ATTENTION DURING ARGUMENT GENERATION AND 
PRESENTATION* 
Ingrid Zukerman, Richard McConachy & Kevin B. Korb 
School of Computer Science and Software Engineering 
Monash University 
Clayton, Victoria 3168, AUSTRALIA 
email: {ingrid,ricky, korb} @csse.monash.edu.au 
Abstract 
We describe the operation of our argumentation system, and discuss its use of attentional focus 
during both content planning and argument presentation. During content planning, attentional focus 
guides an abductive process used to build-up arguments. This process is applied to a model of a user's 
• beliefs and a normative model. During argument presentation, attentional focus supports the generation 
of enthymematic arguments. 
1 Introduction 
In this paper, we describe the operation of our argument-generation system, NAG (Nice Argument Gener- 
ator). We consider its content planning and argument presentation processes, and discuss the attentional 
mechanism used in both of these processes. 
Given a goal proposition, NAG's objective is to generate a nice argument for it, by which we mean 
one which achieves a balance between what is normatively justifiable and what persuades the interlocutor. 
To this end, NAG consults a normative model, which contains our best understanding of the domain of 
discourse, and a model of the user's beliefs. The main modules of the system are shown in Figure 1. 
The Strategist drives the argumentation process. During argument generation, it activates a generation- 
analysis cycle as follows (Section 3). First, it invokes the Attentional Mechanism (Section 4) to activate 
salient propositions, which are used to construct an initial Argument Graph for an argument, or to extend an 
already existing Argument Graph. (An Argument Graph is a network with nodes that represent propositions, 
and links that represent the inferences that connect these propositions.) The Strategist then calls the Gener- 
ator to continue the argument building process (Section 5). The Generator in turn fleshes out the Argument 
Graph by activating Reasoning Agents, which consult several sources of information, and incorporating the 
inferences and propositions returned by these agents into the Argument Graph. This Argument Graph• is 
returned to the Strategist, which passes it to the Analyzer in order to evaluate its niceness and check for 
reasoning flaws (Section 6). If the Analyzer indicates that the argument is not nice enough, i.e., there is not 
sufficient belief in the•goal in the user or the normative model, then the Strategist re-activates the Generator 
in order to find further support for the premises in the argument, and so on.- The generation-analysis cycle 
continues until a sufficiently nice Argument Graph is generated. This graph is then passed to the Argu- 
ment Presenter, which selects an argumentation strategy and determines propositions to be removed from 
the argument, aiming to produce a simpler, enthymematic argument. After each removal, the Presenter acti- 
vates the Analyzer to check whether theargument is still nice and the Attentional Mechanism to determine 
whether the argument can still be followed by the user. 
Thus, the Attentional Mechanism is used in two different stages of the argumentation process. During 
argument generation, it focuses the argument construction process on concepts which are related to the 
"This research was supported in part by Australian Research Council grant A49531227. 
148 
I 
I 
I 
I 
I 
I 
il 
~4 
i i 
I 
I 
! 
I 
Argument :r~a~ffnaa~:l I Argument \[ 
Generator" A~ents "l \[ Analyzer I .......... I Argument I _ _ _ I Argument 
Argument~,,~~ProG°i~ons A~aly~gum~ alysis 
Goal _ I Argument IPresentation Argument Graph l Argument I Argument 
Proposition "\[ Strategist ~ ........... ~1 Presenter t ~ USER 
i "-,q Attentional ~ i . 
User Argument/Inquiry/Goal Proposition 
Figure 1: System Architecture 
goal, avoiding some distractions. During argument presentation, the Attentional Mechanism supports the 
generation of enthymematic arguments. 
2 Related Research 
Charniak and Goldman (1993) describe a Bayesian plan recognition system that uses marker passing as a 
method for focusing attention on a manageable portion of the space of all possible plans. This is analo- 
gous to the way in which NAG uses spreading activation to focus on a small portion of the available data 
during the content planning process. Walker (1996) points out the effect of attentional focus on discourse 
comprehension. This effect is taken into consideration by NAG during argument presentation. 
The approach of "interpretation as abduction" used in \[Hobbs et al., 1993\].aims to recover the premises 
and inferential links which lead to an argument's conclusion. This is similar to NAG's argument analysis. 
The most important difference between NAG and the work by Hobbs et al., in addition to NAG being a 
system that reasons under uncertainty, is that NAG performs both analysis and generation. A generative 
system based on the work of Hobbs et al. is described in \[Thomason et aL, 1996\]. This system deals with 
what can be readily inferred, and so deleted, during communication, but the generated discourse does not 
argue in favour of a proposition. Mehl (1994) describes a system which can turn an existing fully explicit 
argument into an enthymematic one, but it cannot generate an argument from constituent propositions. The 
system described in \[Horacek, 1997\] generates its own arguments and presents them enthymematically. 
However, neither process models explicitly a user's attentional state. 
Like NAG, the systems described in \[Reed and Long, 1997, Huang and Fiedler, 1997\] consider focus of 
attention during argument presentation. NAG differs from these systems in that NAG also uses attentional 
focus to guide the content planning process. In addition, Reed and Long consider attention in order to gener- 
ate additional information that makes a concept salient, and Huang and Fiedler use a limited implementation 
of attentional focus to select which step in a proof should be mentioned next. In Contrast, NAG uses atten- 
tional focus during argument presentation to convert a fully explicit argument into an enthymematic one. 
Finally, Fehrer and Horacek (1997) take advantage of mathematical properties to structure certain types of 
mathematical proofs. They model a user's inferential ability by means of specialized substitution rules, but 
offer no mechanism (such as attention in NAG) to limit the number of applications of their rules. 
3 The Generation-Analysis Cycle 
NAG receives the following inputs: (1) a proposition to be argued for, (2) an initial argument context, and 
(3) two target ranges of degrees of belief to be achieved (one each for the normative model and the user 
model). The argument context is composed of salient propositions and concepts appearing in the discussion 
149 
preceding the argument or in the current Argument Graph. The initial argument context, contexto, is com- 
posed of salient propositions and concepts mentioned in the preamble to the argumen t plus the argument's 
goal. The degrees of belief to be achieved are expressed as ranges of probabilities, e.g., \[0.5, 0.6\], in order to 
be able to represent a variety of goals, e.g., inducing indifference or assent. Two target ranges are required 
since the degree of belief to be reached by the user may differ from that reached by the system. 
When constructing an argument, NAG relies on two collections of information: a normative •model 
composed of different types of Knowledge Bases (KBs), and a user model also composed of different types 
of KBs, which represent the user's presumed beliefs and inferences. A KB represents information in a 
single format, e.g., semantic network (SN), Bayesian network (BN), rule-based system, or database. The 
KBs in the normative and user models are consulted by specialist Reasoning Agents which are activated 
when trying to fill gaps in a partial argument (Section 5). The KBs in the user model are consulted to make 
an argument persuasive for the target audience, while the normative KBs are consulted so that the generated 
argument is normatively correct. This distinction between the normative correctness and the persuasiveness 
of an argument enables the system to control the extent to which it will sacrifice normative correctness in 
order to be persuasive. During argument generation, relevant material from several KBs may need to be 
combined into a common representation. We have chosen BNs for this purpose because of •their ability 
to represent normatively correct reasoning •under uncertainty, and because simple alterations of the normal 
Bayesian propagation rules allow us to model various human cognitive phenomena \[Korb et al., 1997\]. • 
The content planning process produces an Argument Graph which starts from admissible premises and 
ends in the goal proposition. Admissible premises are normatively acceptable propositions that are believed 
by NAG and are either believed the user (sufficiently for the argument to work) or assented to by the user. 
The resultant Argument Graph is then passed to the Presenter. The argument generation process is imple- 
mented by the following algorithm, which is executed by the Strategist. 
Generation-Analysis Algorithm 
1. i +-- 0. 
2. Clamp nodes in the current context, contexti, and perform spreading activation. This yields an Argu= 
ment Graph containing: the clamped nodes, the activated nodes (whose activation exceeds a thresh- 
old), and the links connecting these nodes. (Section 4) 
3. Identify new subgoals in the current Argument Graph. These are nodes which have not been previ- 
ously passed to the Reasoning Agents, and have a path to the goal in the Argument Graph or a high 
level of activation (higher than a subgoaling threshold). 
4. Pass the argument subgoals identified in Step 3 to the Generator, which adds the new information 
returned by its Reasoning Agents to the current Argument Graph. (Section 5) 
5. Pass the Argument Graph generated in Step 4 to the Analyzer for evaluation. (Section 6) 
6. If the Analyzer reports that the current Argument Graph is sufficiently nice, then pass the Argument 
Graph to the Presenter. Otherwise, continue. (Section 7) 
7. i+-i+l. 
8. contexti +-- contexti_l + new nodes connected to the goal during cycle i-1 + 
highly activated nodes that are not connected to the goal. 
9. Go to Step 2. 
4 Simulating Attention 
NAG uses a hierarchical semantic network built on top of BNs in both the user and normative models to cap- 
ture connections between the items mentioned in the discourse (Figure 2 illustrates a three-level semantic- 
Bayesian network). The semantic network portion (upper levels of the pyramid) and the BN portion (base 
of the pyramid) are used by NAG to simulate attentional focus in each model. 
150 
I 
i 
i 
.I 
i 
I 
I 
I 
I Higher level 
/1~ ~ .,concepts like 
I Lower level / / I V,, . ~, concepts like / / ~,/~arge ammaJ 
'dinosaur' J ~ Semantic 
Semantic ~~7~-~---~ 2rid layer 
1st layer - ~- o 
^ - ~ Bayesian 
Proposition, e.g., \[widespread iridium deposited 65 mil yrs BC\] 
Figure 2: Semantic and Bayesian Networks 
The Attentional Mechanism receives as input a context consisting of a set of salient objects. For content 
planning, the initial context consists of the goal proposition and salient concepts and propositions mentioned 
in the preamble of the argument; as the content planning process progresses, the context is extended with 
concepts and propositions included in the argument. For argument presentation, the context initially contains 
the propositions of the first sub-argument to be presented under a particular presentation strategy, and moves 
on to different propositions as the presentation of the argument pro~esses (Section 7). 
We use activation with decay \[Anderson, 1983\], spreading from the current context, to model the focus 
of attention. All items in the semantic-Bayesian networks which achieve a threshold activation level during 
the spreading activation process are brought into the current span of attention. This process passes activation 
through the pyramidal semantic-Bayesian networks, each node being activated to the degree implied by the 
activation levels of its neighbours, the strength of association to those neighbours, and its immediately prior 
activation level (vitiated by a time-decay factor). The spreading activation process ceases when an activation 
cycle fails to activate any new node. By these means we have a direct implementation of attention which 
we use to identify portions of the pyramidal semantic-Bayesian networks that are strongly related to the 
argument being built or presented. 
During content planning, spreading activation is applied to both the semantic-Bayesian pyramid in the 
user model and that in the normative model. This supports the retrieval of information that is semantically 
connected, and hence likely to be useful for building an argument, from each of the models. Further, 
the items in the initial context and in subsequent, extended contexts are clamped, since they are used for 
reasoning, and their level of activation should not fade as the system reasons about the argument. In contrast, 
during argument presentation, spreading activation is performed only in the semantic-Bayesian pyramid in 
the user model, and salient information items are not clamped. This is because we are trying to anticipate 
the effect of the information being presented on the addressee, and information presented earlier will fade 
from the addressee's focus of attention \[Walker, 1996\]. 
The Attentional Mechanism offers the following advantages. During content planning, it allows NAG 
to restrict its search to information semantically or evidentially connected with the propositions already in 
focus, and it enables NAG to analyze its arguments with respect to just these propositions (Section 6). During 
argument presentation, it supports the generation of enthymematic arguments by assessing the omission of 
propositions that are in the addressee's span of attention (as a result of the presentation of other propositions). 
5 Argument Extension 
The Generator activates the Reasoning Agents to collect information relevant to each of the subgoals in the 
current Argument Graph (these subgoals were identified in Step 3 of the Generation-Analysis algorithm). 
The Reasoning Agents determine the relevance of a piece of information to a goal proposition by applying a 
151 
I 
~V5 
I ~'~in~is 
:; :exposure, 
/% 
I r¢ptihrs :have 
Nn ~2 
expl~oc throws l widespread 
~r~.into~ :1 i. .i~,,m 
.atmosphere :J , is ~o=tea 
I 
[~,~ :strike forms : 
N~ 
I 
I 
N~ 
(a) Initial Argument Graph 
N7 
material 
obscures sun 
i 
g herbivorous N~ 
animals start 
N5 
I large ankna~ 
suffer .from 
: exposure,: 
Ns 
I reptile s hav( 
Nlo N. l ~2 ~3 
I shocked quartz -j----j debHs into 1he ~ , iridium " ~~~._~'~ 
J is deposited I J alto°sphere I I isdep osited I |~--~>~ 
I~ I " a.ro~:.. 11 giant crater is I 
I~~'~t st~ke forms•• I-'q discovered in the I 
i~ff3 l J g~anterater ] t Gulf of Mexico ] 
• (b) Argumen t Graph after one cycle 
Figure 3: Argument Graphs for the Asteroid Example during Content Planning 
procedure that is appropriate for the KB under consideration. For instance, if the goal proposition matches 
the consequent of a rule in a rule-based system, the Reasoning Agent for rule-based systems selects this 
rule. The relationships found by the Reasoning Agents represent inferences. The strength of each inference 
is represented by means of conditional probabilities which reflect the degree of belief in the consequent of 
an inference given full belief in its antecedents. 
Example 
Throughout this paper, we consider the generation of an argument for the goal proposition "A large iridium- 
rich asteroid struck Earth about 65-million-yearsBC;' preceded by the preamble "Approximately 65-million- 
years BC the dinosaurs, large reptiles that dominated the Earth for many millions of years, became extinct". 
Initially, the goal proposition and the preamble activate any propositions containing two or more of the 
italicized concepts, i.e., nodes Nr, N13 and N15 (the goal node) in Figure 3(a) (shown in dark grey boxes). 
After clamping the nodes that correspond to this discourse context and performing spreading activation, 
additional nodes become activated in the semantic and Bayesian networks. All the nodes whose activation 
• level exceeds a threshold are added to the Argument Graph. In this example, this yields the nodes shown in 
• light grey boxes in Figure 3(a). The links between the nodes in Figure 3(a) were obtained from the BN, but 
the activation of these nodes involved spreading activation through both the BN and the SN. 
Since none of the nodes in the current Argument Graph (Figure 3(a)) have been passed to the Reasoning 
Agents, the following nodes are passed to these agents (through the Generator): those in the subgraph 
containing the goal node (Nil, N15 and N16), plus the two clamped (highly active) nodes in the graph 
fragments not connected to the goal node (N6 and Nz3). The information returned by the Reasoning Agents, 
which is either causally• or evidentially connected to the nodes passed to the Generator, is then incorporated 
into the Argument Graph (Figure 3(b)). Some of the nodes found by these agents have already been activated 
through spreading activation (shown in light grey in Figures 3(a) and 3(b)), while others are new to the 
Argument Graph (shown in white in Figure 3(b)). In addition, the Reasoning Agents added a new link 
between the previously activated nodes Nlz --~ N12. 
152 
! 
I 
1 
I 
I 
1 
1 
I 
1 
I 
1 
I 
I 
I 
I 
I 
I 
I 
I . 
NI N~ 
~-mo-s~-'pI-a:l--i_.~l. herbivorous 
species die ' I animals starve k 
L___-~ .... , -- 
[-:_ -. :"-- - -I I. large animals :.I 
l".~lrlt7 DecorIP-.S ~ I : -'~;':'-~ " "i:~'t 
.......... ~\ I .. :~1 
I ma,e o, 
I 
.?2::: 
d~sease~attac: ~ 
all d/,,osa,,~s 
i 
Nio N. | NI2 ~ 
[ .,des~'ead I _ lexp'osion throwsI I widespread . ~ 
shocked quartz ~ debris into the I'~ iridium 
is deposited I I atmosphere I I is deposited 
,'v'i4 i N/, l N16 N/, 
,~o,..~ ~ ~.,. ,,,,,.o,; I~~~1 I asteroid strike I I giant crater is I 
; o~ s!~o~ked , I~-~k'~ stake forms I-~'-Idiscovered inthe I 
, n ' I . giant crater ] I Gulf of Mexico I 
Qt28~2 I~ ~OtJ, 
-:" ....... ~ l 
(a) Argument Graph passed to the Presenter 
I 
, most plant 1 
i' specie~ die t i 
.... ~ ..... i 
Nz * N6 
i . 
i Earth becomes 
L____ I ..... \ : / .:...,,-. ................. 
,,v7 i ',, N,8 
N. I N3 
rd ~'°si°nthr°ws] 
ebris into the .~ :l 
atmosphere !~.:1 
N. ~ N~6 87 
t~~ I: ::asteroid '/I giant crater is I 
~~ :sb:ike forms:: I--D'I discovered in the I 
~,~~ ...... I:::glant ctater :: 11 G.,, o, Mexico I 
(b) Argument Graph after pruning 
Figure 4: Argument Graphs for the Asteroid Example during Presentation 
6 Argument Analysis 
The process of computing the anticipated belief in a goal proposition as a result of presenting an argument 
starts with the belief in the premises of the Argument Graph and ends with a new degree of belief in the 
goal proposition. The Analyzer computes the new belief in a proposition by combining the previous belief 
in it with the result of applying the inferences which precede this proposition in the Argument Graph. This 
belief computation process is performed by applying Bayesian propagation procedures to the Bayesian sub- 
network corresponding tothe current Argument Graph in the user model and separately to the subnetwork 
corresponding to the current Argument Graph in the normative model. 
After propagation, the Analyzer returns the following measures for an argument: its normative strength, 
which is its effect on the belief in the goal proposition in the normative model, and its effectiveness, which 
is its effect on the user's belief in the goal proposition (estimated according to the user model). Of course, 
an argument's effectiveness may be quite different from its normative strength. When anticipating an argu- 
ment's effect upon a user, NAG takes into account three cognitive errors that humans frequently succumb 
to: belief bias, overconfidence and the base rate fallacy [Korb et aL, 1997]. 
If the normative strength or effectiveness of the Argument Graph is insufficient, another cycle of the 
Generation-Analysis algorithm is executed, gathering further support for propositions which have a path 
to the goal or have a high activation level (Step 3 of the Generation-Analysis algorithm). In this manner, 
NAG combines goal-based content planning with the associative inspection of highly active nodes. After 
integrating the new sub-arguments into the Argument Graph (Step 4), the now enlarged Argument Graph is 
again sent to the Analyzer (Step 5). Hence, by completing additional focusing-generation-analysis cycles, 
Argument Graphs that are initially unsatisfactory are often improved. 
Example - Continued 
The argument that can be built at this stage consists of nodes NT, N10-N13 and N15-N17. However, only 
N13 is admissible among the potential premise nodes. Thus, the anticipated belief in the goal node in both 
the normative and the user model falls short of the desired ranges. This is reported by the Analyzer to the 
Strategist. Nodes NT, Nlo-N12, N16 and N17 are now added to the context (which initially included N6, 
153 
N13 and N15), and the next cycle of the Generation-Analysis algorithm starts. This process continues until 
the Analyzer reports that the belief in the goa ! proposition is inside the target ranges in both the user and 
the normative model. In this example, this happens after two additional generation-analysis cycles, which 
activate nodes N1, N4 and N14 among others (in dashed boxes in Figure 4(a)). 
7 Argument Presentation 
After a successful Argument Graph has been built, the argument must be structured for presentation to the 
user. This involves selecting an argumentation strategy and pruning unnecessary propositions. These are 
propositions that lend little support to the belief in the goal or support the goal beyond what is required 
(removed by probabilistic pruning), and also intermediate propositions which will be in the addressee's 
focus of attention as a result of information presented earlier, and hence may be omitted (removed by 
semantic suppression). After pruning, the Analyzer (Section 6) is invoked to check whether the belief in 
the goal proposition in the now smaller Argument Graph is still within the target ranges; the Attentional 
Mechanism (Section 4) is invoked to check whether the propositions in the argument are still in focus when 
they are needed. The following greedy algorithm implements this process. 
Presentation Algorithm 
1. Determine an argumentation strategy. (Section 7.1) 
2. Traverse the Argument Graph according to the strategy selected in Step 1, invoking the Attentional 
Mechanism to determine the activation level of intermediate propositions. (Section 7.2) 
3. Alternate between probabilistic pruning and semantic suppression until time runs out or until no prun- 
ing has been successful in the last N consecutive iterations. (Section 7.3) 
At present, the resulting argument is in the form of propositions interleaved with causal or eviden- 
tial relations.• A graphical interface which allows users to build and receive arguments in an annotated 
network form (similar to that Shown in Figure 4(b)) is in preparation. We are also considering the appli- 
cation of microplanning operators for generating paraphrases and aggregations, such as those described in 
\[Huang and Fiedler, 1997\], prior to rendering an argument in English. 
7.1 Determining an Argumentation Strategy 
The argumentation strategy determines the order of presentation of the propositions. Two basic argumen- 
tation strategies are premise-to-goal and goal-to-premise. A premise-to-goal argument simply starts at the 
premises, goes on to support intermediate propositions, and eventually reaches the goal. Two types of goal- 
to-premise arguments are hypothetical, which assume the goal and go from it to the premises (which are suf- 
ficiently believed), and reductio ad absurdum, which start from the negation of the goal and reach a contra- 
diction. NAG selects an argument presentation strategy by examining separately the impact of each individ- 
ual line of reasoning contributing to the belief in the goal in the Argument Graph \[McConachy et al., 1998\]. 
7.2 Traversing the Argument •Graph 
Each argumentation strategy yields a traversal order for the Argument Graph. The premise-to-goal strategy 
is implemented by means of a post-order traversal modified by policies which select the order of presentation 
of sub-arguments. We consider two policies for presenting sub-arguments for any conclusion inside the 
Argument Graph: collective and individual-sequential. In the collective policy, all the current immediate 
antecedents of a conclusion, i.e., those that remain after pruning (Section 7.3), are mentioned immediately 
prior to mentioning this conclusion. For examPle, this policy yields the following argument for node E in 
the Argument Graph in Figure 5(a): "A is evidence for B, which together with D strongly supports E. ''1 
1The phrases that convey the causal and evidential relations in the arguments depend on the numerical values in the conditional 
probability •matrices in the normative BN, the types of the links and the direction of the argument compared to that of the links. 
154 
I 
I 
! 
! 
! 
I 
I 
I 
I 
I 
I 
! 
I 
i 
I 
I 
! 
! 
w-" 
A ~ N6~ Nf~.. 
A ~B.~... B--"~-D...~. A¢~ )N4~ N7~ NII"~ NI5 
c-." H N6~-- N~ ~,f . N,~ 
(a) (b) (c) 
Figure 5: Sample Argument Graphs 
This policy is used when all the antecedents provide similar levels of support for the consequent or when 
the antecedents are not conditionally independent. Since the collective policy requires all antecedents to 
be mentioned prior to stating the consequent, intermediate conclusions drawn earlier in the argument are 
mentioned again prior to their actual usage. For example, in the following argument, which corresponds to 
the Argument Graph in Figure 5(b), proposition D is repeated prior to its use in arguing H: "A, B and C 
cause D. E causes F which in turn causes G. G and D cause H." In the individual-sequential policy, the 
effect of one antecedent on the consequent is mentioned first followed by the effect of the other antecedents, 
e.g., "A is evidence for B, which strongly supports E. D also supports E." •This policy is used when one 
antecedent is conditionally independent from the others, and provides a level of support that is quite different 
from that provided by the others. If an antecedent is selected according to the individual-sequential policy, 
the policy selection considerations are again applied to the remaining antecedents. 
When the collective policy is applied, the sub-arguments are ordered according to the largest-subgraph- 
first heuristic, which is designed to reduce the number of intermediate conclusions that fade from the focus 
of attention (due to decay in their activation). This heuristic states that among several sub-arguments for 
the same consequent, the longer sub-arguments, i.e., those containing the largest number of propositions, 
should be presented before the shorter ones. 
7.2.1 Activation Levels in the Argument Graph 
After determining a traversal Order for the Argument Graph, the activation level of each proposition is 
calculated based on the activation spread from the propositions which NAG has planned to mention so far 
(taking into account a time decay factor). As each proposition is "mentioned", activation spreads from it in 
a pattern which depends on the level of activation of its neighbours and the strength of its association with 
its neighbours (Section 4). 2 For instance, when a highly connected node is activated, it in turn activates a 
large number of nodes, but the activation spread to each node is relatively small, leading to a quick decay in 
the activation process. This process forms a small but densely populated "bubble of attention" around the 
activated node \[Taylor, 1998\]. In contrast, when a node with a few connections is activated, the activation 
spread to each neighbouring node is stronger, hence the process takes longer to decay, yielding a larger but 
more sparsely populated bubble of attention. Such larger bubbles normally encompass several reasoning 
steps in the Argument Graph, thereby enabling NAG to remove intermediate conclusions (Section 7.3). 
7.3 Pruning the Argument Graph 
The Presenter interleaves probabilistic pruning with semantic suppression. It starts with probabilistic prun- 
ing, which iteratively removes from the Argument Graph premises that have a relatively small contribution 
to the belief in their consequents, and also removes entire lines of reasoning which provide more support 
for the goal than what is Strictly necessary in order to achieve a belief within the target ranges. After each 
removal, the Analyzer is called to check whether the belief in the goal proposition is still within the target 
ranges. In addition, affected subgraphs in the Argument Graph are reordered according to the policies men- 
tioned in Section 7.2 (reordering is necessary if their relative sizes have changed as a result of the removal of 
2Nothing is actually mentioned at this stage. However. in order to anticipate the effect of an argument, NAG pretends that 
propositions have actually been mentioned. 
•155 
propositions), and the Attentional Mechanism is called to determine whether the remaining propositions still 
have sufficient activation for the addressee to be able to follow the resulting argument. Probabilistic pruning 
fails when the Analyzer reports that the anticipated belief in the goal is outside a target range or when the 
level of activation of a proposition falls below a certain threshold. In this case, the last removed proposi- 
tion is reinstated, and semantic suppression is activated. This process iteratively omits from the Argument 
Graph intermediate conclusions whose level of activation was already high before they were "mentioned", 
so long as the probabilities of the links from their antecedent to their consequent are sufficiently strong. The 
rationale for this omission is that propositions which are readily inferred, because they have a high level of 
activation and the user attributes them a very high probability on the basis of a simple inference from just 
mentioned premises, will continue to contribute much of their support to the conclusions that follow them. 
For instance, if the bubble of attention around node E in the Argument Graph in Figure 5(b) includes nodes 
F and G, then node F may be omitted, yielding an argument such as "A, B and C cause D. E leads to 
G, which together with D causes H". Semantic pruning fails when the level of activation of any required 
subsequent proposition falls below a threshold. In this case, the last removed proposition is reinstated, prob- 
abilistic pruning is reactivated, and the interleaved pruning continues. These pruning processes continue 
until time runs out Or until both pruning methods have failed several consecutive times. 3 
Example - Continued 
After traversing the Argument Graph according to the premise-to-goal strategy, pruning is performed. Prob- 
abilistic pruning removes node N9 in Figure 4(a) as its probability is very low, and hence so is its effect on its 
consequent N6. The branch containing nodes Nlo and NI4 is removed despite the high probability of these 
nodes, because after their omission the probability of the goal proposition is still withinthe target ranges. 
The branch containing Ns and N6 is removed for the same reason (without removing node N6, since it is 
still connected to other nodes in the Argument Graph). Probabilistic pruning fails after this step, so semantic 
suppression is attempted, which removes nodes N2 ~ N3 and N12. Both types of pruning fail after this step, 
yielding the Argument Graph in Figure 4(b). The expanded tree which represents the final ordering of this 
Argument Graph appears in Figure 5(c). The post-order traversal of this tree yields an argument which may 
be roughly paraphrased as follows. 
Dinosaurs became extinct about 65 million years ago. This could have been caused in part by most 
plant species dying. This could also have been caused in part by reptiles having trouble breeding. 
This and mostplant species dying was likely caused by the Earth becoming much cooler, which could 
have been caused by material obscuringthe Sun. 
Many 65 million-year-old iridium deposits (which were found) and material obscuring the Sun could 
have been caused by an explosion throwing debris into the atmosphere. 
A giant crater which was discovered in the Gulf of Mexico was likely caused by an asteroid strike 
forming a giant crater. This and the explosion debris in the atmosphere were likely caused by a large 
iridium-rich asteroid striking Earth 65 million years BC. 
8 Conclusion 
NAG uses items in the user's focus of attention to guide a series of generation-analysis cycles which result 
in the generation of an Argument Graph that is both normatively acceptable and persuasive. Attentional 
focus also supports the generation of enthymematic arguments. 
NAG was written in Common Lisp. It was tested on five sample scenarios with KBs containing up to 
200 propositions. The simulation of attention Via spreading activation generally led to a significant speed- 
up in content planning times, with little effect on the generated arguments \[Zukerman et al., 1998\]. Content 
planning times slowed down when extremely slow decay factors and low activation thresholds were used 
3An alternative approach t O combining probabilistic pruning and semantic suppression is described in \[McConachy et aL, 1998\]. 
156 
I 
I 
I 
II 
I 
,I 
:l 
! 
! 
! 
i| 
i1 
i 
I 
i 
,I 
I 
i 
and when extremely fast decay factors and high activation thresholds were used. The former incorporated 
into the Argument Graph nodes that were only marginally related to the goal, while the latter incorporated 
too few nodes, resulting in essentially a goal-based search for an argument. 
We conducted a preliminary Web-based evaluation of NAG's content planning component. This evalua- 
tion consisted of giving respondents pre-test questions regarding certain propositions in an asteroid argument 
similar to that discussed in this paper, and then presenting them with hand-generated renditions of an aster- 
oid argument which took into account the responses to these questions. A post-test was used to determine 
the effect of the argument. This test showed a clear tendency among the respondents to shift belief towards 
the targets as a result of NAG's argument. A more rigorous evaluation of the content planning component 
and an evaluation of the argument presentation component will be performed in the next months. 

References 
\[Anderson, 1983\] Anderson, J. R. (1983). The Architecture of Cognition. Harvard University Press, Cambridge, 
Massachusetts. 
\[Charniak and Goldman, 1993\] Charniak, E. and Goldman, R. P. (1993). A Bayesian model of plan recognition. 
Artificial Intelligence, 64(1):50-56. 
\[Fehrer and Horacek, 1997\] Fehrer, D. and Horacek, H. (1997). Exploiting the addressee's inferential capabilities 
in presenting mathematical proofs. In IJCAI97- Proceedings of the Fifteenth International Joint Conference on 
Artificial Intelligence, pages 959-964, Nagoya, Japan. 
\[Hobbs et al., 1993\] Hobbs, J. R., Stickel, M. E., Appelt, D. E., and Martin, P. (1993). Interpretation as abduction. 
Artificial Intelligence, 63(1-2):69-142. 
\[Horacek, 1997\] Horacek, H. (1997). A model for adapting explanations to the user's likely inferences. User Model- 
ing and User-Adapted Interaction, 7(1): 1-55. 
\[Huang and Fiedler, 1997\] Huang, X. and Fiedler, A. (1997). Proof verbalization.as an application of NLG. In 
HCAI97 - Proceedings of the Fifteenth International Joint Conference on Artificial Intelligence, pages• 965-970, 
Nagoya, Japan. 
\[Korb et al., 1997\] Korb, K. B., McConachy, R., and Zukerman, I. (1997).• A cognitive model of argumentation. 
In Proceedings of the Nineteenth Annual Conference of the Cognitive Science Society, pages 400--405, Stanford, 
Califomia. 
\[McConachy et al., 1998\] McConachy, R., Korb, K. B., and Zukerman, I. (1998). Deciding what not to say: An 
attentional-probabilistic approach to argument presentation. In Proceedings of the Twentieth Annual Conference of 
the Cognitive Science Society, Madison, Wisconsin. 
\[Mehl, 1994\] Mehl, S. (1994). Forward inferences in text generation. • In ECAI94 - Proceedings of the Eleventh 
European Conference on Artificial Intelligence, pages 525-529, Amsterdam, The Netherlands. 
\[Reed and Long, 1997\] Reed, C. and Long, D. (1997). Content ordering in the generation of persuasive discourse. In 
IJCAI97- Proceedings of the Fifteenth International Joint Conference on Artificial Intelligence, pages 1022-1027, 
Nagoya, Japan. 
\[Taylor, 1998\] Taylor, J. (1998). Constructing the relational mind. Psyche, 4. http://psyche.cs.monash.edu.au/v4. 
\[Thomason et al., 1996\] Thomason, R. H., Hobbs, J. R., and Moore, J. D. (1996). Communicative goals. In Pro- 
ceedings of ECAI-96 Workshop - Gaps and Bridges: New Directions hi Planning and NLG, pages 7-12, Budapest, 
Hungary. 
\[Walker, 1996\] Walker, M. A. (1996). The effect of resource limits and task complexity on collaborative planning in 
dialogue. Artificial Intelligence, 85:181-243. 
\[Zukerman et al., !998\] Zukerman, I., McConachy, R., and Korb, K. B. (1998). Bayesian reasoning in an abductive 
mechanism for argument generation and analysis. In AAAI98 - Proceedings of the Fifteenth National Conference 
on Artificial Intelligence, Madison, Wisconsin. 
