The HealthDoc Sentence Planner 
Leo Wanner 
Department of Computer Science 
University of Waterloo 
Waterloo 
Ontario N2L 3G1 
Canada 
tel: +1-519-885-1211 ext. 5344 
fax: +1-519-885-1208 
email: lwanner@logos.uwaterloo.ca 
Eduard Hovy 
Information Sciences Institute 
University of Southern California 
4676 Admiralty Way 
Marina del Rey, CA 90292-6695 
U.S.A. 
tel: +1-310-822-1511 ext. 731 
fax: +1-310-823-6714 
email: hovy@isi.edu 
Abstract 
This paper describes the Sentence Planner (sP) 
in the HealthDoc project, which is concerned 
with the production of customized patient- 
education material from a source encoded in 
terms of plans. The task of the sP is to trans- 
form selected, not necessarily consecutive, plans 
(which may vary in detail, from text plans spec- 
ifying only content and discourse organization 
to fine-grained but incohesive, sentence plans) 
into completely specified specifications for the 
surface generator. The paper identifies the sen- 
tence planning tasks, which are highly interde- 
pendent and partially parallel, and argues, in 
accordance with \[Nirenburg et al., 1989\], that' 
a blackboard architecture with several indepen- 
dent modules is most suitable to deal with them. 
The architecture is presented, and the interac- 
tion of the sentence planning modules within 
this architecture is shown. The first implemen- 
tation of the sP is discussed; examples illustrate 
the planning process in action. 
1 Sentence Planning 
1.1 Introduction 
Most current models of text generation in- 
clude a phase of content selection and orga- 
nization, usually performed by a text planner 
or schema application engine, followed by a 
phase of grammatical surface-form rendering, 
performed by a sentence generator. 
In practice, it is usually found that the sen- 
tence generator requires more detailed linguis- 
tic information than text planners or schema 
appliers can provide \[Meteer, 1991; Rambow 
and Korelsky, 1992; Hovy, 1992; Panaget, 1994; 
Wanner, 1994\]. So further planning is required. 
Following \[Rambow and Korelsky, 1992\], we call 
this additional planning task sentence planning 
(even though some operations may cross sen- 
tence boundaries). A sentence planner (sP) 
must specify one of the various possible alter- 
native phrasings at roughly the sentence and 
clause level. By transforming and augmenting 
its input, the sentence planner produces repre- 
sentations detailed enough for the surface gen- 
erator to operate deterministically. Consider an 
example where the lack of sentence planning re- 
sults in an awkward text: 
(o) In some instances, an implant 
wears out, loosens, or fails. If 
an implant wears out, loosens, or 
fails, it will have to be removed. 
More appropriate alternatives can be gen- 
erated when different sentence planning tech- 
niques are used: 
(1) Alternative reference: 
In some instances, an implant 
wears out, loosens, or fails. If 
this happens, it will have to be 
removed. 
(2) Alternative lexical choice: 
In some instances, an implant 
wears out, loosens, or fails. 
If replacement is needed, it will 
have to be removed. 
(3) Removal of redundancy 
(aggregation): 
In some instances, an implant 
wears out, loosens, or fails, 
and \[\] will have to be removed. 
In this paper we describe the sentence plan- 
ner in the HealthDoc project. HealthDoc \[Di- 
Marco et al., 1995\] was established in early 
1995 with the goal of generating customized 
patient-education documents. It combines ex- 
isting generation technology--the sentence gen- 
erator KPML \[Bateman, 1995\] 1 and its input no- 
tation sPL \[Kasper, 1989\]--and new systems, 
such as the sentence planner described here. 
The sentence planner embodies a design that we 
hope has some general applicability in bridging 
the gap between text planners and sentence gen- 
erators. Its input is a specification of the desired 
output content (a patient document) written in 
Tezt Source Language (TSL), see Subsection 2.3; 
its output consists of one or more SPL expres- 
sions. Its general operation is to recombine, en- 
hance, and refine TSL expressions until they are 
adequately specific SPL expressions. 
1.1 Sentence Planning Tasks 
After analysis of a number of patient-education 
documents, including those on diabetes, choles- 
terol, and hormone replacement therapy, we 
have identified the following most important 
sentence planning tasks: 
• Fine-graln discourse structuring: Dis- 
course relations (aST relations, for example) 
that conjoin clause-size pieces in the TSL are 
still open to considerable variation of expres- 
sion, such as the inclusion of a discourse marker 
or the lexical or implicit communication of the 
discourse relation. See, for example, \[Scott and 
de Souza, 1990\] for treatment of ELABORATION, 
\[Vander Linden and Martin, 1995\] of PURPOSE, 
and \[Grote et al., 1995\] of CONCESSION. 
• Sentence grouping and sentence content 
determination: Individual sentences must be 
delimited; temporal, spatial, and causal nuances 
of predicates must be determined, and so on 
\[Meteer, 1991; Pustejovsky, 1995; Stede, 1996\]. 
• Clause-internal structuring: The order 
of clause constituents, taxis, and projectivity 
of propositions within the clause \[Hovy, 1992; 
DiMarco and Hirst, 1993; Panaget, 1994\] must 
be determined; within each sentence, the the- 
matized and focused elements must be identi- 
fied \[Iordanskaja, 1992\]; redundancy must be re- 
moved \[Dalianis and Hovy, 1996a; Dalianis and 
Hovy, 1996b\]. 
• Reference planning (endophoric lexical 
1KPML stands for Komet/Pcnman MultiLingual and 
is a development of the Penman system \[Penman 
Project, 1989\]. 
choice): The particular form of coreference (in- 
cluding anaphora, deixis, and ellipis) and refer- 
ence must be chosen in order to maintain dis- 
course cohesion \[McDonald, 1978; Tutin and 
Kittredge, 1992; Dale and Reiter, 1995\]. 
• Exophoric lexical choice: As argued in 
\[Nirenburg et al., 1989; Meteer, 1991; Wan- 
ner, 1994\], lexical choice other than linguistic 
reference should also be considered a sentence 
planning task, since lexical units predetermihe 
the syntactic structure of a clause, and since 
salience may be realized by lexical means. 
2 The Sentence Planner 
2.1 Constraints and Desiderata 
Given the nature of the task (namely, the trans- 
formation and/or augmentation of partial sen- 
tence specifications into full ones), a set of con- 
straints emerge for the design of the Sentence 
Planner. We believe these constraints to be 
fairly general; that is, that sentence planners 
developed for applications other than Health- 
Doc will have much the same structure. These 
constraints are the following: 
1. The sP must transform an underspecified 
input of deep semantics into a suitably specified 
output of shallow semantics. 
2. The sP must modularize sentence planning 
tasks as far as possible, to facilitate design clar- 
ity, development, and maintenance. Since the 
sentence planning tasks listed above are not 
single-step operations, since they do not have 
to be performed in strict sequence \[De Smedt, 
1990; Reithinger, 1992\], and since the planner's 
operation is non-deterministic (early choices 
may undergo subsequent revision), this suggests 
that each sentence planning task should be im- 
plemented by a separate module or by several 
modules. 
3. The intermediate steps of the sP should 
be accessible and easily interpretable to people 
building the sP, to enable cross-module inter- 
connection and debugging. 
4. The sP must be extensible, allowing new 
modules to be introduced as a need for them 
is identified. 
5. The level of sophistication of the knowledge 
within a module must not be ccnstrained by 
the sP architecture, so that the modules might 
be crude initially but then can incrementally be 
refined without impeding throughput. To facili- 
tate this, the rules and knowledge resources em- 
ployed by the sP modules should be represented 
as declaratively as possible. 
Constraints I and 3 suggest that the inter- 
mediate form(s) of the data during sP opera- 
tion be some sort of SPL-in-emergence; that is, 
a frame continually evolving from the more ab- 
stract input to the final SPL output(s). One 
way to achieve this is to see sP modules as tree- 
transformation engines (viewing SPL and pre- 
SPL expressions as trees2). This means that 
their effects must be written as tree-rewriting 
rules in the following general form (see Sec- 
tion 2.3): 
\[pre~sPil ~ pre-sPi2\] 
Naturally, each module must know which 
tree transformation rule to apply to any given 
pre-sPI, under any given conditions. A suit- 
able mechanism is provided by system networks, 
used just as in KPML'S grammar \[Matthiessen 
and Bateman, 1991\]. Each module contains a 
feature (system) network that discriminates ar- 
bitrarily finely to a desired state. At any point 
in the network, the selection of a feature with an 
associated tree-rewriting rule causes application 
of that rule to the current pre-sP5. Thus tree- 
rewriting rules are realization statements in the 
sP modules (several other realization operators 
are also supported). 
Constraints 2, 4, and 5 suggest that the sP 
• employ a blackboard architecture. As has al- 
ready been argued by \[Nirenburg et hi., 1989\], 
a blackboard is best suited to accommodate the 
flexible order in which modules can take action. 
It also supports the addition of new modules 
without requiring the revision of the interfaces 
of existing modules. 
2.2 Sentence Planner Architecture 
The architecture of the HealthDoc sentence 
planner is shown in Figure I. Solid arrows indi- 
cate data flow; dashed arrows indicate control 
flow. The components are: 
1. A set of modules: Discourse Structuring, 
Content Delimitation, Sentence Structuring, 
Aggregation, Exophoric Lexical Choice, and 
2Strictly speaking, SPL expressions and their origins 
are directed acyclic graphs, not trees; this does not affect 
the design in any way. 
Knowledge Sources 
Paficm data I 
sp~=So~o= k .__Lt__, / 
I I 
I Exopbonc LC I 
I EndophoncLC I 
Figure 1: The Architecture of the Sentence 
Planner in HealthDoc. 
Endophoric Lexical Choice. 3 
2. Knowledge sources: the lexicon (essen- 
tially KPML's lexicon extended by colloca- 
tional \[Wanner and Bateman, 1990\] and qualia 
\[Pustejovsky, 1995\] information), the seman- 
tic and lexicogrammatical resources as dis- 
cussed in \[Wanner, 1994\], the Penman Upper 
Model \[Bateman, 1990\], the HealthDoc Domain 
Model, and the Reader Model of the patient. 
Not shown in Figure I is the history of choices 
made in the course of processing. 
3. The blackboards: the main blackboard, 
which contains the latest pre-SPL expression(s) 
and their derivation history, and the control 
blackboard, which contains bookkeeping infor- 
mation such as the flags that signal the status 
(running/idle/etc.) of each module. 
4. The Administrator: the top-level process that 
invokes modules, updates pre-sPL expressions, 
manages parallel alternative expressions, etc. 
5. The Tree Transformer: the engine that 
matches the left-hand sides of tree transfor- 
aWe also foresee an Ordering module. 
mation rules to pre-SPL expressions and, when 
they match, unifies all variables and replaces 
the matched tree fragments with the right-hand 
sides of the rules. 
6. The Network Traverser: a process that tra- 
verses the system network of each module, 
handles the choice criteria functions (typically, 
these criteria pertain either to the input pre-SPL 
or to one of the knowledge resources), and upon 
reaching tree transformation rules, hands them 
off to the Tree Transformer. 
2.3 The Planning Process 
The sentence planning process transforms an in- 
put TSL expression(s) into one or more SPL ex- 
pressions. 
Input and output notation. TSL is cur- 
rently under development. When fully devel- 
oped, its degree of detail will be variable from 
one-to-one equivalent with SPL, to an abstract 
form that contains only the deep semantic frame 
for a predication (thereby being a notation in 
which, for example, commit suicide has sui- 
cide as head, rather than commit). The flex- 
ible degree of detail of the TSL will allow ei- 
ther more semantic or more surface-oriented 
sentence planning. 
For illustration, compare the following two 
TSLS of varying degrees of abstraction for sen- 
tences (1) to (3) above: 
A more abstract TSL expression: 
((D1 / disjunction 
:domain (W / wearout 
:undergoer (I1 / implant)) 
:range (D2 / disjunction 
:domain (L / loosen 
:undergoer I1) 
:range (F / fail 
:undergoer I1)) 
:circumstance (O / occasional)) 
(C1 / condition 
:domain (R / remove 
:patient I1 
:mode necessity) 
:range D1)) 
Note that coreference information is here en- 
coded by using the same variables. 
A more specific TSL expression: 
((D1 / disjunction 
:domain (W / wearout 
:actor (I1 / implant 
:number singular) 
:tense present) 
:range (D2 / disjunction 
:domain (L 
:range (F / 
:nonorienting-action (I2 / 
:theme I2) 
(C1 / condition 
:domain (It / remove 
:actee I1 
:modality must) 
:range D1)) 
loosen 
:actor I1 
:tense present) 
fail 
:actor I1 
:tense present)) 
three-d-location 
:lex instance 
:number plural 
:determiner some) 
In the more specific TSL expression, the deep 
semantic roles have been replaced by surface 
semantic roles (:actor, :actee), and syntac- 
tic information (tense) and textual information 
(theme) have been added. To see the difference 
between the TSL and SPL, consider the $PL ex- 
pression for the first sentence in (0)-(3): 
SPL output expression: 
(D1 / disjunction 
:domain (W / nondirected-action 
:lex wear-out 
:tense present 
:actor (I1 / object 
:lex implant 
:number plural)) 
:range (D2 / disjunction 
:domain (L / nondirected-act. 
:lex loosen 
:tense present 
:actor I1) 
:range (F / nondirected-act. 
:lex fail 
:tense present 
:actor I1)) 
:nonorienting-action (I2 / three-d-location 
:lex instance 
:number plural 
:determiner some)) 
Overall planning process. The planning 
process starts with the Administrator, which 
places a pre-sPL fragment onto the blackboard 
4 
and activates a module. 
Linguistically, it is not possible to prespecify 
a fixed sequence in which the modules should 
apply \[De Smedt, 1990; Nirenburg et al., 1989\]. 
In some cases, mutual dependencies exist be- 
tween different linguistic phenomena (i.e., also 
planning tasks) that cause race conditions or 
deadlock. We briefly discuss conflict resolution 
strategies in Section 4. In general, though, we 
define a (partial) default sequence for the mod- 
ules: the Discourse Structuring and Sentence 
Structuring modules run first, and in parallel. 4 
They are followed by the Content Delimitation 
module, and finally by the Exophoric and En- 
dophoric Choice modules, also in parallel. This 
is in accordance with the increasing delicacy of 
phenomena the modules deal with. However, 
we will also experiment with other sequences. 
A user-specifiable Agenda that determines the 
ordering of module application has been imple- 
mented. 
Upon activation, a module removes a pre- 
SPL from the blackboard, refines and enriches it, 
and replaces it on the blackboard. The outputs 
of parallel modules are unified and the unified 
pre-SPL expression becomes the working copy. 
The output of non-parallel modules become the 
working copy immediately. After all modules 
have run, the constructed sPL on the blackboard 
is passed to KPML for realization. 
In our current implementation, all modules 
except Discourse Structuring are defined uni- 
formly using system networks. 5 This allows us 
to adopt an already developed and well-known 
representation, and the machinery that pro- 
cesses the information encoded in terms of this 
representation. For example, in \[Panaget, 1994; 
Wanner, 1994\], this machinery has been ap- 
plied to construct an SPL expression from a 'text 
plan'. Unlike this work, which builds up SPL ex- 
pressions anew using the text plan as a source 
of information to guide processing, our sP trans- 
forms the text plan itself into the SPL. We be- 
lieve that such a transformation is more trans- 
parent to the sP builders, enabling them to in- 
spect and manipulate directly the pre-sPL ex- 
4 "In parallel" here means in arbitrary order. 
SHowever, each module is an independent black box. 
This enables separate and parM.lel developraent of each 
module using any formalism. 
pressions formed during the planning process. 
The types of planning operations required re- 
main the same in both cases. 
To implement transformation rules in the 
framework of system networks, we define three 
new realization operators: REWRITE, ADD, and 
SUPPLANT. Each operator has a different ef- 
fect on the fragment(s) of pre-SPL that match 
its left hand side: RBWRIT~ alters its content or 
structure, ADD adds new information but alters 
nothing, and SUPPLANT replaces the matched 
portion with something else (see \[Jakeway et 
al., 1996\]). The transformation rules invoke the 
Tree Transformer when the features they are as- 
sociated with are chosen. The Tree Transformer 
then applies the rules to the current pre-SPL ex- 
pression. 
Four general types of transformation are gen- 
erally required during sentence planning (sym- 
bols preceded by ? are wildcards and match any 
variable or symbol in the pre-SPL): 
1. Augment a pre-SPL expression: 
(ADD ((?X / CAUSS) -+ (:dmarker causation))) 
in which any fragment of the current pre-SPL 
containing as head the deep semantic type 
CAUSE is augmented by the addition of the role 
:dmarker with the filler causation (i.e., a cue 
word signaling causation in the surface form). 
2. Modify a pre-sPL expression: 
(R~WRITB ((?X / RST-RELATm~) ~ (7~ / RST-CAUSE))) 
in which the head of a pre-SPL fragment is 
changed from RST-RELATION to RST-CAUSE. 
3. Remove a portion of a pre-SPL expression: 
(SUPPLANT ((?X / CAUSE 
:rolel y 
:function (:rolel agent)) 
(?x / CAUSE 
:actor y))) 
in which the intermediate roles :role1 and :func- 
tion are removed as soon as the process type 
(and, therefore, also the name of the agent role) 
are determined, and replaced by appropriate in- 
formation. 
4. Move a portion of a pre-SPL expression: 
(SUPPLANT ((?X / PROCESS 
:situation y) -+ 
(?x / y))) 
in which the intermediate role :situation is re- 
moved Pnd its filler moved to occupy the head of 
the fragment rooted at '?x'. (This occurs when 
it has been decided that the predicate 'y' 
e.g., MOVE--is to be expressed as the verb \[to\] 
move, rather than as the support verb construc- 
tion make a move; now MOVE must be promoted 
to be the head of ?x in the emerging SPL). Note 
that 'remove a fragment' always implies 'move 
a fragment'. 
2.4 Inter-Module Conflict Resolution 
If a module is activated but is not able to make 
all the decisions it needs to, or if it makes deci- 
sions that are known to be possibly incompati- 
ble with those made by other modules later on, 
there are in general three options for how to 
proceed: 
1. The module must suspend operation until all 
information required is available. 
2. The module must make decisions somewhat 
blindly and allow backtracking when a decision 
turns out later to be incompatible. 
3. The module must not make the decision but 
produce all alternatives in parallel, to be win- 
nowed out by later processing. 
We do not discuss the first two options since 
they are standard AI search techniques. The 
third option is inspired by the treatment of 
alternatives in statistics-based NLP and the 
way alternative options are handled in MTM 
\[Mel'~uk, 1981\]. In this option, we allow a mod- 
ule to replace an input with multiple alterna- 
tive outputs instead of only one when it cannot 
make a choice. All such alternative pre-SPLs are 
spawned and propagated onto the blackboard, 
so that other modules can work on them all, as 
parallel alternatives. Should one of the alter- 
natives later turn out to be incompatible with 
some module's decision, that alternative simply 
'dies' and is removed from the blackboard. If, 
on the other hand, all alternatives remain viable 
to the end, then the sP has produced more than 
one valid locution from the input. This option 
is only feasible if the networks of the modules 
are not very detailed, i.e., do not lead to an 
excessive number of alternatives. Although we 
will experiment with all three modes, our pri- 
mary intention is to employ mode 1; currently, 
for implementational reasons, we use mode 2 
in its simplest instantiation: when the changes 
made by a module turn out to be incompatible 
with those of other modules, the module starts 
again. 
3 The Modules 
This section describes the functionality of each 
module. As an example, we show the creation 
of the SPL expressions for the sentences (1) to 
(3). Lack of space prevents us from discussing 
the criteria and heuristics that are responsible 
for the individual choices. 
3.1 Discourse Structuring Module 
The Discourse Structuring module decides upon 
the way a discourse relation is communicated. 
So far, three major distinctions are made: 
1. Marker/no-marker: For example, the 
CONDITION relation can be marked by if, in 
case, etc. 
2. Explicit/implicit: CONDITION can be com- 
municated explicitly by discourse means and/or 
lexical means (such as the verb necessitate), or 
implicitly, obtainable via inference. 
3. Nucleus/satellite salience: In the case of 
CONDITION, salience can be shifted by change of 
the order of the condition and effect arguments. 
The pre-SPL expression created by the Dis- 
course Structuring module reflects the choices 
made regarding sentences 1 to 3 above in the 
type and number of roles to introduce, the role 
filler information, etc. The following fragment 
shows the result of the discourse structuring 
module for the sample sentences: 
(D1 
(C1 / condition 
:domain (It / remove 
:actee I1 
:modality must) 
:range D1) 
:dmarker condition 
:range-ordering first)) 
The fragment D1 is not changed. For fragment 
C1, the use of a discourse marker has been de- 
termined. Also, due to the salience of the condi- 
tion part of the utterance, the ':range' role will 
be expressed first. 
3.2 Sentence Structuring Module 
The Sentence Structuring module determines 
the structure of the sentences to be encoded in 
the SPL. This includes: 
1. Sentence boundaries: If two separate sen- 
tences are to be produced, the SPL is split into 
one per sentence and built up sequentially. 
2. Global sentence structure: A sentence 
can be a hypotactic clause complex, a parat- 
actic clause complex, or a simple clause. To de- 
termine this, the Sentence Structuring module 
evaluates whether the predications in the pre- 
SPL are to be communicated as a sequence or 
as a composite, complex event. A sequence of 
events can further be emphasized by a marker. 
In our example, the SPL under construction 
undergoes the following changes: 
((D1 / disjunction 
:domain W 
:range (D2 / disjunction 
:domain L 
:range F 
:dmarker disjunction) 
:dmarker - 
:nonorienting I2 
:theme I2) 
c1) 
It has been determined that the first sPL will 
contain a paratactic clause complex and that 
there will be a disjunction marker between the 
L and F fragments. Since a CONDITION that 
is represented by the roles :domain and :range 
is expressed in KPML as a hypotactic clause 
complex, fragment C1 remains unchanged (note 
that the aggregation module still has not run to 
make the changes for (3)). 
3.3 Content Delimitation Module 
The Content Delimitation module determines 
the material to be included into each separate 
SPL expression. At the present stage, this in- 
cludes the following: 
1. Constituents of a predication: Con- 
stituents that are to be encoded in the pre-SPL. 
Depending on the contextual and linguistic con- 
straints, roles that are listed in the input might 
be suppressed and additional ones might be in- 
troduced. 
2. Causal~ temporal~ aspectual nuances: 
Nuances of the predication that are to be en- 
coded in the pre-sPL. 
The Content Delimitation module primar- 
ily introduces realization statements into the 
pre-sPL expression that constrain the exophoric 
choice module. 
In our example, which starts from a relatively 
specific TSL, the content delimitation module 
does not make visible changes: all roles present 
in the pre-SPL are determined to be realized. 
Starting from the abstract TSL, intermediate 
roles :situation would have been introduced in 
the fragments labeled by the variables W, L, F, 
and R. This syntactically neutral role :situation 
enables the Exophoric Choice module to gener- 
ate different internal clause structures. 
3.4 Aggregation Module 
The Aggregation module eliminates redundancy 
in the pre-sPL by grouping entities that are ar- 
guments of the same relation, process, etc., to- 
gether. The actions of the Aggregation module 
affect, as a rule, the resulting syntactic struc- 
ture. In our example, redundancy is apparent 
in pre-SPL fragments W, L, and F, since their 
only internal difference is their type, as well as 
the repetition of the whole D1 in fragment CI. 
The actions of the Aggregation module result 
in the following changes within pre-SPL C1 for 
sentence (3): 
(C / conjuntion 
:domain (It / remove 
:actee I1 
:lex zeroellipsis 
:modality must) 
:range D1 
:range-ordering first) 
3.5 Exophoric Lexical Choice Module 
The Exophoric Lexical Choice module chooses 
lexical units for those entities specified in the 
pre-SPL that are new in the discourse. More 
precisely, it makes the pre-SPL more concrete 
along three lines: 
1. Lexicalization of discourse structure 
relations: Discourse relations (and their cue 
words) may be realizable by lexical means. In 
our example, the CONDITION marker in (1), (2) 
is lexicalized as if; the DISJUNCTION marker as 
or, and the CONJUNCTION marker as and. 
2. Internal clause structuring: The inter- 
nal clause structure is predetermined by, among 
other means, the valency schema of the lexical 
unit that is chosen to serve as a head of a clause 
or phrase. With the choice of a head lexical 
unit, the salience of the arguments is also pre- 
determined (see, e.g., \[Wanner and Bateman, 
1990\]). One of the choices the exophoric lexi- 
cal choice module makes while generating (2), 
is the replacement of the fragment D1 in the 
CONDITION part by If replacement is needed. 
This choice can be made because the KB con- 
tains the process of replacement as a possible 
consequence of an implant being worn out, loos- 
ened, or having failed. It is not motivated by the 
presence of the coreference link in the TSL. The 
pre-SPL reflects this choice as follows: 
(D1 
(C1 / condition 
:domain (It / remove 
:actee I1 
:modality must) 
:range (N / need 
:actee (ItE / replacement)) 
:dmarker condition 
:range-ordering first)) 
3. Lexicallzation of the entities: this is 
traditionally considered to be the task of the 
lexical choice process. We do not discuss this 
issue here. 
3.6 Endophoric Lexical Choice Module 
The Endophoric Lexical choice module chooses 
lexical units for entities that have already been 
introduced in the discourse either verbatim or 
by related entities. If an entity has been intro- 
duced verbatim, its next mention can be real- 
ized as a personal pronoun, generalized name, 
definite name, deictic pronoun, or ellipsis. If a 
related entity has been introduced, the new lexi- 
cal unit depends on the relation between the two 
entities; compare \[Tutin and Kittredge, 1992\]. 
In our example, if the Exophoric module 
runs first, the Endophoric module ends up only 
pronominalizing implant in the last clause. If 
instead the Endophoric module runs first, the 
SPL produced is (1) rather than (2), i.e., the 
Endophoric Choice module chooses the phrase 
this happens to refer to an implant wears out, 
loosens, or fails. If we assume this variant, the 
pre-SPL expression for the second sentence is 
changed to: 
(C1 / condition 
:domain (It / remove 
:actee I1 
:modality must) 
:range (H / happen 
:actor (I1 / implant 
:lex this)) 
:dmarker condition 
:range-ordering first) 
4 Related Research 
Related work falls into two areas. The first is 
sentence planning as a task, including the orga- 
nization of the planning process. The second 
covers specific subtasks of sentence planning. 
Since we have already provided extensive refer- 
ences to work in the second area, and our focus 
in this paper is not on the detailed presentation 
of these subtasks, we refrain from discussing it 
further. 
In the first area, our sP appears, at first 
glance, to closely resemble DIOCENES \[Niren- 
burg et al., 1989\]: both systems contain a black- 
board with different sentence planning tasks 
performed by separate modules. However, sig- 
nificant differences exist with respect to pro- 
cessing strategies, including blackboard man- 
agement and conflict resolution; the assignment 
of different subtasks to modules; the organiza- 
tion of the modules; and the organization of 
knowledge resources. This issue is discussed in 
\[Jakeway et al., 1996\]. 
In other related work \[Appelt, 1985; Meteer, 
1991; Horacek, 1992\], several sentence plan- 
ning tasks are treated within the process of 
text planning. \[Rambow and Korelsky, 1992; 
Panaget, 1994\] have a separate sentence plan- 
ning component, but they do not separate the 
specific subtasks of sentence planning into dis- 
tinct submodules. Necessarily, some subtasks, 
such as content delimitation, exophoric, and en- 
dophoric choice, then play a less prominent role. 
5 Conclusion 
Individual sentence planning tasks have been 
the focus of much previous research. A few 
sentence planners have combined some of these 
tasks in a single engine. With the HealthDoc 
sentence planner, we are attempting to build an 
architecture that supports both the addition of 
new sentence planning tasks, in the form of new 
modules, as well as continued growth in sophis- 
tication and coverage of individual task perfor- 
mance. This is an ambitious goal. We believe 
that a blackboard architecture with separate 
modules, a situation discr:.mination mechanism 
such as feature networks, and a continuously 
transformable internal representation, from TSL 
input to SPL output using tree transformation 
operators, to be a promising avenue of research, 
given the complexity of the problems facing text 
generators. 
Acknowledgments 
We would like to thank Bruce Jakeway for im- 
plementations as well as the other members of 
the Sentence Planning Group, Chrysanne Di- 
Marco, Phil Edmonds, and Daniel Marcu, for 
many fruitful discussions. Special thanks to 
John Wilkinson, who was one of the "Key Sen- 
tence Planners" during the first phase of our 
work. 

References 
\[Appelt, 1985\] D.E. Appelt. Planning natural 
language utterances. Cambridge University 
Press, Cambridge, England, 1985. 
\[Bateman, 1990\] J.A. Bateman. Upper Model- 
ing: Organizing Knowledge for Natural Lan- 
guage Processing. In 5th International Work- 
shop on Natural Language Generation, Pitts- 
burgh, PA., 1990. 
\[Bateman, 1995\] J. Bateman. KPML: 
KOMET-Penman Multilingual Linguistic Re- 
source Development Environment. In Pro- 
ceedings of the 5th European Workshop on 
Natural Language Generation, Leiden, 1995. 
\[Dale and Reiter, 1995\] R. Dale and E. Reiter. 
Computational Interpretation of the Gricean 
Maxims in the Generation of Referring Ex- 
pressions. Cognitive Science, 19(2):233-263, 
1995. 
\[Dalianis and Hovy, 1996a\] 
H. Dalianis and E.H. Hovy. Aggregation in 
Natural Language Generation. In G. Adorni 
and M. Zock, editors, Trends in natural lan- 
guage generation: An Artificial Intelligence 
perspective. Springer Verlag, Berlin & Heidel- 
berg, 1996. 
\[Dalianis and Hovy, 1996b\] 
H. Dalianis and E.H. Hovy. On Lexical Ag- 
gregation and Ordering. In Proceedings of the 
8th International Workshop on Natural Lan- 
guage Generation. Herstmonceux, 1996. 
\[De Smedt, 1990\] K. De Smedt. IPF: An Incre- 
mental Parallel Formulator. In Robert Dale, 
Christopher S. Mellish, and M. Zock, editors, 
Current research in natural language genera- 
tion. Academic Press, 1990. 
\[DiMarco and Hirst, 1993\] C. DiMarco and 
G. Hirst. A Computational Theory of Goal- 
Directed Style in Syntax. Computational Lin- 
guistics, 19(3):451-499, 1993. 
\[DiMarco et al., 1995\] C. DiMarco, G. Hirst, 
L. Wanner, and J. Wilkinson. Healthdoc: 
Customizing patient information and health 
education by medical condition and personal 
characteristics. In Proceedings of the Work- 
shop on Patient Education, Glasgow, 1995. 
\[Grote et al., 1995\] B. Grote, N. Lenke, and 
M. Stede. Ma(r)king Concessions in English 
and German. In Proceedings of the 5th Euro- 
pean Workshop on Natural Language Gener- 
ation, Leiden, 1995. 
\[Horacek, 1992\] H. Horacek. An Integrated 
View of Text Planning. In Proceedings of 
the 6th International Workshop on Natural 
Language Generation, Trento, Italy, 1992. 
Springer-Verlag. 
\[Hovy, 1992\] E.H. Hovy. Sentence Planning Re- 
quirements for fautomated Explanation gen- 
eration. In Proceedings of the Workshop 
on Explanation Facilities for Model-Based 
Expert Systems, DIAMOD-Bericht no. 23, 
Sankt Augustin, Germany, 1992. 
\[Iordanskaja, 1992\] L. Iordanskaja. Commu- 
nicative Structure and its Use during Text 
Generation. International Forum on In- 
formation and Documentation, 17(2):15-27, 
1992. 
\[Jakeway et al., 1996\] B. Jakeway, E. Hovy, 
and L. Wanner. Specification of the Health- 
Doc sentence planner. Internal Note, CS 
Department, University of Waterloo and 
USC/ISI, Waterloo and Marina del Rey, 
1996. 
\[Kasper, 1989\] R. Kasper. SPL: A sentence 
plan language for text generation. Technical 
report, Information Sciences Institute, Uni- 
versity of Southern California, 1989. 
\[Matthiessen and Bateman, 1991\] C.M.I.M. 
Matthiessen and J.A. Bateman. Text gener- 
ation and systemic-functional linguistics: ex- 
periences .from English and Japanese. Frances 
Pinter Publishers and St. Martin's Press, 
London and New York, 1991. 
\[McDonald, 1978\] D. D. McDonald. Subse- 
quent References: Syntactic and Rhetorical 
Constraints. In Theoretical Issues in Natural 
Language Processing--2 (TINLAP). ACM, 
New York, 1978. 
\[Mel'Suk, 1981\] Igor Mel'Suk. Meaning-text 
Models: a Recent Trend in Soviet Linguis- 
tics. Annual Review of Anthropology, 10:27- 
62, 1981. 
\[Meteer, 1991\] M.W. Meteer. Bridging the 
Generation Gap Between Text Planning and 
Linguistic Realization. Computational Intel- 
ligence, 7(4):296- 304, 1991. 
\[Nirenburg et al., 1989\] 
S. Nirenburg, V. Lesser, and E. Nyberg. Con- 
trolling a Language Generation Planner. In 
Proceedings of the Joint Conference on Arti- 
ficial Intelligence, pages 1524-1530, Detroit, 
1989. 
\[Panaget, 1994\] F. Panaget. Using a Textual 
Representation Level Component in the Con- 
text of Discourse or Dialogue Generation. In 
Proceedings of the 7th International Work- 
shop on Natural Language Generation, Ken- 
nebunkport, 1994. 
\[Penman Project, 1989\] Penman Project. PEN- 
MAN documentation: the Primer, the User 
Guide, the Reference Manual, and the Nigel 
manual. Technical report, USC/Information 
Sciences Institute, Marina del Rey, Califor- 
nia, 1989. 
\[Pustejovsky, 1995\] J. Pustejovsky. The Gener- 
ative Lexicon. MIT Press, Cambridge, 1995. 
\[Rambow and Korelsky, 1992\] O. Rambow and 
T. Korelsky. Applied Text Generation. In Ap- 
plied Conference on Natural Language Pro- 
cessing, Trento, Italy, 1992. 
\[Reithinger, 1992\] Norbert Reithinger. 
Eine parallele Architektur zur inkrementellen 
Generierung multimodaler DialogbeitrEge. In- 
fix Verlag, St. Augustin, 1992. 
\[Scott and de Souza, 1990\] D. Scott and C. S. 
de Souza. Getting the Message Across in RSW- 
Based Generation. In R. Dale, C. Mellish, 
and M. Zock, editors, Current Research in 
Natural Language Generation, pages 47-73. 
Academic Press, London, 1990. 
\[Stede, 1996\] M. Stede. A generative perspec- 
tive on verbs and their readings. In Pro- 
ceedings of the 8th International Workshop 
on Natural Language Generation, Herstmon- 
ceux, 1996. 
\[Tutin and Kittredge, 1992\] 
A. Tutin and R. Kittredge. Lexical Choice 
in Context: Generating Procedural Texts. In 
Proceedings of COLING 92, pages 763-769, 
1992. 
\[Vander Linden and Martin, 1995\] 
K. Vander Linden and J.H. Martin. Express- 
ing rhetorical relations in instructional text: 
A case study of the purpose relation. Com- 
putational Linguistics, 21(1):29-57, 1995. 
\[Wanner and Bateman, 1990\] L. Wanner and 
J.A. Bateman. A Collocational Based Ap- 
proach to Salience-Sensitive Lexical Selec- 
tion. In 5th Natural Language Generation 
Workshop, Pittsburgh, PA., 1990. 
\[Wanner, 1994\] L. Wanner. Building Another 
Bridge over the Generation Gap. In Pro- 
ceedings of the 7th International Workshop 
on Natural Language Generation, Kenneb- 
unkport, 1994. 
