The Role of Cognitive Modeling in Achieving 
Communicative Intentions 
Marilyn A. Walker* 
Mitsubishi Electric Research Laboratories 
201 i Broadway 
Cambridge, MA 02139 
USA 
walk~r~merl, com 
Abstract 
A discourse planner for (task-oriented) dialogue must 
be able to make choices about whether relevant, but 
optional information (for example, the "satellites" in 
an RST-based planner) Should be communicated. We 
claim that effective text planners must explicitly model 
aspects of the Hearer's cognitive state, such as what the 
hearer is attending to and what inferences the hearer 
can draw, in ord@r to make these choices. We argue 
that a mere representation of the Hearer's knowledge is 
inadequate. We support this claim by (1) an analysis 
of naturally occurring dialogue, and (2) by simulating 
the generation of discourses in a situation in which we 
can vary the cognitive parameters of the hearer. Our 
results show that modeling cognitive state can lead to 
more effective discourses (measured with respect to a 
simple task). 
1 Introduction 
Text planning is the task for a speaker (S) of decid- 
ing what information to :communicate to a hearer (H) 
and how and when to communicate it. Over the last 
few years a consensus has emerged that the text plan- 
ning task should be formulated in terms of commu- 
nicative goals or intentions \[19, 25, 23, 16\]. Consider, 
for example, the RST-based planners developed at ISI 
\[13, 21, 14\]. These planners use the discourse relations 
proposed by Rhetorical S'tructure Theory (RST) \[18\] as 
plan operators, by interpreting the requirements on the 
related segments as preconditions, and the resultant ef- 
fect of the discourse relation as a postcondition in a 
traditional AI planning architecture. 
Two types (at least) of discourse relations have been 
identified in the literature. A SUBJECT-MATTER rela- 
tion \[18\] or SEMANTIC relation \[14\] simply reflects a 
relation that exists independently in the world, such as 
causation. Each subject-matter relation can be seen as 
a rhetorical strategy for the linguistic realization of a 
*Walker was partially funded by ARO grant DAAL03- 
89-C0031PRI and DARPA grant N00014-90-J-1863 at the 
University of Pennsylvania and by Hewlett Packard, U.K. 
t Rambow was supported by NATO on a NATO/NSF 
postdoctoral fellowship in France. 
Owen Rambow t 
Universit4 Paris 7 and CoGenTex, Inc. 
TALANA, UFR Linguistique, Case 7003 
2, Place Jussieu 
75251 Paris Cedex 05, France 
rainbow©linguist, j ussieu, fr 
range of communicative intentions \[22\]. A PRESENTA- 
TIONAL RELATION \[18\] or INTERPERSONAL relation \[14\] 
holds between two discourse segments such that the jux- 
taposition increases H's STRENGTH of belief, desire, or 
intention. Each presentational relation maps directly 
to a communicative intention. Examples of presenta- 
tional relations include the MOTIVATION relation, which 
increases H's desire to perform an action, hopefully per- 
suading H to form an intention to do the action. 
Both subject-matter and presentational relations re- 
late two clauses: (1) the NUCLEUS which realizes the 
main point; and (2) the SATELLITE which is auxiliary optional 
information. For example in the MOTIVATION 
relation shown in figure 1, the SATELLITE is the belief 
which provides motivation to do the action realized by 
the proposal or suggestion in the NI;CLEUS. Since the 
SATELLITE information may or may not be realized, pre- 
vious text planners have run in either verbose or terse 
mode, in which either all or no satellite information is 
realized \[22\]. 
If an approach to text planning based on the notion 
of communicative intention is to succeed, it requires 
an appropriate representation of communicative goals, 
and of all mental states required for reasoning about 
these goals. This is especially true in the case of pre- 
sentational relations. We can immediately observe that 
since such relations affect the degree of strength of H's 
belief, desire of intention, we need a gradual representa- 
tion of mental attitudes. To our knowledge, no current 
text planner uses such a gradual representation. Sec- 
ond, it has been widely assumed that a model of what 
the hearer knows determines exactly when to include 
optional information in verbose mode: include optional 
information unless the hearer knows it. However, in our 
analysis of a corpus of 55 naturally-occurring dialogues, 
information that the hearer knew was frequently real- 
ized \[35\]. Consider the following short natural dialogue, 
part of a discussion about which Indian restaurant to 
go to for lunch: 
(1) a. Listen to Ramesh. 
b. He's Indian. 
Clearly, S wants to MOTIVATE H to accept his pro- 
posal with (lb). However, in this situation all of the dis- 
171 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
course participants already knew that Ramesh was In- 
dian. We hypothesize that example (1) shows that there 
are cognitive processing motivations for S's choice to in- 
clude information that is already known to the hearer, 
such as (lb), and that a model of H's cognitive pro- 
cesses are required for a text planner to appropriately 
decide when to include optional information. 
The remainder of this paper is structured as follow. 
We start out by describing in more detail the prob- 
lem facing text planners based on communicative goals 
(Section 2). In Section 3 we briefly review cognitive 
theories of deliberation and inference and relate these 
to an account of working memory. Next, in Section 4 we 
present the Design-World experimental environment, in 
which we embed our cognitive model. In Section 5, we 
present some examples of modeling experiments that 
suggest what sort of information S must access in order 
to generate efficient discourse. Finally, in Section 6 we 
briefly discuss possible implications for text planning 
architectures. 
2 Achieving Communicative Goals 
In this paper, we focus on presentational relations, and 
we use the MOTIVATION relation as a prototypical pre- 
sentational discourse relation to illustrate our points. 
The MOTIVATION relation, whose effect is to increase 
H's desire to perform an action, is shown in Figure 1. 
We will call the nucleus of MOTIVATION the PROPOSAL, 
and any information that can serve as the satellite of 
MOTIVATION we will call a WAR.RANT. 
What sort of representation is needed in order to use 
MOTIVATION for discourse planning? The effect of pre- 
sentational relations is always to increase H's belief, de- 
sire, or intention. Thus we will need (in the case of 
MOTIVATION) some sort of representation of degree of 
desire. In a first attempt at using MOTIVATION, we will 
use utility theory \[6\] and simply associate utilities with 
proposed actions. Under this view, an agent's strength 
of desire to perform an action is the utility he or she be- 
lieves performing the action will yield, where "utility" 
is a quantifiable variable. In section 6 we will discuss 
the limitations of this approach. 
However, even though utility theory can be used as 
the theoretical underpinning of the MOTIVATION rela- 
tion, it will not in general be sufficient because it does 
not take into account the way in which H's beliefs in- 
teract with his attentional state, and the way that H's 
cognitive limitations interact with the demands of the 
task. 
Consider the following scenario. An agent S wants an 
agent H to accept a proposal P. In the situation where H 
always deliberates and H knows of options which com- 
pete with proposal P, H cannot decide whether to ac- 
cept P without a warrant. 1 Previous work has assumed 
1Elsewhere we consider scenarios in which H always ac- 
cepts S's proposal without a warrant and in which H never 
knows of competing options to P \[36\]. 
that the warrant can be omitted if it is already believed 
by H. Presumably the speaker in (2) will not say It's 
shorter if she believes that the hearer knows that the 
Walnut St. route was shorter. 
(2) a. Let's walk along Walnut St. 
b. It's shorter. 
However, consider again (1), repeated here as (3): 
(3) a. Listen to l:tamesh. 
b. He's Indian. 
The warrant in (3b) was included despite the fact 
that it was common knowledge among the conversants. 2 
Our hypothesis is that 3 shows that speakers distinguish 
between information that H knows and information that 
is salient for H \[28\]. Thus even if H knows a warrant 
for adopting S's proposal, if the warrant is not SALIENT 
for H, then S may choose to include it with a proposal. 
We define SALIENT as available in current Working 
Memory, referred to below as Attention/Working Mem- 
ory or AWM. A model of the hearer H's attentional 
state will distinguish between those discourse entities 
and beliefs which are currently available in working 
memory, and thus salient, and those that are not. In 
section 3, we introduce an operationalization of AWM 
and discuss how S can model what is salient for H. 
When a warrant is not SALIENT, H must either infer 
the warrant information, or retrieve it from long term 
memory, or obtain it from an external source in order to 
evaluate S's proposal. Thus S's communicative choice 
as to whether to include the warrant satellite may de- 
pend on S's model of H's attentional state. Further- 
more, it is possible that, even though the warrant is 
not salient, merely a trivial processing effort is required 
to retrieve it, so that it is not worth saying. Another 
possibility is that processing the warrant utterance re- 
quires effort that can be traded off against other pro- 
cesses such as retrieval and inference. In other words, 
S may decide that it is easier just to say the warrant 
rather than require H to infer or retrieve it. We will 
call a text planning strategy that always includes the 
warrant satellite the Explicit-Warrant strategy. 
We see that in addition to S modeling H's knowledge, 
H's attentional state and expectations about other as- 
pects of H's cognitive processes may also influence S's 
text planning decisions, and S cannot simply represent 
H's beliefs as a set of pairs of propositions and associ- 
ated utility. 3 In sum, the choice is hypothesized to de- 
pend on cognitive properties of H, e.g. what H knows, 
H's attentional state, and H's processing capabilities, as 
well as properties of the task and the communication 
channel. 
2After inferring the intended relation the hearer still 
must decide whether s/he believes that Indians know of good 
Indian restaurants \[38\]. 
3How S would have access to the necessary information 
is a separate issue, briefly discussed in section 6. 
172 
7th International Generation Workshop • Kennebunkport, Maine - June 21-24, 1994 
relation name: 
constraints on Nucleus: 
constraints on Satellite: 
constraints on the 
Nucleus + Satellite combination: 
the effect: 
MOTIVATION 
presents an action in which H is the actor 
unrealized with respect to the context of N 
(a proposal in our terminology) 
none 
H's comprehending the Satellite increases H's desire to 
perform action presented in the Nucleus 
H's desire to perform action presented in the Nucleus is increased 
Figure 1: RST definition of Motivation Relation 
In this paper, we explore some cognitive issues in- 
volved in planning to include satellite information found 
in RST presentational' relations and the representa- 
tional demands that arise for text planning tasks. We 
will argue that S must :model H's cognitive state in a 
much more detailed manner than previously assumed 
and put forth a proposal about how S might access the 
information required in order to do so. In order to 
provide evidence for our claim, we will use the cogni- 
tive modeling methodology developed in \[35\], in which 
communicative strategies on the part of S can be repre- 
sented and their effects can be empirically tested. This 
architecture allows us to identify parameters in H's cog- 
nitive state that affect S's communicative decisions and 
therefore must be mod@led. The simulation/modeling 
environment is called Design-World. 
3 Modeling Cognitive Limits 
In Section 2 we proposed some cognitive factors, mo- 
tivated by proposals in naturally occurring dialogue, 
that may provide limits on whether agents can opti- 
mally deliberate proposed actions or make inferences. 
We hypothesized that these factors will determine when 
Explicit-Warrant is an effective strategy. Here we 
briefly present a way of cognitively modeling agents' 
limited attention and the relationship of limited atten- 
tion to deliberation and inference. 
It is well known that human agents have limited at- 
tentional capacity \[20\] and it has been argued that lim- 
ited attention plays a major role in theoretical and sci- 
entific reasoning \[29, 15,! 32\], ie. in deduction, and be- 
lief and intention deliberation. We hypothesized that 
example 2 shows that a warrant must be SALIENT for 
both agents in order to be used in deliberation, i.e. for it 
to motivate H effectivelyl This fits with the psycholog- 
ical theories mentioned above, that only salient beliefs 
are used in deliberation and inference. 
In Design-World, salience is modeled by the AWM 
model, adapted from \[17\]. While the AWM model is ex- 
tremely simple, Landauer shows that it can be parame- 
terized to fit many empirical results on human memory 
and learning \[2\]. AWM consists of a three dimensional 
space in which propositions acquired from perceiving 
the world are stored in chronological sequence accord- 
ing to the location of a moving memory pointer. The 
sequence of memory loci used for storage constitutes a 
random walk through memory with each locus a fixed 
distance from the previous one. If items are encoun- 
tered multiple times, they are stored multiple times \[12\]. 
When an agent retrieves items from memory, search 
starts from the current pointer location and spreads out 
in a spherical fashion. Search is restricted to a partic- 
ular search radius: radius is defined in Hamming dis- 
tance. For example if the current memory pointer locus 
is (0 0 0), the loci distance 1 away would be (0 1 0) (0 
-10) (0 0 1) (0 0-1) (-1 0 0) (10 0). The actual lo- 
cations are calculated modulo the memory size. The 
limit on the search radius defines the capacity of atten- 
tion/working memory and hence defines which stored 
beliefs and intentions are SALIENT. 
The radius of the search sphere in the AWM model 
is used as the parameter for Design-World agents' 
resource-bound on attentional capacity. In the exper- 
iments below, memory is 16x16x16 and the radius pa- 
rameter varies between 1 and 16. Agents with an AWM 
of 1 have access to 7 loci, and since propositions are 
stored sparsely, they only remember the last few propo- 
sitions that they acquired from perception. Agents with 
an AWM of 16 can access everything they know. 4 
The AWM model also gives us a way to measure (1) 
the number of retrievals from memory in terms of the 
number of locations searched to find a proposition; (2) 
the number of inferences that the agents make as they 
means-end reason and draw content-based inferences; 
and (3) the number of messages that the agents send to 
one another as they carry out the dialogue. The amount 
of effort required for each of these cognitive processes 
are parameters of the model. These cost parameters 
support modeling various cognitive or text planning ar- 
chitectures, e.g. varying the cost of retrieval models 
4The size of memory was determined as adequate for 
producing the desired level of variation in the current task 
across all the experimental variables, while still making it 
possible to run a large number of simulations involving 
agents with access to all of their memory in a reasonable 
amount of time. In order to use the AWM model in a differ- 
ent task," the experimenter might want to explore different 
sizes for memory. 
173 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
different assumptions about memory. Since these cog- 
nitive processes are the primitives involved in text plan- 
ning, this framework can be used to model many differ- 
ent architectures rather than the results being specific 
to a particular text-planning architecture. 
The retrieval parameter alone allows us to model 
many different assumptions about memory. For exam- 
ple, if retrieval is free then all items in working memory 
are instantly accessible, as they would be if they were 
stored in registers with fast parallel access. If AWM 
is set to 16, but retrieval isn't free, the model approxi- 
mates slow spreading activation that is quite effortful, 
yet the agent still has the ability to access all of mem- 
ory, given enough time. If AWM is set lower than 16 
and retrieval isn't free, then we model slow spreading 
activation with a timeout when effort exceeds a certain 
amount, so that an agent does not have the ability to 
access all of memory. Thus the AWM parameter sup- 
ports a distinction between an agent's ability to access 
all the information stored in its memory, and the effort 
involved in doing so. 
It does not make sense to fix absolute values for the 
retrieval, inference and communication cost parame- 
ters in relation to human processing. However, Design- 
World supports exploring issues about the relative costs 
of various processes. These relative costs might vary de- 
pending on the language that the agents are communi- 
cating with, properties of the communication channel, 
how smart the agents are, how much time they have, 
and what the demands of the task are \[24\]. Below we 
vary the relative cost of communication and retrieval. 
The advantages of the AWM model is that it has 
been shown to reproduce, in simulation, many results 
on human memory and learning. Because search starts 
from the current pointer location, items that have been 
stored most recently are more likely to be retrieved, 
predicting recency effects \[2\]. Because items that are 
stored in multiple locations are more likely to be re- 
trieved, the model predicts frequency effects \[17\]. Be- 
cause items are stored in chronological sequence, the 
model produces natural associativity effects \[1\]. 
The overall agent architecture is modeled on the 
IRMA agent architecture \[3\] with the addition of AWM. 
See figure 2. As figure 2 shows, AWM interacts with 
other processing because deliberation and means:end 
reasoning only operate on salient beliefs. This means 
that limits in AWM produces a concomitant inferen- 
tial limitation, i.e. if a belief is not salient it cannot be 
used in deliberation or means-end-reasoning. Thus mis- 
takes that agents make in their planning process have a 
plausible cognitive basis. Agents can both fail to access 
a belief that would allow them to produce an optimal 
plan, as well as make a mistake in planning if a belief 
about how the world has changed as a result of planning 
is not salient. Depending on the preceding discourse, 
and the agent's attentional capacity, the propositions 
that an agent knows may or may not be salient when a 
proposal is made \[28\]. 
W 
I rrE oN/woRxING I 
Io= ---J--- , \/ 
• /options 
Figure 2: Design-World version of the IRMA Agent Ar- 
chitecture for Resource-Bounded Agents with Limited 
Attention (AWM) 
4 Experimental Environment: 
Design-World 
Design-World is an experimental environment for test- 
ing the relationship between ways of realizing com- 
municative intentions and agents' cognitive capabili- 
ties, similar to the single-agent TileWorld simulation 
environment \[26, 11\]. Design-World agents can be 
parametrized as to discourse strategy, e.g. whether to 
use the Explicit-Warrant strategy, and the effects of 
this strategy can be measured against a range of cog- 
nitive and task parameters. In section 4.1, we describe 
the Design-World domain and task. In section 4.2, we 
describe two alternate discourse strategies. In section 
4.3, we discuss how performance is evaluated and com- 
pared. Finally, in section 5 we present the experimental 
results. 
4.1 Design World Domain and Task 
The Design-World task requires two agents to carry out 
a dialogue in order to negotiate an agreement on the de- 
sign of the floor plan of a two room house \[30, 39\]. The 
DESIGN-IIOUSE plan requires the agents to agree on how 
to DESIGN-R.OOM-1 and DESIGN-R.OOM-2. At the begin- 
ning of the simulation, both agents know the structure 
of the DESIGN-IIOUSE plan. Each agent has 12 items 
of furniture that can be used in the plan, with utility 
scores ranging from 10 to 56. A potential final design 
plan negotiated via a (simulated) dialogue is shown in 
figure 3. 
174 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
KEY: 
F=FUCHSIA 
G = GREEN 
P= PURPLE 56 IN 155 
( ) 
III)11) 
F 
J 
I 
54 D 
I 
52 53 
ROOM#1 ROOM#2 
DESIGN WORLD COLLABORATIVE PLAN: 4M p0ints 
Figure 3: One Final State for Design-World Standard 
Task: Represents the Collaborative Plan Achieved by 
the Dialogue, 434-points 
Negotiating an agreement consists of a cycle in which: 
(1) individual agents perform means-end reasoning 
about OPTIONS to use various furniture pieces that they 
have in the floor plan; (2) individual agents deliberate 
about which options are preferable; (3) then agents use 
the OPTIONS identified as the content of PROPOSALS to 
other agents to PUT a piece of furniture into one of the 
rooms; (4) then these proposals are ACCEPTED or RE- 
JECTED by the other agent, or acceptance/rejection is 
postponed by a CLARIFICATION. 
Dialogue (4) illustrates agents' communication for 
part of designing room-!: Both the artificial language 
that the agents communicate with and a gloss generated 
from that language in ifalics are shown: ~ 
(4) 1: BILL: Let's put the green rug in the study. 
(propose agent-bill agent-kim option-10: put-act 
(agent-bill green rug room-l)) 
2: KIM: Then, let's put the green lamp there. 
(propose agent-kim, agent-bill option-33: put-act 
(agent-ldm green lamp room-l)) 
3: BILL: Next, let's put the green couch in the study. 
(propose agent-bill agent-kim option-45: put-act 
(agent-bill green couch room-l)) 
4: KIM: No, instead let's put the purple couch in the 
study. 
(reject agent-kim agent-bill option-56: put-act (agent- 
kim purple couch room-l)) 
On receiving a proposal, an agent deliberates whether 
to ACCEPT or r~EJECT the proposal \[6\]. Each furniture 
5The generation of the gloss was not a focus of this study 
and was done via adhoc methods. 
item has a value that contributes to an evaluation of 
the final plan. The values on the furniture items range 
from 10 to 56, and both agents' furniture items range 
over these values. Agents know the values of all the 
furniture items at the beginning of the dialogue. The 
values of the furniture items are used to MOTIVATE the 
agents to ACCEPT a proposal, as well as providing a 
way of objectively evaluating agents' performance. In 
other words, we define each potential action to have an 
associated SCORE; agents can evaluate whether their 
desire to do an action is increased by comparing the 
score of the proposed action with other actions that 
they know about. 
For example, at the beginning of the dialogue, Agent- 
Kim has stored in memory the proposition that (score 
green-rug 56). When she receives Bill's proposal as 
shown in (4-1), she evaluates that proposal in order to 
decide whether to accept or reject it. As part of evalu- 
ating the proposal she will attempt to retrieve the score 
proposition stored earlier in memory. Thus the proposi- 
tions about the scores of furniture items are VCARJ1.ANTS 
for supporting deliberation. 
Agents REJECT a proposal if deliberation leads them 
to believe that they know of a better option. For exam- 
ple, in (4-4) Kim rejects the proposal in (4-3), for pur- 
suing option-45, and proposes option-56 instead. The 
form of the rejection as a counter-proposal is based on 
observations about how rejection is communicated in 
naturally-occurring dialogue as codified in the GOLLAB- 
OR.ATIVE PLANNING PRINCIPLES \[37\]. 
Proposals i and 2 are inferred to be implicitly AC- 
CEPTED because they are not rejected \[37\]. If a pro- 
posal is ACCEPTED, either implicitly or explicitly, then 
the option that was the content of the proposal becomes 
a mutual intention that contributes to the final design 
plan \[27, 34, 30\]. 
The model of AWM discussed above plays a critical 
role in determining agents' performance. Remember 
that only salient beliefs can be used in means-end rea- 
soning and deliberation, so that if the warrant for a 
proposal is not salient, the agent cannot properly eval- 
uate a proposal. 
4.2 Varying Discourse Strategies 
Agents are parametrized for different discourse strate- 
gies by placing different expansions of discourse plans in 
their plan libraries. In Design-World the only discourse 
plans required are plans for PROPOSAL, REJECTION, 
ACCEPTANCE, CLAR.IFICATION, OPENING and CLOSING. 
The only variations discussed here are variations in the 
expansions of PROPOSALS. 
The All-Implicit strategy is an expansion of a dis- 
course plan to make a PROPOSAL, in which a PROPOSAL 
decomposes trivially to the communicative act of Pn.o- 
POSE. In dialogue (4), both Design-World agents com- 
municate using the All-lmplicit strategy, and propos- 
als are expanded to the PROPOSE communicative acts 
shown in 1, 2, and 3 in dialogue (4). The All-Implicit 
175 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
strategy never includes warrants in proposals, leaving 
it up to the other agent to retrieve them from memory. 
The Explicit-Warrant strategy expands the PRo- 
POSAL discourse act to be a WAR.R.ANT followed by a 
PROPOSE utterance \[33\]. 6 Since agents already know 
the point values for pieces of furniture, warrants are al- 
ways optional in the experiments here. In RST terms, 
an agent with the Explicit-Warrant strategy always 
chooses to MOTIVATE every proposal. For example (5-1) 
is a W'AR.R.ANT for the proposal in (5-2): 
(5) 1: TED: Putting in the green rug is worth 56. 
(say agent-ted agent-ben bel-10: score (option-10: 
put-act (agent-ted green rug room-l) 56)) 
2: TED: So, let's put the green rug in the study. 
(propose agent-ted agent-ben option-10: put-act 
(agent-ted green rug room-l)) 
3: BEN: Putting in the green lamp is worth 55. 
(say agent-ben agent-ted bel-34: score (option-33: 
put-act (agent-ben green lamp room-l) 55)) 
4: BEN: So, let's put the green lamp in the study. 
(propose agent-ben agent-ted option-33: put-act 
(agent-ben green lamp room-l)) 
The fact that the green rug is worth 56 points is mo- 
tivation for adopting the intention of putting the green 
rug in the study. Whether it is good motivation de- 
pends on what other options the agent knows about and 
what their utilities are. The Explicit-Warrant strategy 
models naturally occurring examples such as those in 
2 because the score information used by the hearer to 
deliberate whether to accept or reject the proposal is 
already mutually believed. 
4.3 Evaluating Performance 
Remember that we incorporate cognitive modeling into 
Design-World so that attentional capacity and the cost 
of various cognitive processes are parameters. To eval- 
uate PERFORMANCE, we compare the Explicit-Warrant 
strategy with the All-Implicit strategy while we vary 
agents' attentional capacity, and the cost of retrieval, 
inference and communication. Evaluation of the result- 
ing DESIGN-IIOUSE plan is parametrized by (1) COMM- 
COST: cost of sending a message; (2) INFCOST: cost 
of inference; and (3) I~.ETCOST: cost of retrieval from 
memory: 
PERFOllMANCE = Task Defined I~AW SCOaE 
- (COMMCOST × total messages) 
- (INFCOST × total inferences) 
- (I~.ETCOST × total retrievals) 
I~.AW SCOaE is task specific: in the Standard task 
we simply summarize the point values of the furniture 
pieces in each PUT-ACT in the final design. 
6The ordering of these two acts as given lets us have a 
simple control regime for processing utterances. The reverse 
ordering would require the agent to check whether it has 
more messages before processing the current message. 
We simulate 100 dialogues at each parameter set- 
ting for each strategy. Because the AWM model is 
probabilistic, the agents do not perform identically on 
each trial, and their performance over the 100 dialogues 
defines a performance distribution. In order to com- 
pare two strategies, we test whether the differences in 
the performance distributions are significant, using the 
Kolmogorov-Srnirnov (KS) two sample test \[31\]. 
A strategy A is BENEFICIAL aS compared to a strat- 
egy B, for a set of fixed parameter settings, if the 
difference in distributions using the Kolmogorov- 
Smirnov two sample test is significant at p < .05, 
in the positive direction, for two or more AWM 
settings. 
A strategy is DETRIMENTAL if the differences go in 
the negative direction. Strategies may be neither BEN- 
EFICIAL or DETRIMENTAL, there may be no difference 
between two strategies. 
5 Experimental Results on Providing 
Motivation 
This section discusses a few experimental results on the" 
Explicit-Warrant discourse strategy, which we compare 
with the All-Implicit strategy. Here we simply test the 
effect of whether the warrant is salient or not, whether 
there is any processing effort associated with retrieving 
the warrant from long term memory, and whether the 
cost of processing an utterance is high. 
Differences in performance between the Explicit- 
Warrant strategy and the All-Implicit strategy are 
shown via a DIFFERENCE PLOT such as figure 4. In fig- 
ure 4 performance differences are plotted on the Y-axis 
and AWM settings are shown on the X-axis. If the plot 
is above the dotted line for 2 or more AWM settings, 
then the Explicit-Warrant strategy may be BENEFICIAL. 
Each point represents the difference in the means of 100 
runs of each strategy at a particular AWM setting. 
5.1 Explicit Warrant reduces Retrievals 
Dialogues in which one or both agents use the Explicit- 
Warrant strategy are more efficient when retrieval has 
a cost. 
Figure 4 shows that the Explicit-Warrant strategy is 
detrimental at AWM of 3,4,5 for the Standard task, in 
comparison with the All-Implicit strategy, if retrieval 
from memory is free (KS 3,4,5 > .19, p < .05). This 
is because making the warrant salient displaces infor- 
mation about other pieces when agents are attention- 
limited. 
However, Figure 5 shows that Explicit-Warrant is 
beneficial when retrieval has an associated processing 
cost. By AWM values of 3, performance with Explicit- 
Warrant is better than All-Implicit because the war- 
rants intended to motivate the hearer and used by the 
hearer in deliberation are made salient with each pro- 
posal (KS for AWM of 3 and above > .23, p < .01). At 
AWM parameter settings of 16, where agents have the 
176 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
cost-ted-benbill-kimC:= 1 , I= 1 , FI=O 
oo 
  o 
..... , ,, ,~ 
0 2 .4, 8 10 1 1 1 
Affentlon/Worklng Memory 
Figure 4: If Retrieval is Free, Explicit-Warrant is detri- 
mental at AWM of 3,4,5: Strategy 1 of two Explicit- 
Warrant agents and strategy 2 of two All-Implicit 
agents: Task = Standard, commcost = 1, infcost = 
1, retcost -- 0 
ability to search a huge belief space for beliefs to be used 
as warrants, the saving in processing time is substan- 
tial. Again at the lowest AWM settings, the strategy 
is not beneficial because it displaces information about 
other pieces from AWM, However in Figure 5, in con- 
trast with Figure 4, retrieval has an associated cost. 
Thus the savings in retrieval balance out with the loss 
of raw score so that the strategy is not DETRIMENTAL. 
5.2 Explicit Warrant is detrimental if 
Communication is Expensive 
Finally we can amplify the results shown in Figure 4 by 
positing that in addition to there being no processing 
effort for retrieving from memory, processing the ad- 
ditional warrant utterance requires a lot of processing 
effort. Figure 6 shows that if communication cost is 10, 
and inference and retrieval are free, then the Explicit- 
Warrant strategy is DETR.IMENTAL (KS for AWM 1 to 
5 > .23, p< .01). This isl because the Explicit-Warrant 
strategy increases the number of utterances required to 
perform the task; it doubles the number of messages 
in every proposal. If communication is expensive com- 
pared to retrieval, processing additional warrant utter- 
ances is highly detrimental if there would be no effort 
involved in reirieving them, i.e., if they are essentially 
already salient. 
cost - ted-ben bill-kim C= 1 , I = 1 , R = 0.01 
$ 
°Y 
o ~ 
~_~._ o_o/°. ~ ........... 
I I I I I I t 116 
2 4 6 8 10 12 1.4, 
Attention/WorkinG Memory 
Figure 5: Retrieval costs: Strategy 1 is two Explicit- 
Warrant agents and strategy 2 is two All-Implicit 
agents: Task = Standard, commcost = 1, infcost = 
1, retcost = .01 
6 Implications for Text Planning 
The experiments reported in the previous section show 
that there is a direct relation between H's attentional 
state and the advisability of including warrants in a text 
plan. There are two ways in which we have modeled 
different aspects of H's attentional state in the experi- 
ments reported in the previous sections: we have var- 
ied the cost of retrieval, and we have varied the size of 
AWM. We can think of H's attentional state as compris- 
ing a (small) active working memory, and a larger long- 
term memory. We vary whether an agent has the ability 
to retrieve an item by varying the radius of AWM. We 
vary the amount of effor~ involved in retrieving an item 
by varying the cost of retrieval \[17, 24, 2\]. If a war- 
rant for the proposal is in short-term memory and can 
be accessed virtually cost-free, then figure 4 shows that 
generating the warrant explicitly can actually be detri- 
mental, since it can displace other information. The ef- 
fect is further magnified if communication is very costly, 
as shown in figure 6. (Cost of communication may by 
increased by a large number of factors, such as the lin- 
guistic complexity of the generated message, the fact 
that H is not fully competent in the language of com- 
munication, or noise in the channel of communication.) 
Thus, if S knows that information that can serve as a 
warrant is salient to H, then no warrant should be gen- 
erated. On the other hand, if a cost is in fact associated 
with retrieval, as in the experiment reported in figure 5, 
177 
7th International Generation Workshop • Kennebunkport, Maine ° June 21-24, 1994 
cost - ted-ben bill-kirn C= 10, I = 0, R = 0 
m 
0 2 4 6 8 10 1 14 1 
Art ention/VVorking Memory 
Figure 6: If Communication is Expensive: Communica- 
tion costs can dominate other costs in dialogues. Strat- 
egy 1 is two Explicit-Warrant agents and strategy 2 is 
two All-Implicit agents: Task = Standard, commcost = 
10, infcost = 0, retcost = 0 
then we can see that generating a warrant is beneficial, 
especially as the size of AWM increases and agents have 
the ability to access all of long term memory. 
We conclude that for a text planner to decide whether 
or not to communicate certain non-essential informa- 
tion, such as a warrant as part of a MOTIVATE relation, 
depends not just on the effect of the relation (which 
must of course match the discourse goal), but also on 
the attentional state of H, and on the other factors such 
as the cost of communication. A maximally efficient 
text planner will need to have access to: 
• a model of H's attentional state; 
• an algorithm that, given the attentional state model 
and additional parameters such as the costs of 
communication and retrieval determines whether H 
knows accessible information which can serve as a 
warrant. 
Here "accessible information" means either that the 
information is already salient or that it can be retrieved 
at a reasonable cost, given the costs of communication 
and retrieval. If we assume that the algorithm defines 
a binary predicate NOT-ACCESSIBLE, we can for- 
malize the RST relation MOTIVATE as a plan operator 
Motivation as given in Figure 7. The format follows 
the format of the plan operators given in \[22\], except 
that, for simplicity, we conflate the intentional and the 
rhetorical levels/ This plan operator is of course only 
meant to be suggestive, and we are not committed to 
any details. 
Of course, procedure NOT-ACCESSIBLE cru- 
cially relies on a proper model of H's attentional state 
and on an algorithm that accesses it. We intend to in- 
vestigate these issues in future work, but sketch some 
possible solutions here. An obvious candidate for the 
model is the AWM model used in the simulations it- 
self. (In fact, it is quite plausible that speakers use their 
own attentional state as a model for thai of the hearer.) 
The algorithm could then be defined very straightfor- 
wardly in terms of a three-dimensional boolean matrix, 
indexed on distance in memory, communication cost, 
and retrieval cost. The value for a given triple indicates 
whether or not information stored at this distance is ac- 
cessible. The values in the table are determined using 
the simulation environment. Presumably, this process 
weakly corresponds to the acquisition of proper text 
planning strategies by human agents. 
While an obvious candidate for the model of H's at- 
tentional state is the AWM model itself, certain aspects 
of this model do not exactly match some widely believed 
and intuitively motivated observations aboUt hierarchi- 
cal discourse structure \[10, 18\]. However, hierarchical 
structure interacts with attentional state in ways that 
have not been fully explored in the literature to date. In 
particular, if a discourse segment consisting of a nucleus 
with a hierarchically complex satellite that is extremely 
long, then a further satellite to the same nucleus may 
well require repetition of the nucleus \[35\]. Neither RST 
nor the model of \[10\] accounts for such effects. We con- 
clude that it is not a priori obvious that hierarchical 
structure contradicts our model. We will investigate 
this issue further. 
Throughout this paper, we have used the MOTIVATE 
relation in order to motivate our claims. However, simi- 
lar observations apply to other presentational relations, 
such as BACKGROUND, or I~VIDENCE as shown in 5: s 
(6) a. Clinton has to take a stand on abortion rights 
for poor women. 
b. He's the president. 
Here 6b is already known to the discourse partici- 
pants, but saying it makes it salient. The fact that 
it is already known makes S's attempt to convince H 
of 6a more likely to succeed. A discussion of the EVI- 
DENCE relation, however, is complicated by the need to 
find a proper representation of the degree of strength of 
ZMoore and Paris argue that for presentational relations 
(such as MOTIVATE), there is a one-to-one mapping between 
intentional and rhetorical structure. Therefore, conflating 
them is theoretically justified. 
SWe also believe that whether a known proposition is 
salient is an issue for supporting content-based inferences, 
and thus cognitive modeling may be required for text plan- 
ning of subject-matter relations as well. 
178 
7th International Generation Workshop.. Kennebunkport, Maine • June 21-24, 1994 
NAME: 
EFFECT; 
CONSTRAINTS: 
SATELLITE 
NUCLEUS: 
MOTIVATION) 
(DESIRE ?hearer (DO ?hearer ?act) ?utility-act) 
(AND (AGENT ?act ?hearer)) 
(UNREALIZED ?act) 
(NOT-ACCESSIBLE ?hearer (UTILITY ?act ?utility-act)) 
(BEL ?hearer (UTILITY ?act ?utility-act)) 
(BEL ?hearer (WANT ?speaker (DO ?hearer ?act))) 
Figure 7: The MOTIVATION plan operator 
belief, since utility theory is not appropriate as a repre- 
sentation of degree of belief \[9\]. In other work, we have 
developed a version of Gallier's theory of belief revision 
which takes into account the effect of limited working 
memory on agent's belief revision processes \[7, 8, 5, 35\]. 
This theory could be used for the RST EVIDENCE rela- 
tion. 
However, the use of a simple evaluation function for 
the representation of gradual strengths (of desire, belief, 
etc.) is in itself problematic. In this paper, we used the 
theoretical construct of utility as the basis for degree 
of desire for the MOTIVATION relation. However, obser- 
vations of human dialogue show that there are many 
evaluation functions in the real world that can be the 
basis for the MOTIVATION relation. Furthermore, these 
evaluation functions are incomparable and competing, 
as shown by (7), asserted by a speaker while walking to 
work: 
(7) I don't like going down that way. 
It may be shorter, but I don't like it. 
The speaker's desire for an aesthetic environment on 
her walk has in this case overridden her desire for the 
fastest route to work. However if she were late, the 
efficiency evaluation function might dominate. In the 
real world or in real text planning domains, the issue 
of multiple competing evaluation functions on poten- 
tial intended actions must be addressed. Observe that 
this issue is crucially related to the proper theoretical 
discussion of presentational relations. 
Finally, we would like to observe that there is ev- 
idence that the results presented here are domain- 
independent. The task that the agents are performing 
in Design-World is a simple task without any complex 
constraints, which we would expect to be a subcom- 
ponent of many other tasks. The model of limited re- 
sources we used was cogtjitively based, but the cost pa- 
rameters allow us to model different agent architectures, 
and we explored the effects of different cost parameters. 
The Explicit-Warrant strategy is based on simple rela- 
tionships between different facts which we would expect 
to occur in any domain, i.e. the fact that some belief 
can be used as a 'WAIl.RANT for accepting a proposal 
should occur in almost tiny task. Furthermore, our re- 
sults are confirmed by naturally occurring discourses in 
a wide variety of domains \[38, 4, 35\]. 
7 Conclusion 
In this paper, we have argued that a text planner that 
is based on the notion of communicative intention and 
on plan operators that explictly represent such inten- 
tions must also incorporate a sophisticated model of 
the heater's attentional state, and the ability to use 
this model in order to make decisions about whether 
or not to include optional information (satellites of pre- 
sentational relations). We have motivated this claim 
using naturally occurring dialogues, and by experimen- 
tal results from a simulation environment which imple- 
ments a simple, but psychologically plausible model of 
attentional state. Future work includes extending the 
analysis to the whole range of presentational relations, 
defining precisely a hearer model that can be used in 
text planning and an associated algorithm that can be 
used by the plan operators, and a theoretical investi- 
gation of the interaction between textual hierarchy and 
attentional state. 

References 
\[1\] J. R. Anderson and G. H. Bower. Human Associa- 
tive Memory. V.H. Winston and Sons, 1973. 
\[2\] A. Baddeley. Working Memory. Oxford University 
Press, 1986. 
\[3\] M. Bratman, D. Israel, and M. Pollack. Plans and 
resource bounded practical reasoning. Computa- 
tional Intelligence, 4:349-355, 1988. 
\[4\] A. Cawsey. Planning interactive explanations. In- 
ternational Journal of Man.Machine Studies, 1992. 
\[5\] A. Cawsey, J. Galliers, S. Reece, and K. S. Jones. 
Automating the librarian: A fundamental ap- 
proach using belief revision. TR-243, Cambridge 
Computer Laboratory, 1992. 
\[6\] J. Doyle. Rationality and its roles in reasoning. 
Computational Intelligence, November 1992. 
\[7\] J. R. Galliers. Belief revision and a theory of com- 
munication. TR 193, University of Cambridge, 
Computer Laboratory, 1990. 
\[8\] J. R. Galliers. Autonomous belief revision and 
communication. In P. Gardenfors, editor, Belief 
Revision, pages 220 - 246. CUP, 1991. 
\[9\] P. Gardenfors. Knowledge in flux : modeling the 
dynamics of epistemic states. MIT Press, 1988. 
\[10\] B.J. Grosz and C. L. Sidner. Attentions, intentions 
and the structure of discourse. Computational Lin- 
guistics, 12:175-204, 1986. 
\[11\] S. Hanks, M. E. Pollack, and P. R. Cohen. Bench- 
marks, testbeds, controlled experimentation and 
the design of agent architectures. AI Magazine, 
December 1993. 
\[12\] D. L. Hintzmann and R. A. Block. Repetition and 
memory: evidence for a multiple trace hypothesis. 
Journal of Experimental Psychology, 88:297-306, 
1971. 
\[13\] E. H. Hovy. Planning coherent multisentential 
text. In Proc. 26th Annual Meeting of the ACL, 
Association of Computational Linguistics, pages. 
163-169, Buffalo, 1988. ACL. 
\[14\] E. H. Hovy. Automated discourse generation using 
discourse structure relations. Artificial Intelligence 
Journal, 63:341-385, 1993. 
\[15\] D. Kahnemem, P. Slovic, and A. Tversky. Judg- 
ment under uncertainty : heuristics and biases. 
Cambridge University Press, 1982. 
\[16\] R. Kittredge, T. Korelsky, and O. Rambow. On the 
need for domain communication knowledge. Com- 
putational Intelligence, 7(4):305-314, 1991. 
\[17\] T. K. Landauer. Memory without organization: 
Properties of a model with random storage and 
undirected retrieval. Cognitive Psychology, pages 
495-531, 1975. 
\[18\] W. Mann and S. Thompson. Rhetorical structure 
theory: Description and construction of text struc- 
tures. In G. Kempen, editor, Natural Language 
Generation, pages 83-96. Martinus Nijhoff, 1987. 
\[19\] K. R. McKeown. Discourse strategies for generat- 
ing natural language text. Artificial Intelligence, 
27(1):1-42, September 1985. 
\[20\] G. A. Miller. The magical number seven, plus or 
minus two: Some limits on our capacity for pro- 
cessing information. Psychological Review, pages 
81-97, 1956. 
\[21\] J. D. Moore and C. L. Paris. Planning text for 
advisory dialogues. In Proc. 27th Annual Meeting 
of the Association of Computational Linguistics, 
Vancouver, 1989. ACL. 
\[22\] J. D. Moore and C. L. Paris. Planning text 
for advisory dialogues: Capturing intentional and 
rhetorical information. Computational Linguistics, 
19(4), 1993. 
\[23\] J. D. Moore and M. E. Pollack. A problem for 
rst: The need for multi-level discourse analysis. 
Computational Linguistics, 18(4), 1992. 
\[24\] D. A. Norman and D. G. Bobrow. On data-limited 
and resource-limited processes. Cognitive Psychol- 
ogy, 7(1):44-6, 1975. 
\[25\] C. L. Paris. Tailoring object descriptions to a 
user's level of expertise. Computational Linguis- 
tics, 14(3):64-78, 1988. 
\[26\] M. E. Pollack and M. Ringuette. Introducing the 
Tileworld: Experimentally Evaluating Agent Ar- 
chitectures. In AAAIgO, pages 183-189, 1990. 
\[27\] R. Power. Mutual intention. Journal for the The- 
ory of Social Behaviour, 14, 1984. 
\[28\] E. F. Prince. Toward a taxonomy of given-new 
information. In Radical Pragmatics, pages 223- 
255. Academic Press, 1981. 
\[29\] L. Ross and C. A. Anderson. Shortcomings in the 
attribution process: On the origins and mainte- 
nance of erroneous social assessments. In Judgment 
under uncertainty : heuristics and biases. Cam- 
bridge University Press, 1982. 
\[30\] C. Sidner. Using discourse to negotiate in col- 
laborative activity: An artificial language. AAAI 
Workshop on Cooperation among Heterogeneous 
Agents, 1992. 
\[31\] S. Siegel. Nonparametric Statistics for the Behav- 
ioral Sciences. McGraw Hill, 1956. 
\[32\] M. Solomon. Scientific rationality and human rea- 
soning. Philosophy of Science, September 1992. 
\[33\] D. D. Suthers. Sequencing explanations to enhance 
communicative functionality. In Proceedings of the 
15th Annual Meeting of the Cognitive Science So- 
ciety, 1993. 
\[34\] M. A. Walker. Redundancy in collaborative dia- 
logue. In Fourteenth International Conference on 
Computational Linguistics, pages 345-351, 1992. 
\[35\] M. A. Walker. Informational Redundancy and Re- 
source Bounds in Dialogue. PhD thesis, University 
of Pennsylvania, 1993. 
\[36\] M. A. Walker. Discourse and deliberation:testing 
a collaborative strategy. In Coling 94, 1994. 
\[37\] M. A. Walker and S. Whittaker. Mixed initiative 
in dialogue: An investigation into discourse seg- 
mentation. In Proc. 28th Annual Meeting of the 
ACL, page s 70-79, 1990. 
\[38\] B. Webber and A. Joshi. Taking the initiative in 
natural language database interaction: Justifying 
why. In COLING84, pages 413-419, 1982. 
\[39\] S. Whittaker, E. Geelhoed, and E. Robinson. 
Shared workspaces: How do they work and when 
are they useful? IJMMS, 39:813-842, 1993. 
