the former. 
516 
While the two critiques do not conflict, RTPI's 
rules exploit the relation between the commu- 
nicative goals in their respective text plans to 
produce a more concise and coherent message. 
In particular, one of RTPI's rules recognizes the 
interaction between an initial plan to get the 
user to perform an action As, and a second plan 
that gets the user to recognize a dependency be- 
tween As and another action. This rule creates 
a text plan for the message: 
Do a peritoneal lavage immediately as part of 
ruling out abdominal bleeding. Use the results 
of the peritoneal lavage to decide whether to 
reassess the patient in 6 to P4 hours. 
TraumaTIQ critiques: 
Performing local visual exploration of all ab- 
dominal wounds is preferred over doing a peri- 
toneal lavage for ruling out a suspicious ab- 
dominal wall injury. 
Please remember to check for laparotomy scars 
before you do a peritoneal lavage. 
Message from RTPI integrated plan: 
Performing local visual exploration of all ab- 
dominal wounds is preferred over doing a peri- 
toneal lavage for ruling out a suspicious ab- 
dominal wall injury. However, if you do a 
peritoneal lavage, then remember to first check 
for laparotomy scars. 
5.2 Trailing Comments 
Occasionally when several text plans are inte- 
grated into a single text plan, another text plan 
that overlaps with the integrated plan will re- 
main outside the new plan because the scoring 
function for the applicable rule was too low to 
allow it to combine. This is typically because an 
effort to integrate such a text plan would create 
a message so complex that the heuristic deemed 
it inappropriate. 
However, once concepts have been introduced 
in the integrated text plan, focusing heuristics 
(McKeown, 1985) suggest that other text plans 
containing these concepts be included in the in- 
tegrated plan as well. Rather than restructure 
the result of our transformation (against the ad- 
vice of our heuristic), we append them to the 
end of the message. Thus we refer to them as 
trailing comments. 
Unfortunately, when the communicative goal 
is to get the user to perform an action, trailing 
comments that refer to such actions have the po- 
tential to erroneously suggest new instances of 
actions. Our solution to this problem is imple- 
mented in the text realization templates, where 
we (1) make the focused action the subject of 
the sentence, reflecting its given status in the 
discourse, (2)utilize clue words to call atten- 
tion to its occurrence earlier in the message and 
to the new information being conveyed, and (3) 
subordinate other concepts presented with the 
focused concept by placing them in a phrase in- 
troduced by the cue words "along with". In 
one such example from the trauma domain, the 
main text plan contains the communicative goal 
of getting the user to perform several actions, 
including a laparotomy. A SEQUENCE relation 
is used to adjoin an overlapping text plan as 
a trailing comment, and this additional com- 
municative goal is realized in English as (clue 
words underlined): 
Figure 6: Conflict resolution. 
Moreover., doing the laparotomy is also indi- 
cated, along with repairing the left diaphragm, 
to treat the lacerated left diaphragm. 
6 Algorithm 
RTPI performs rule-based integration of a set 
of RST-style trees. Rules are applied in an or- 
der designed to maximize derived benefit. The 
system first applies the rules that resolve con- 
flict, since we hypothesize that the presence of 
conflict will most seriously hamper assimilation 
of a message. Next, the rules that exploit rela- 
tions between text plans are tried because they 
enhance coherence by explicitly connecting dif- 
ferent communicative goals. Then the aggrega- 
tion rules are applied to improve conciseness. 
Finally, the rules for trailing comments reduce 
the number of disconnected message units. 
The algorithm is both greedy and anytime 
(Garvey and Lesser, 1994); it takes the best re- 
sult from a single application of a rule to a set of 
text plans, and then attempts to further apply 
rules to the modified set. The rule instantiation 
with the highest heuristic score is chosen and 
the rule's operator is applied to the trees using 
those bindings. Since the rules are designed to 
apply incrementally to a set, every application 
of a rule results in an improvement in the con- 
ciseness or coherence of the tree set, and the 
tree set is always a viable set of text plans. The 
user can thus set a time limit for processing of 
a tree set, and the algorithm can return an im- 
proved set at any time. In practice, however, 
the processing has never taken more than 1-2 
seconds, even for large (25 plans) input sets. 
517 
7 Results 
We tested RTPI using the corpus of critiques 
generated by TraumaTIQ. A set of critiques was 
extracted from the middle of each of 48 trauma 
cases, and RST-style text plans were automati- 
cally generated for all the critiques. Then RTPI 
ran each set, and messages resulting from a 
template-based realization of RTPTs text plans 
were analyzed for conciseness and coherence. 
We are currently using templates for sentence 
realization since we have been working in the 
domain of trauma care, where fast real-time re- 
sponse is essential. 
There was a 18% reduction in the aver- 
age number of individual text plans in the 48 
sets examined. The results for individual sets 
ranged from no integration in cases where all of 
the text plans were independent of one another, 
to a 60% reduction in sets that were heavily 
inter-related. More concise messages also re- 
sulted from a 12% reduction in the number of 
references to the diagnostic and therapeutic ac- 
tions and objectives that are the subject of this 
domain. The new text plans also allowed some 
references to be replaced by pronouns during 
realization, making the messages shorter and 
more natural. 
To evaluate coherence, messages from twelve 
cases 1 were presented, in randomly ordered 
blind pairs, to three human subjects not affili- 
ated with our project. The written instructions 
given to the subjects instructed them to note 
whether one set of messages was more compre- 
hensible, and if so, to note why. Two subjects 
preferred the new messages in 11 of 12 cases, 
and one subject preferred them in all cases. All 
subjects strongly preferred the messages pro- 
duced from the integrated text plan 69% of the 
time. 
8 Summary 
Integration of multiple text plans is a task that 
will become increasingly necessary as indepen- 
dent modules of sophisticated systems are re- 
quired to communicate with a user. This pa- 
per has presented our rule-based system, RTPI, 
for accomplishing this task. RTPI aggregates 
communicative goals to achieve more succinct 
text plans, resolves conflict among text plans, 
and exploits the relations between communica- 
tive goals to enhance coherence. 
RTPI successfully integrated multiple text 
plans to improve conciseness and coherence 
in the trauma care domain. We will fur- 
ther explore the application of RTPTs domain- 
independent rules by applying the system to a 
1The evaluation examples consisted of the first eleven 
instances from the test set where RTPI produced new 
text plans, plus the first example of conflict in the test 
set. 
different domain. We would also like to develop 
more domain-independent and some domain- 
dependent rules, and compare the fundamental 
characteristics of each. 

References 
Douglas E. Appelt. 1985. Planning english 
referring expressions. Artificial Intelligence, 
26(1):1-33. 
Charles B. Callaway and James C. Lester. 
1997. Dynamically improving explanations: 
A revision-based approach to explanation 
generation. In Proceedings of the 15th Inter- 
national Joint Conference on Artificial Intel- 
ligence, Nagoya, Japan, August. IJCAI. 
Alan Garvey and Victor Lesser. 1994. A survey 
of research in deliberative real-time artificial 
intelligence. The Journal of Real-Time Sys- 
tems, 6. 
A. Gertner and B. L. Webber. 1996. A Bias To- 
wards Relevance: Recognizing Plans Where 
Goal Minimization Fails. In Proceedings of 
the Thirteenth National Conference on Arti- 
ficial Intelligence, Portland, OR. 
Eduard Hovy. 1991. Approaches to the plan- 
ning of coherent text. In Natural Lan- 
guage Generation in Artificial Intelligence 
and Computational Linguistics, pages 153- 
198. Kluwer. 
William C. Mann and Sandra A. Thompson. 
1987. Rhetorical structure theory: A the- 
ory of text organization. Technical Report 
ISI/RS-87-190, ISI/USC, June. 
Kathleen R. McKeown. 1985. Text Gener- 
ation. Cambridge University Press, Cam- 
bridge, New York. 
Johanna Moore and Cecile Paris. 1993. Plan- 
ning text for advisory dialogues: Capturing 
intentional and rhetorical information. Com- 
putational Linguistics, 19(4):651-695. 
Johanna D. Moore, 1995. Participating in Ex- 
planatory Dialogues, chapter 3. MIT Press. 
Leo Wanner and Eduard Hovy. 1996. The 
HealthDoc sentence planner. In Proceedings 
of the International Workshop on Natural 
Language Generation, pages 1-10. 
Bonnie L. Webber, Ron Rymon, and John R. 
Clarke. 1992. Flexible support for trauma 
management through goal-directed reason- 
ing and planning. Artificial Intelligence in 
Medicine, 4:145-163. 
Ingrid Zukerman and Richard McConachy. 
1995. Generating discourse across several 
user models: Maximizing belief while avoid- 
ing boredom and overload. In Proceedings of 
the International Joint Conference on Artifi- 
cial Intelligence, pages 1251-1257. 
