Strategies for Sequencing as a Planning Task 
Daniel D. Suthers 
Learning Research and Development Center 
University of Pittsburgh 
3939 O'Hara Street 
Pittsburgh PA 15260 
suthers+@pitt.edu 
Abstract. The paper summarizes an ongoing investiga- 
tion of the discourse planning tasks concerned with the se- 
quencing of utterances and their parts. Content selection 
provides some important constraints on sequencing, most 
notably those derived from the preconditions and effects of 
planning operators. However these operators underconstraln 
sequencing, especially below the granularity at which they 
interface with a domain knowledge source. Further con- 
straints are available from the integrative processes of the 
heaxer or reader and from working memory \]imlts. Applica- 
tion of these constraints is a matter for discourse planning 
because the choices relate to one's communicative goals. The 
planning task is one of translating functionally relevant re- 
lationships between units to be ordered into ordering con- 
straints. A collection of strstegies for this task are presented. 
Some of the strategies were used in an earlier implemented 
system; many are justified by prior psycholinguistic research. 
Also discussed include current efforts to extend the work to 
focus structure in general, and to address the handling of 
conflicts between strategies. 
Introduction 
The extent to which the components of a text or utter- 
ance succeed in carrying out their intended function de- 
pends in part on the sequence in which they are realized. 
For example, a critical aspect of understanding an ex- 
planation is to integrate the concepts and propositions 
in the explanation with existing knowledge. Sequenc- 
ing decisions should attempt to facilitate this integra- 
tion and otherwise enhance the intended functionality 
of the segments of the explanation. Superficially, the 
"sequential structure = of discourse is simply the order 
in which its elements are positioned in a linear medium. 
However, some of the ordering may be arbitrary. In a 
theoretical analysis, it is more useful to define the se- 
quential structure of discourse as a partial ordering 
that has specific justifications. This paper provides a 
collection of such justifications in the form of strate- 
gies for translating functionally significant relationships 
between discourse elements into palrwise ordering I con- 
straints between those elements. 
We begin with a discussion of the nature of the se- 
1 "Ordering" and "sequencing" are used interchangeably. 
quencing task and the advantages of explicit operators 
for this task. A number of strategies for the coher- 
ent ordering of an explanation are then presented, some 
of which were used in an earlier implemented system 
\[Suthers 1993a\], and many of which are justified by 
prior psycholinguistic research. Directions for further 
research are also discussed, including current efforts to 
extend the work to focus structure in general (i.e., sub- 
ordination structure as well as sequential structure), 
and to address the handling of conflicts between strate- 
gies (e.g., between centering theory and McKeown's fo- 
cus preferences). 
Approaches to the Sequencing Task 
Previous work in generation has handled sequencing de- 
cisions in a number of ways. Schematic approaches spec- 
ify allowable orderings implicitly in terms of the tran- 
sitions of finite state automata \[McKeown 1985\]. Non- 
determinism in these automata has been addressed us- 
ing focus preferences for selecting from a content pool, 
these preferences being embodied in a selection mecha- 
nism. Other approaches exploited the structure of do- 
main knowledge with mechanisms for traversing data 
structures representing this knowledge \[Paris & McK- 
eown 1986, Sibun 1992\]. Planning approaches initially 
utilized more local yet still schematic specifications of 
ordering, expressed as preconditions or optional satel- 
lites for plan operators \[Cawsey 1989, Hovy 1988, Moore 
1989\]. More recently, partial order causal link (POCL) 
planning is being applied to discourse planning, with 
partial ordering derived in a principled manner from the 
relationships between preconditions and postconditions 
of plan steps \[Young et gl. 1994\]. 
Content selection processes provide some important 
constraints on sequential structure, most notably in the 
form of satisfaction-precedence relations derived from 
the preconditions and effects of discourse planning op- 
erators. However these processes underconstraln se- 
quential structure. This is especially true below the 
granularity at which the operators interface with a do- 
main knowledge source, because the latter is partially 
responsible for providing collections of related content 
that can't be specified by domain-independent opera. 
tors. For example, a specification that some distinguish- 
29 
7th International Generation Workshop • Kennebunkport, Maine * June 21-24, 1994 
ing attributes of an entity should be expressed might 
bind to several nonexclusive alternatives, or a specifl- 
cation for a description of the constituents of an ob- 
ject or process might result in retrieval of a collection 
of propositions. Some sequencing decisions above the 
granularity of access to domain knowledge may be un- 
derconstrained as well, for example the order in which 
to express a list of multiple reasons for a conclusion. 
Further constraints are available, for example from rea- 
soning about the integrative processes of the hearer or 
reader and the impact of working memory limits on 
these processes. Application of these constraints is a 
matter for discourse planning because the choices relate 
to one's communicative goals, as discussed below. Se- 
quential structure is too important to treat arbitrarily 
and too context-sensitive to treat in a schematic man- 
ner. 
Overall, ordering heuristics have not typically been 
made explicit as causally efficacious plan operators. Ex- 
plicit ordering operators facilitate the expression and 
study of alternate theories of sequential structure, and 
have benefits for planning as well. For example, they 
enable a p|A=~er to tell whether it can achieve a com- 
municative goal by ordering its utterances in a certain 
way, and they provide a handle for changing strategies 
in different discourse situations. Suthers \[1993a\] treated 
sequencing as a distinct planning task and used sequenc- 
ing operators that derived ordering constraints from rel- 
evant relationships between the elements being ordered. 
However, prior planning approaches (including that just 
cited) have not made the effects of ordering decisions ex- 
plicit. Effects can help choose between conflicting oper- 
ators, and must be considered in reasoning about what 
communicative goals are achieved by particular order- 
ings. 
Choice Between Conflicting Operators. When 
ordering operators conflict, the operators express prin- 
cipled reasons for choosing between alternate sequential 
structures. The choice is a matter for discourse planning 
because it relates to one's goals. The intended effects 
of conflicting ordering operators can he used to select 
between them provided that these effects can be related 
to contextual factors such as superordinate goals and 
stylistic preferences \[DiMarco & Hirst 1993, McCoy & 
Cheng 1991, Hovy 1990\]. For example, a common con- 
~ict is between a sequential structure that makes a sin- 
gle entity salient and a sequential structure that flows 
smoothly from one entity to another (~dovetailing~). 
Entity salience might be preferred when the entity in 
question is a topic of the current segment, while dove- 
tailing might be preferred when communicating a re- 
lational structure among equally important entities or 
to make a transition to a new topic. (A definition of 
~topic ~ is forthcoming.) Or consider the ordering of ex- 
amples with respect to the generality they exemplify. If 
the examples are presented first, the reader has an op- 
portunity to engage in inductive inference towards the 
generality, yet may fail to see the relevance of the ex- 
amples. Presented after a generality, examples provide 
concrete instances under which the generality may be 
indexed. The choice depends on whether the speaker 
or writer is trying to get the hearer or reader to engage 
in active induction or trying to ease the comprehension 
process. 
Achieving Communicative Goals by Sequential 
Structure. Some planners (e.g., POCL planners) can 
notice when a goal can be achieved by actions that have 
already been planned. However, these planners can only 
take advantage of existing actions when the explicit ef- 
fects of a single action meets the goal in question. It 
should be possible to extend these planners to use or- 
dering operators in an opportunistic manner to identify 
ways in which goals can be achieved by feIicltous or- 
dering of multiple actions, e.g., to achieve communica~ 
rive goals by implicit relations \[Lascarides & Oberlander 
1992, Mann & Thompson 1983\]. If an operator's effect 
matches to an active unsatisfied goal and the opera~ 
tor's constraints match existing utterance components 
or planning can satisfy these constraints with new utter- 
ances, then mere installation of the ordering constraint 
can be used to achieve the communicative gc~al. For 
example, suppose a discourse planner has the goals of 
describing a number of events and the causal relations 
between them. Communication of the causal relations 
might be achieved implicitly by describing the events in 
their causal order. 
Sequencing as Exploiting Relationships. The se- 
quencing of two elements of a discourse can only be 
decided if there exists some relationship between the 
elements which has implications for their order of ex- 
pression. The sequencing task is relational, not merely 
selective. Approaches that treat sequencing as a se- 
lective task, for example by using predicates that se- 
lect the most preferred element out of a set of remain- 
ing candidates, are forced to generate a full ordering 
(the sequence of selected items). The structural as- 
pects of sequencing are not a natural consequence of 
the model. When sequencing is treated as the accumu- 
lation of explicit constraints between elements, partial 
orderings can be constructed. Derivation of these con- 
straints from relationships between the elements to be 
ordered addresses the structural aspect of sequencing 
directly. 
The remainder of this paper will present a number of 
sequencing strategies that illustrate how relationships 
30 
7th Intemational Generation Workshop • Kennebunkport, Maine * June 21-24, 1994 
between elements yield sequencing constraints. But first 
some notational preliminaries are required. 
Notation for Ordering Strategies 
Two aspects of sequential structure are distinguished: 
precedence and juxtaposition. Precedence indicates 
that one segment should occur sometime before another 
segment. Precedence is significant when prior commu- 
nication of the contents of one segment facilitate the 
intended functionality of the contents of another seg- 
ment. For example, in technical writing definitions of 
terms usually precede their use. Juxtaposition indi- 
cates that one segment should occur nest to another 
segment in the sequential realization of the explana- 
tion. Juxtaposition is significant when the contents of 
both segments must be in focus of attention at the same 
time in order for the segments to fulfill their intended 
communicative function. For example, statements of 
similarity and difference are usually juxtaposed when 
making a comparison so that the relative significance 
of the similarities and differences can be weighed. Any 
constraint on sequential structure must involve one of 
precedence or juxtaposition. Succession indicates the 
simultaneous presence of both constraints (i.e., that one 
segment should occur immediately before and ~ezt to 
another segment). 
Sequential constraints are placed between text plan 
elements at three granularities. Inter-lntentlonal 
constraints are placed between intentions to perform 
rhetorical and communicative acts \[Suthers 1993a\], 
and constrain the ordering of the utterance segments 
that achieve these intentions. For example, an inter- 
intentional constraint would be used to ensure that a 
description of the structure of a device precede an ac- 
count of how the device carries out its function. Inter- 
propositional constraints are placed between proposi- 
tions and constrain the ordering of sentences, adjectives, 
relative clauses, and other surface realizations of the 
propositions. For example, an inter-propositlonal prece- 
dence constraint between (Parallel plates-l) and (Made- 
of plates-1 metal-l) would allow any of "parallel metal 
plates," Uparal|el plates made of metal" or "The plates 
are parallel. The plates are made of metal." but not 
"metal parallel plates," "metal plates that are parallel," 
or ~The plates are made of metal. The plates are paral- 
lel." (Clearly, this constraint leaves other realization de- 
cisions open.) Intra-proposltional constraints, placed 
between roles of a proposition, controls voice (i.e., which 
role filler is expressed as the subject of a clause). ~ For 
example, an intra-propositional precedence constraint 
2I.u the implemented system of Sut\]aers \[1993a\], subjects were 
always surface-initial. Subject and su.rface-inltial are not con- 
fiatecl in forthco~.ing revisions. 
could select between C'The plates are made of metal." 
vs. "Metal is what the plates are made of." Juxta- 
position constraints between role fillers discourage the 
insertion of a subordinate clause between the realization 
of the role fillers. 
The ordering strategies areexpressed in the form of 
rules for translating other kinds of relationships into 
ordering relations. The general form of the rules is: 
If $1 bears relation R to $2 then 51 ~'~ $2 
where $1 and $2 are segments and S, ~'~ $2 is one of 
the following: 
$1 ~ $2 for precedence ($1 occurs sometime before 
s2), 
$1 \]~,=t $2 for juxtaposition ($1 and 52 are next to 
each other in either order), and 
$1 suc~ 52 for succession ($1 and $2 are juxtaposed 
and $1 precedes $2). 
Propositions are denoted by (P rl r2 *) where P is a 
predicate, the ri are role fillers, and * denotes 0 or more 
additional role fillers. No ordering of rl, r2, and • is 
implied by this notation; in particular • may represent 
other role fillers that can be expressed before or between 
rl and r~ as well as after them. The notation for intra- 
propositional ordering is 
(P r, =d -~ 1, 2 *). 
This constrains the realizations of the fillers of rl and 
r~ to be sequenced in a manner respecting relation ord. ~re~ 
For example, if (P r, r2) is expressed as a clause, 
then rl will be the surface-initial subject, s 
Two predicates and a function are needed to express 
some of the ordering strategies: 
End-p(P), true when P is a proposition at the end of 
a chain of ordered propositions: 
p~ ~'~ ...P. ~'~ p. 
Toplc-p(t), true when t is the argument of or a con- 
stituent of the argument of a rhetorical goal scoping 
over the propositions to be ordered. For example, 
in Describe((Structure capacitor-l)) both capacitor-1 
and (Structure capacitor-l) are %opics." 
Famillarlty(c) = l when a possibly fallible oracle in- 
dicates that concept c is assumed to be familiar to 
the questioner at level l, a member of a partially 
ordered set of levels, perhaps using categories such 
as in Prince's \[1981\] taxonomy, l = false for unfa- 
miliar c. 
8The notation falls to constrain the rea\]Jsatlon of the predicate 
relative to its role fillers, a deficiency that will be acld.ressecl in a 
future revision. 
31 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
Table 1: Rules for Ordering Strategies 
Supplemental Con.~traints 
Antithesis 
Background 
Enrichment 
Evidence 
a 
b 
Exemplification 
a 
b 
Motivation 
Preview 
Summary 
If A is an an~ith.es/J of T 
If B is backgro=~d for F 
If E is an enoch.merit of S 
If E provides evidence for assertions in H 
(go=l-d~uen t=k) 
If E provides an ezamp/e of G 
(illustrative ~e) 
(inductive use) 
If M provides a motivation for S 
If P provides a preuieus of B 
If S is a s=mmar~d of B 
ju¢tt then T , J A. 
then B p,e~ F. 
then S ,~c~ E. 
then E =t~c~ H 
or H '~'=? E. 
then G a~,e= E 
or E v"~ G. 
then M ,,c~ S. 
then P ~ B. 
then 13 P"~ S. 
Memor31 and Processing CovJtrain~ 
Topic Initial 
Topic as Subject 
Dovetailing 
If Topic-p(~), (P1 ~ *) is unordered 
and (P= *) does not contain t 
If Topic-p(t) in (Pz t *) 
If R = (P1 z y ,), End-p(R), 
and (P2 Y z ,) is unordered 
then (Pz t *) v,,~ (P2 *). 
then (P1 t V,e~ ,) 
then (P1 z g ,) ,==c (P2 Y p,e~ z *). 
Dora=in Knowledge Constrains 
Natural Ordering 
Differentia 
Context 
If (N z tt) where N is a Natural-Ordering 
and z is in the predecessor role of N, 
If (Subsumption c s) and s is 
differentiated within c by (P s *) 
If C is context in which S holds 
then (N z v,=~ b') 
then (Subsumption c s) ,ue~ (p s *). 
then G ,uej= S 
EpLctemie Cordezt Constr¢ints 
Familiar First If (P f n *), Famitiarity(P), and 
Familiarity(f) > Familiarity(n) then (P f P"~ n *). 
(See Su~hers \[1993a, b\] for ezch=nge-leuel co~strair~ts on model cl~oice.) 
Table 1 lists the strategies. Selected strategies are dis- 
cussed below. 
Supplemental Constraints 
The most obvious constraints on sequential structure 
are those derived from the inclusion of supplemental 
material. Supplemental material facilitates the under- 
standing or acceptance of other segments of an ex- 
planation in specific ways, the success of which is of- 
ten affected by order of presentation. Suthers \[1993a\] 
used a collection of supplemental relations, these be- 
ing Urhetorical" relations that are primarily intentional 
rather than informational \[Moore & Pollack 1992\] and 
in which one can pre-identify a "nucleus" that is more 
essential to the goals of the discourse than the other re- 
lata. Ordering strategies are associated with each sup- 
plemental relation in Table 1. Some relations give rise to 
an unambiguous ordering, and thus have only one rule 
in the table. The ordering implications of others are 
complicated by possible differences in tutorial strategy 
and individual differences in learning style. One advan- 
tage of separating ordering decisions from supplemental 
relationships is the ability to model stylistic differences 
by changing ordering strategies independently of sup- 
plemental strategies. 
Background. Background material is that which 
functions to enable the comprehenaion of nuclear fore- 
ground material. Thus background functions best if it 
precedes the foreground. Succession is not necessary as 
long as the delay between background and foreground 
is small enough that the background will not have been 
forgotten when the foreground is encountered. 
32 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
Exemplification. Strategic variation is possible in 
the ordering of examples and illustrations. An explainer 
can encourage a questioner to engage in inductive in- 
ference by giving examples before the generalizations 
or concepts that they exemplify, as expressed by the 
E ~-e~ G constraint of version b of the rule. Alternately, 
the example can be given immediately after the con- 
cept or generalization being exemplified, as expressed 
by the G ,~ef E constraint of version a. Under this 
strategy, the questioner does not have to guess the gen- 
eralization and will appreciate why the example was in- 
troduced. When multiple examples are present other 
ordering strategies are available for inter-exemplar or- 
dering: see Mittal & Paris \[1993\] and Rissla~d \[1978\]. 
Motivation. A motivation segment is intended to 
point out the utility of another segment of an expla- 
nation so that the hearer will appreciate the relevance 
of the motivated segment enough to take it seriously. 
Motivation to attend doesn't work retroactively, so the 
motivating segment should occur prior to the motivated 
segment. Succession is preferred, but not necessary. 
Previews and Summaries. By definition, a preview 
precedes the main body of an explanation. To serve the 
function of preparing the questioner for the sequence 
of utterances to follow, a previewed text should be the 
immediate successor of the preview, because a preview 
sets up an expectation that the subsequent segments 
will be those mentioned in the preview. Violation of this 
expectation with intervening material can cause confu- 
sion. A summary is similar to a preview in that both 
provide skeletal characterization of the main body of an 
explanation, though summaries can refer back to con- 
tent that was not available at the time of a preview. 
The pedagogical utility of a summary is in repetition 
and consolidation. Succession is not as important for 
summaries. In fact, a summary might be used because 
there is some extra material between the segments re- 
lated by the summary relation: the summary functions 
to refocus on the main points after the digression. 
Memory and Processing Constraints 
The ordering operators of this section rely on a few pro- 
cessing assumptions that have been supported in the 
psycholinguistic literature. Memory (retention and re- 
trieval) is better for "integrated" items, i.e. those that 
the subject can relate to other prior knowledge \[Keenan 
et al. 1984, Kintsch & van Dijk 1978\]. New integrative 
links are constructed in a limited working memory dur- 
ing comprehension \[Kintsch & van Dijk 1978\]. There is 
a cost (allocation of attention, probability of error) asso- 
ciated with changing the contents of this working mere- 
ory, and the longer an item is kept in working memory 
the more likely it is to be encoded in long term mem- 
ory. Finally, subjects attempt to identify thematically 
central entities and use these as the default locus of 
integration when attempting to integrate new material 
\[Carpenter & Just 1977\]. 
Comments on "topic" and "theme" may be helpful 
at this point. Lavid & Hovy (unpublished working pa- 
per) define "theme" as "that element that informs the 
listener as discourse unfolds how to relate the incom- 
ing information to what is already known." The topic 
predicate is not intended as a generally applicable defini- 
tion of thematic elements outside its use in the ordering 
rules. It merely provides candidates for being made the- 
matic dements though appropriate ordering and other 
devices. 
When a questioner asks a question, the topics of the 
query are brought into the focus of attention. Their fo- 
cal status motivates the relevance of assertions made 
and concepts introduced by the explainer in the re- 
sponse. If assertions or concepts that had no apparent 
relation to the topics were introduced, the questioner 
might be unable to integrate them and could become 
confused due to the conversational implicature of the 
apparent change in subject. For example: 
"What killed the dinosaurs?" 
"Many rocks at the KT-boundary have an un- 
usual concentration of iridium ... 
(The iridium poisoned tltemf The speaker 
doesn't want to talk about dinosaur dentisef)" 
In contrast, the following explanation changes focus of 
attention from the question's topic to other concepts 
and propositions in a well connected manner: 
"The dinosaur extinctions may have been 
caused by a huge meteorite. Evidence for such 
a meteorite is provided by an unusual con- 
centration of iridium found in KT-boundary 
rocks .... " 
This example illustrates the next three rules: the re- 
sponse starts with a proposition about the topic, the 
topic is in the subject position, and new concepts are 
introduced by their relation to prior concepts. 
Topic Initial and Topic as Subject. The "topic 
initial" rule specifies that propositions involving a topic 
entity t are to be expressed before propositions not in- 
volving a topic. As illustrated above, in situations where 
the topic has been pre-identified, this rule ensures con- 
tinuity with the recipient's expectations. In other situ- 
ations, first-mention aids the subject's identification of 
the topic \[Kieras & Boviar 1981, Gernsbacher & Harg- 
reaves 1988\], helping to ensure that facts about a partic- 
ular entity among the many mentioned are remembered. 
33 
7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 
The "topic as subject" rule marks the topic as the 
surface-initial subject of any clause it occurs in. Repe- 
tition in the subject position is another way to facilitate 
the bearer's or reader's identification of the topic \[Kieras 
& Boviar 1981\]. This rule can also help smooth over vi- 
olations of the following "dovetailing" rule by returning 
to a concept that is presumably easy to reactivate. 
Dovetailing. This strategy relies on a combination 
of "argument overlap" \[Kintsch & van Dijk 1978\] and 
"given-new" \[Haviland & Clark 1974\]. The inter- 
propositional P1 ,~,c~ P2 constraint introduces a propo- 
sition P2 when it involves s concept or proposition that 
has been brought into immediate focus of attention by 
another ordering decision to express P1 (i.e., there is ar- 
gument overlap). The intra-propositional y ~re¢ , z con- 
straint makes the surface-initial subject of each propo- 
sition be the role flUer by which it was introduced (i.e., 
proceeds from given to new). Dovetailing is intended 
to minimize working memory changes and maintain 
the connectedness of the subject's evolving conceptual 
model or "text base." 
Dovetailing can be iterated on z in P2 to produce 
chains. However, overapplication of dovetailing risks 
obtaining stream-of-consciousness texts which lack the- 
matic development. Topic salience and dovetailing can 
be in conflict. One manifestation of this conflict is 
the conflict between McKeown's \[1985\] preference of 
"change" over "maintain" over "return" in order to 
avoid having to reintroduce entities one has more to say 
about (based on Sidner \[1979\]) and Gordon et a/'s \[1993\] 
preference of "continuing" over "retaining" over "shift- 
ing" in order to maintain local coherence. Further work 
is required to identify how the choice of one strategy 
over the other depends on register and task demands, 
and to examine the interaction with other factors such 
as location in the discourse. For example, the author 
expects that topic salience will be preferred in contexts 
where the theme is being established (e.g., early in a 
document, paragraph, or other new discourse segment) 
while dovetailing might be preferred once the theme is 
established and a transition to a subordinate theme is 
needed. 
Domain Knowledge Constraints 
Now we consider constraints derived from relationships 
in the domain knowledge being expressed. 
Natural Orderings. Temporal and causal relations 
are normally experienced in a particular direction, for 
example from prior to posterior events or from cause to 
effect. The assumption that our cognitive apparatus is 
adapted to more easily use these relations in the ~for- 
ward" direction suggests that predicates categorized as 
"natural orderings" \[Bienkowski 1986\] be expressed with 
the prior event or cause as the subject, for example ``X 
caused y" is preferred to "y is caused by z," all other 
things being equal. This strategy is consistent with psy- 
cholingulstic evidence indicating that reverse causal and 
temporal ordering inhibits comprehension \[Irwin 1980\] 
and disrupts thematic processing \[Townsend 1983\]. It 
also facilitates the hearer's identification of implicitly 
expressed temporal relations \[Lascarides & Oberlander 
1992\]. Related strategies are available for spatial de- 
scriptions \[Linde 1974, Sibun 1992\]. 
Differentia and Context. The "differentia" relation 
holds between two propositions when one proposition 
(P s .) differentiates a subclass s from other subclasses 
of a class c. An explainer chooses the statement (P s *) 
from amongst all the possible predicates one could apply 
to s because P distinguishes s from the other subdivi- 
sions of e that the questioner might know about. The 
questioner cannot recognize or assess this significance 
of (P s *) unless he or she has been informed of the 
"contrast class" against which the claim (P s *) is be- 
ing made. This rule suggests that the contrast class c 
be introduced first, for example: 
"An electric field is a kind of force field that 
applies a force to a charged object." 
is preferred over 
UAn electric field applies a force to a charged 
object and is a kind of force field." 
Mere precedence is insufficient because the genus pro- 
vides the context in which the differentia is meaning- 
ful. Succession places them both in focus of attention 
at once. This strategy can be generalized to the level 
of sibling communicative acts or rhetorical intentions 
\[Suthers 1993a\], and to other context/statement rela- 
tionships, as suggested by the "Context" rule in Table 1. 
"Epistemic Context" Constraints 
Suthers \[1993a,b\] discusses how the "epistemic context" 
(the knowledge available to the explainer and questioner 
and the knowledge shared in prior dialogue) influences 
the choice between alternate domain models on which 
to base an explanation. Some of the "preferences" pre- 
sented in these publications address sequential concerns. 
For example, when preferences to "say something new," 
"minimize new propositions," and "elaborate on focal 
models" are applied together in a dialogue about some 
phenomenon, incremental construction of increasingly 
elaborate domain models of the phenomenon will result 
\[Suthers eta/. 1992\]. In general, the epistemic context 
34 
7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 
provides important constraints on sequential structure 
across multiple exchanges but has less impact on se- 
quential structure within a single utterance. This pa- 
per does not discuss constraints across exchanges. Only 
one constraint originating in assumptions about hearer 
familiarity is discussed. 
FAm;llar First. New concepts can be introduced in 
relation to familiar ones using any domain relation. 
Suppose concept f is familiar and n is new. Then 
any proposition (P f n .) will do the job, provided 
the predicate P itself is familiar. (An unfamiliar pred- 
icate won't be much help in integrating an unfamiliar 
concept.) The strategy assumes an ordering of famil- 
iarity levels. If f is more familiar than n, it installs 
an intra-propositional constraint that fl should be ex- 
pressed surface-initial, yielding expressions of form "F 
is P-related to n." (not "n is P-X-related to .f'). Once 
familiar-first has been applied, dovetailing can be used 
to introduce other unfamiliar concepts. 
The assumption behind familiar-first is that it is eas- 
ier to retrieve s known concept and integrate a new 
concept in relation to it than it is to construct a new 
concept from scratch and subsequently retrieve a known 
concept to which it can be integrated. However, at and 
below the clause level the delay between the introduc- 
tion of r~ and f may be so small that this strategy does 
not have a discernible effect. Also, if n is a topic, "topic 
as subject" may be more relevant to the speaker's goals. 
These are questions for empirical work. 
Closing Comments 
The sequencing of expository texts and speech should 
be chosen to enhance the intended functionality of each 
textual unit and to facilitate the questioner's integra- 
tion of the communicated information. This aspect of 
the planning task is usefully seen as one of translat- 
ing functionally relevant relationships between textual 
units into ordering constraints. The strategies presented 
in this paper were derived from examination of exam- 
ple explanations and found to be necessary to enable an 
automated explanation generator to produce coherently 
sequenced explanations \[Suthers 1993a\]. (Space con- 
straints necessitate leaving a full example to the work 
just cited.) Many of the strategies were subsequently 
found to correspond to results in the psycholinguistic 
literature. 
As discussed previously, the strategies can conflict. 
Suthers \[1993a\] handled conflicts with a simple pref- 
erence ordering. This approach is inadequate because 
conflict resolution is expected to depend on contextual 
factors such as the speaker or writer's goals and the re- 
lationships of the elements to be ordered with respect 
to discourse segment boundaries. The author is cur- 
rently planning psycholinguistic experiments to test the 
impact of the strategies through reading time and re- 
call studies, with particular concern for how the resolu- 
tion of conflicts between topic salience and other heuris- 
tics should be sensitive to discourse context. Another 
question for future work is the extent to which the se- 
quencing task fits top-down models of planning such as 
DPOCL \[Young eta/. 1994\] vs. requiring s distinct 
mechanism for the application of conflicting heuristics. 
The modeling of working memory limits and memory 
decay in DPOCL operators would require generalizing 
the POCL notion of a %hreat" to be a matter of de- 
gree rather than absolute. The author suspects that 
important generalities will be easier to capture if fac- 
tored out and expressed as explicit sequencing operators 
rather than manifested in variations of multiple decom- 
positional operators. 
The foregoing work is being extended to include the 
subordination structure Of an utterance as well as its 
sequential structure. "Focus trees" \[McCoy & Cheng 
1991, Hovy & MCCoy 1989\] will be used to represent the 
combined sequential and subordination structure, with 
partial orderings placed between siblings at each level of 
the hierarchy. The nodes of a focus tree represent units 
of a text at various granularities ranging from inten- 
tionally defined segments of the sort discussed by Grosz 
& Sidner \[1986\] down through clausal propositions to 
predicates and entity references. "Focusing operators," 
including versions of the sequencing heuristics of this pa- 
per rewritten in the focus tree notation, wiU manipulate 
the tree structure by subordinating one or more subtree 
to another or by installing cross-links between siblings 
to constrain possible traversals. This reformulation is 
expected to be an improvement for several reasons. It 
allows the expression of heuristics for the subordina- 
tion aspects of focus structure, not just sequencing, and 
thematic development can take place at multiple gran- 
ularities. Only one ordering relation, precedence, is re- 
quired. Juxtaposition is handled by grouping within 
a subordinate structure rather than by a different or- 
dering relation that requires special interpretation. An 
availability metric can be defined in terms of distance 
to search back through the tree. Finally, the notation 
can be used in the analysis of texts, and promises to 
support application of the sequencing heuristics to text 
revision as well as text generation. 
Acknowledgements 
This research was conducted while supported by grant 
MDR-9155715 from the National Science Foundation Ap- 
plications of Advanced Technology program. The author 
thanks the anonymous reviewers for thoughtful comments. 
35 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 

References 
\[Bienkowski 1986\] M. Bienkowski. A Computational 
• Model for Extemporaneous Elaborations, CSL Itcport 
1, Cognitive Science Laboratory, Princeton University, 
Princeton, N.J, 1986. 
\[Carpenter & Just 1977\] P. A. Carpenter & M. A. Just. 
Integrative processes in comprehension. In D. Laberge 
S. J. Samuels, Basic Processes in Reading: Percep. 
tion and Comprehension, pp. 217-241. Hg|~dale, N.J.: 
LEA, 1977. 
\[Cawsey 1989\] A. Cawsey. Generating Explanatory Dis- 
course: A Plan-Based, Interactive Approach. Ph.D. 
Dissertation, University of Edinburgh, 1989. 
\[DLMarco & Hixst 1993\] C. DLMexco & G. HJxst. A compu- 
tational theory of goal-dixected style in syntax. Com- 
putational Linguistics 19:451-499, 1993. 
\[Gordon et al. 1992\] P. Gordon, B. Gross, & L. Gillion. 
Pronouns, names, and the centering of attention in dis- 
course. Cognitive Science 17:311-347, 1992. 
\[Gross & Sidner 1986\] B. Grosz& C. Sidner. Attention, in- 
tention, and the structure of discourse. Computational 
Linguistics 12:175-204, 1986. 
\[Haviland & Clark 1974\] S. E. Haviland & H. H. Clark 
1974. What's new? Acquiring new information as a 
process in comprehension. Journal of Verbal Learning 
and Verbal Behavior 13:512-521, 1974. 
\[Hovy 1988\] E. Hovy. Approaches to the planning of coher- 
ent text. Presented at the 4th International Workshop 
on Text Generation. Catalina Island, California, July 
1988. 
\[Hovy 1990\] E. Hovy. Pragmatlcs and natural language 
generation. Artificial Intelligence 43:153-197, 1990. 
\[Hovy & McCoy 1989\] E. Hovy & K. McCoy. Focusing 
your ItST: A step toward generating coherent multi- 
sentential text. Proc. 11th Annual Conference of the 
Cognitive Science Societg, pp. 667-674. Ann Arbor, 
M.ichigan, 1989. 
\[Keenan etal. 1984\] J. Keenan, S. Balllet, & P. Brown. The 
effects of causal cohesion on comprehension and mem- 
ory. Journal of Verbal Learning and Verbal Behavior 
23:115-126, 1984. 
\[Kintsch & van Dijk 1978\] W. Kintsch & T. A. van Dijk. 
Toward a model of text comprehension and production. 
Psychological Review 85:363-394, 1978. 
\[Lascarides & Oberlandez 1992\] A. Lascarides & J. Ober- 
lander. Abduclng Temporal Discourse. In It. Dale, E. 
Hovy, D. It6sner, & O. Stock (Eds.) Aspects of Au- 
tomated Natural Language Generation, pp. 167-182. 
Berlin: Springer-Verlag, 1992. 
\[Linde 1974\] C. Linde. The Linguistic Encoding of Spatial 
Information. Doctoral Dissertation, Columbia Uni- 
versity, 1974. 
\[Mann & Thompson 1983\] W. C. Mann & S. A. Thompson. 
Relational Propositions in Discourse. ISI/RR-83-115. 
Information Sciences Institute, University of Southern 
California, Maxina del trey, California, 1983. 
\[McCoy & Cheng 1991\] K. McCoy & J. Chen 8. Focus of 
attention: ConstrR~i~g what can be said next. In C. 
Paxis, W. Swaztout, & W. Mann (Eds.) Natural Lan- 
guage Generation in Artificial Intelligence and Com- 
putational Linguistics, pp. 103-124. Boston: Kluwer 
Academic, 1991. 
\[McKeown 1985\] K. McKeown. Discourse strategies for 
generating natural language text. Artificial Intelligence 
27:1-41, 1985. 
\[MRtal & Paris 1993\] V. Mittal & C. Paris. Automatic doc- 
ument generation: The interaction of text and exam- 
ples. Proc. Thirteenth International Joint Conference 
on Artificial Intelligence (IJCAI.9$), pp. 1158-1163. 
Chambery, Fzance, 1993. 
\[Moore 1989\] 3. Moore. A Reactive Approach to Ezplana- 
tion in Expert and Advice-giving Systems. Ph.D. Dis- 
sertation, University of California, Los Angeles, 1989. 
\[Paris & McKeown 1986\] C. Paris & K. McKeown. Dis- 
course strategies for describing complex physical ob- 
jects; Prec. NATO Advanced Research Workshop on 
Natural Language Generation. Nijmegen, The Nether- 
lands, August 19-23, 1986. 
\[Prince 1981\] E. F. Prince. Toward a taxonomy of given- 
new information. In Peter Cole (Ed.), Radical Prag- 
matics, pp. 223-255. NY: Academic Press, 1981. 
\[Rissland 1978\] Published under E. B.issland Michener. Un- 
derstanding Understanding Mathematics. Cognitive 
Science 2:361-383, 1978. 
\[Sibun 1992\] P. Sibun. Generating text without trees. 
Computational Intelligence 8, 1992. 
\[Sidner 1979\] C. Sidner. Focusing in the comprehension of 
definite anaphora. In M. Brady & It. Berwick (Eds.), 
Computational Models of Discourse, pp. 267-330. 
Cambridge, Mass: MIT Press, 1979. 
\[Suthcrs 1993a\] D. Suthers. An Analysis of Explanation 
and its Implications for the Design of Explanation Plan- 
ners. Ph.D. Dissertation, Dept. of Computer Science, 
University of Massachusetts, 1993. 
\[Suthers 1993b\] D. Suthers. Preferences for Model Selec- 
tion in Explanation. Proc. Thirteenth International 
Joint Conference on Artificial Intelligence (IJCAI-93), 
pp. 1208-1213. Chambery, France, 1993. 
\[Suthers etal. 1992\] D. Suthers, M. Cornell, & B. Woolf. 
Steps ~om explanation planning to model construction 
dialogues. Proc. Tenth National Conference on Artifi- 
cial Intelligence (AAAI-9£), pp. 24-30. San Jose, CA, 
1992. 
\[Young et al. 1994\] It. Young, J. Moore, & M. Pollack. 
Towards a principled representation for discourse plans. 
To appear in Proc. Sizteenth Annual Conference of the 
Cognitive Science Societ~h Atlanta, GA, 1994. 
