Proceedings of the Fourth International Natural Language Generation Conference, pages 114–121,
Sydney, July 2006. c©2006 Association for Computational Linguistics
 
Generation of Biomedical Arguments for Lay Readers  
 
Nancy Green 
Department of Computer Science 
University of North Carolina Greensboro 
Greensboro, North Carolina 27402-6170 USA 
nlgreen at uncg.edu 
 
 
 
Abstract 
This paper presents the design of a discourse 
generator that plans the content and organi-
zation of lay-oriented genetic counseling 
documents containing arguments, and an 
experiment to evaluate the arguments. Due 
to the separation of domain, argument, and 
genre-specific concerns and the methodol-
ogy used for acquiring a domain model, this 
approach should be applicable to argument 
generation in other domains. 
1 Introduction 
The goal of our research is to develop methods by 
which intelligent systems can help lay audiences to 
understand biomedical and other kinds of scientific 
arguments. We have been studying how one type 
of lay-communication and biomedical-domain ex-
pert, the genetic counselor, presents written argu-
ments in patient letters, standard documents 
summarizing information and services provided to 
the client (Baker et al., 2002). Clinical genetics 
involves causal probabilistic reasoning, e.g., diag-
nosis of a genetic basis for a health problem or 
prediction of inheritance risks. The patient letter is 
designed to document the experts’ reasoning for 
medical and legal purposes, as well as to provide 
an explanation that a lay client can understand.  
    This paper presents, for the first time, the design 
of a discourse generator that plans the content and 
organization of genetic counseling patient letters 
containing arguments; and an experiment that we 
performed to evaluate the arguments.  The dis-
course generation process involves three modules: 
a qualitative causal probabilistic domain model, a 
normative argument generator, and a genre-
specific discourse grammar. In (Green, 2005), we 
reported a corpus study that produced a reliable 
biomedical coding scheme. In subsequent work-
shop papers (Green et al., 2004; 2005), we intro-
duced our use of qualitative probabilistic 
constraints and provided a brief description of the 
biomedical domain model. We have also provided 
informal descriptions of argument patterns in the 
corpus (Green, to appear; 2006). However, we 
have not previously published the design of the 
discourse generator, including the discourse 
grammar and argument generator, and their rela-
tionship to the domain model. 
    The theoretical significance of this work is 
three-fold. First, it is empirically based, i.e., based 
on analysis of arguments in a corpus of genetic 
counseling patient letters, since the goal is to pro-
duce the same kinds of normative arguments as are 
used in expert-lay communication. Second, the 
normative argument generator creates an inten-
tional-level representation of the arguments in the 
text, which provides a foundation for an intelligent 
system’s ability to engage in follow-up discussion 
about the arguments that have been presented. Fi-
nally, due to the separation of domain, argument, 
and genre-specific concerns in the design, and due 
to the methodology used to acquire a domain 
model, it should be possible to apply this approach 
to lay-oriented argument generation in other do-
mains. The practical significance of this work is 
that it is major step in the design of a deployable 
system to generate the first draft of genetic coun-
seling patient letters. As genetics plays an increas-
ingly important role in medicine, there is a need for 
tools to aid in dissemination of patient-tailored in-
formation. 
114
 
    In the next section, we give an overview of a 
prototype generation system, whose main compo-
nents are described in more detail in sections 3-5; 
an example of the generation process is given in 
section 6; an experiment to evaluate the generated 
arguments is presented in section 7; and related 
work is summarized in section 8. 
2 System Overview  
We are developing a prototype system for genetic 
counselors that will synthesize the first draft of a 
patient letter. The deployed system will consist of 
a graphical user interface for the genetic counselor, 
a domain model/reasoner, an argument generator, a 
discourse grammar, and a linguistic realizer. Proto-
types of all components except the linguistic real-
izer have been implemented. Although this paper 
focuses on discourse generation and its relationship 
to the domain model, as background we now de-
scribe the flow of information through the system. 
The domain model (section 3) is initialized with 
generic information on clinical genetics. Through a 
user interface providing menus and other non-free-
text input devices, the counselor will provide stan-
dard clinical information such as a patient’s symp-
toms and information about his family tree; test 
results; preliminary diagnosis (before testing); and 
final diagnosis (after test results are known). The 
system uses this information to transform its ge-
neric domain model into a specialized domain 
model of the patient and his family. 
In this genre, a patient letter must provide not 
only the above information, but arguments for the 
diagnosis and other inferences made by the medi-
cal experts. The discourse generation process 
works as follows. A discourse grammar (section 4) 
encodes the high-level topic structure of letters in 
this genre. The discourse grammar rules generate a 
derivation instantiated from the domain model with 
information specific to a patient’s case. For each of 
the writer’s claims about the case for which a nor-
mative  argument must be provided according to 
standard practice, the discourse grammar invokes 
the argument generator.  
The argument generator (section 5) uses non-
domain-specific argument strategies that are in-
stantiated with information from the domain 
model. The argument generator returns a structured 
representation of an argument in which the com-
municative function of information, e.g., as data or 
warrant, is identified. As illustrated in section 6, in 
future interactive systems  knowledge of commu-
nicative function could be used to support follow-
up discussion. In the current prototype, this knowl-
edge is used to determine presentation order, e.g., 
that data supporting a claim is to be presented be-
fore that claim. One of the goals of the experiment 
described in section 7 was to evaluate this order-
ing. In the final system, the output of discourse 
generation will be transformed by a linguistic real-
izer into the first draft of a letter. 
3     Domain Model 
In a previous study of the corpus (Green, 2005), 
we identified a small set of categories (e.g. geno-
type, test result, symptom) with good inter-rater 
reliability that can be used to describe the biomedi-
cal content of a genetic counseling letter as a 
causal probabilistic network (Korb and Nicholson, 
2004). A prototype domain model has been manu-
ally constructed covering representative genetic 
disorders using only these categories of variables. 
By restricting a domain model to these categories, 
the result should reflect the simplified conceptual 
model of genetics used by genetic counselors in 
communication with their lay clients; this facili-
tates generation since the generator will not have to 
distinguish what information in the domain model 
is appropriate to communicate to a lay audience. 
Another benefit of restricting a domain model in 
this way is that it reduces the knowledge acquisi-
tion effort of choosing variables and determining 
network topology; any genetic disorder in the 
scope of the coding scheme (over 4500 single-gene 
autosomal disorders) would be modeled in terms of 
a small number of variable types and a standard 
topology. Thus, it should be straightforward to 
semi-automatically construct a domain model cov-
ering many different genetic disorders. 
Figure 1 shows part of a domain model after it 
has been updated with information about a particu-
lar patient’s case. The nodes labeled GJB2 
(mother), GJB2 (father), GJB2 (child) are geno-
type variables, representing the mother’s, father’s, 
and child’s GBJ2 genotype, respectively. (A geno-
type is a pair of alleles of a gene; one allele is in-
herited from each parent. An individual who has 
two mutated alleles of the GJB2 gene usually ex-
periences hearing loss.) The nodes labeled hearing 
loss (child) and non-syndromic (child) are vari-
115
 
ables representing the child’s symptoms. The node 
labeled test result (child) is a variable representing 
the results of testing the child’s GJB2 genotype.  
The most likely states of the variables are 
shown beside the nodes in Figure 1; T
1
 and T
2
 rep-
resent the time at which the (experts’) belief is 
held, before or after the child’s genetic test results 
are known, respectively. The information recorded 
in the network about this particular case is that the 
child was observed to have hearing loss and no 
features of a genetic syndrome; the preliminary 
diagnosis, i.e. before testing, was that the cause of 
hearing loss is having two mutated alleles of GJB2; 
the test results were negative, however; thus, the 
current diagnosis is some other (unspecified) auto-
somal recessively inherited genetic condition, rep-
resented by the genotype variable labeled other 
genotype (child). In addition, the parents are hy-
pothesized to be carriers (i.e. to each have one mu-
tated allele) of that genotype, represented by the 
variables labeled other genotype (mother), other 
genotype (father).  
     Although a causal probabilistic network used to 
perform diagnosis or risk calculation would require 
specification of numeric probabilities, the role of 
the network in our system is to qualitatively model 
the reasoning that the medical experts have per-
formed outside of the system. Also, we found that 
in the corpus numeric probabilities were provided 
only when citing epidemiological statistics or risks 
calculated according to Mendelian inheritance the-
ory (which does not require Bayesian probability 
computation). Thus, instead of using numeric 
probabilities for domain reasoning, the domain 
model uses qualitative constraints based upon for-
mal relations of qualitative influence, product syn-
ergy, and additive synergy (Druzdzel and Henrion, 
1993).  
     In addition to being adequate for natural lan-
guage generation, this approach greatly reduces  
knowledge acquisition effort; it should be straight-
forward to semi-automatically acquire the qualita-
tive constraints of a full-scale domain model due to 
regularities in this domain and the use of a re-
stricted set of variable types as described above. 
For example, qualitative constraints between geno-
types of parents and child would be determined by 
whether a genotype follows an autosomal domi-
nant or recessive inheritance pattern. 
     We now describe some of the qualitative do-
main constraints. An influence relation holds be-
tween a node in a causal graph and its direct 
descendant. A has a positive qualitative influence 
on B, written S
+
(state(A,V
A
), state(B,V
B
)), if the 
state of A reaching a threshold value V
A
 makes it 
more likely that the state of B reaches value V
B
. 
For example, if having two mutated alleles of a 
genotype A normally results in the appearance of a 
symptom B, this could be described as 
S
+
(state(A,2), state(B,yes)). Each arc in Figure 1 
implicitly represents an S
+
 relation. 
Product and additive synergy describe converg-
ing connections, i.e., the relation between a set of 
variables {A, B} and their direct descendant C in a 
graph. A and B have negative product synergy 
with respect to state V
C
 of C, written                    
X
-
({state(A,V
A
), state(B,V
B
)}, state(C,V
C
)), if 
either the state of A reaching a threshold V
A 
or the 
state of B reaching a threshold V
B 
makes it more 
likely that the state of C reaches V
C
. This type of 
relationship characterizes mutually exclusive alter-
native diagnoses that could account for the same 
symptom; it also characterizes autosomal dominant 
inheritance, an inheritance pattern where inheriting 
one mutated allele of a genotype (from either par-
ent) is usually sufficient to cause health problems. 
In Figure 1, the possible alternative causes of the 
symptoms are indicated by the X
-
 annotations. 
On the other hand, autosomal recessive inheri-
tance, an inheritance pattern where inheriting two 
mutated alleles (one from each parent) is usually 
necessary to cause health problems, is character-
ized by zero product synergy (X
0
); A and B have 
zero product synergy with respect to state V
C
 of C,  
X
0
({state(A,V
A
), state(B,V
B
)}, state(C,V
C
)), if the 
state of A reaching a threshold V
A 
and the state of 
B reaching a threshold V
B 
makes it more likely that 
the state of C reaches V
C
. For example, if the 
mother’s, father’s, and child’s genotype are repre-
sented by variables A, B, and C, respectively, then 
X
0
({state(A,1), state(B,1)}, state(C,2)) can repre-
sent the constraint that if the child’s genotype C 
has two mutated alleles, then one mutated allele 
must have come from each parent. In Figure 1, the 
autosomal recessive inheritance pattern of GJB2 
and the other hypothesized genetic disorder are 
indicated by the X
0
 annotations. 
Other qualitative constraints used in the domain 
model are based on negative qualitative influence 
(S
-
), positive product synergy (X
+
), and negative 
additive synergy (Y
-
). In addition, the domain 
model stores epidemiological statistics as probabil-
116
 
ity statements composed of variables used in the 
network, e.g., the frequency of hearing loss due to 
GJB2. This type of information can be used as 
backing in an argument (see section 5) but does not 
play a role in domain reasoning. 
4 Discourse Grammar 
A discourse grammar was written based upon our  
analysis of the corpus and a description of standard 
practice in genetic counseling (Baker et al., 2002). 
The current grammar is intended to cover letters on 
single-factor autosomal genetic disorders. Thanks 
to the regularities in this domain and in this genre, 
the grammar consists of a small number of rules. 
The starting rule of the grammar represents the 
main sections of a letter in their standard order: 
opening, referral, preliminary diagnosis, testing, 
final diagnosis, origin of genetic condition, inheri-
tance implications, prognosis/treatment, and clos-
ing. One or more grammar rules describe each of 
these sections.  
     Grammar rules may request the domain rea-
soner for case-specific information to be included 
in the letter. In addition, when the grammar pro-
vides a choice of rules, rule selection is based upon 
case-specific information provided by the domain 
reasoner. For example, one rule for reporting the 
final diagnosis handles cases in which the patient’s 
test results confirm the preliminary diagnosis, and 
another rule those cases where the preliminary di-
agnosis has been disconfirmed by test results; the 
domain reasoner returns the information needed to 
choose between those two rules. 
     The process described so far creates an initial 
outline of the information to be presented (in non-
linguistic form), including various claims requiring 
an argument. Each of those claims is passed to the 
argument generator described in the next section. 
For example, the letter shown in Figure 2 contains 
seven claims labeled C
1
 to C
7
; argument generation 
adds information labeled D
1
 to D
7
, W
1
 to W
7
, and 
B
1
 to B
4
. The information returned by the argument 
generator is added to the outline, completing the 
structure that will be transformed by the linguistic 
realizer into text.      
5 Argument Generation 
Given a claim, the argument generator uses argu-
ment strategies to construct a normative argument 
for the claim from information provided by the 
domain reasoner. The strategies are non-domain-
specific in the sense that they refer to formal prop-
erties of the qualitative causal probabilistic domain 
model rather than to genetics.  
    According to Toulmin’s model of normative 
argument structure (1998), an argument for a claim 
can be analyzed in terms of various functional 
components: the data, warrant, and backing. The 
data are the facts used to defend a claim. The war-
rant is a principle that licenses the claim given the 
data. An optional backing may be used to justify 
the warrant, e.g., by giving the facts upon which 
the warrant is based. To derive the argument 
strategies used in the system, we analyzed the ar-
guments in the corpus in terms of Toulmin’s 
model; the resulting strategies describe mappings 
from formal properties of the domain model to the 
data and warrant supporting a claim and to the 
backing of a warrant. Several strategies are para-
phrased below for illustration.  
    Strategy 1. Argument for belief in causal claim, 
based on effects: An argument for the claim that it 
is believed to some extent at time T
i 
that 
state(A,V
A
) holds and that state(A,V
A
) is responsi-
ble for the states of variables B
1
..B
i
, i.e., 
state(B
1
,V
B1
) .. state(B
i
,V
Bi
), consists of the (pre-
supposed) data that state(B
1
,V
B1
) .. state(B
i
,V
Bi
) 
hold, and optionally other data that state(B
j
,V
Bj
) .. 
state(B
k
,V
Bk
) hold, where the warrant is a positive 
influence relation S
+
(state(A,V
A
), state(B
p
,V
Bp
)) 
for each B
p
 in   { B
1
 .. B
i
 , B
j
 .. B
k
}. 
     Strategy 2. Argument for decrease in belief to 
unlikely that state of causal variable is at or over 
threshold value, based on absence of predicted 
effect: An argument for the claim that there has 
been a decrease in belief, from time T
1 
to T
2
, to the 
belief at T
2
 that it is unlikely that state(A,V
A
) 
holds, consists of the (newly acquired) data that it 
is unlikely that state(C,V
Ci
) holds for all V
Ci
$V
C
, 
where the warrant is a positive influence relation 
S
+
(state(A,V
A
), state(C,V
C
)). 
    Strategy 3. Argument for increase in belief in 
causal claim, based on decrease in belief in alter-
native cause: An argument for the claim that there 
has been a increase in belief, from time T
1 
to T
2
, to 
the belief at T
2
 that it is believed to some extent 
that state(A,V
A
) holds and that state(A,V
A
) is re-
sponsible for the states of variables state(B
1
,V
B1
) .. 
state(B
i
,V
Bi
), consists of the (presupposed) data 
that state(B
1
,V
B1
) .. state(B
i
,V
Bi
) hold, and the 
117
 
(newly acquired) data that it is unlikely that 
state(Alt,V
Alt
) holds for all V
Alt
$V
threshold
, where the 
warrant is a negative product synergy relation     
X
-
({state(A,V
A
),state(Alt,V
threshold
)},state(B,V
B
)) 
for each B in {B
1
 .. B
i
}. 
    Strategy 4. Argument for belief in joint respon-
sibility, based on effect. An argument for the claim 
that it is believed to some extent at time T
i 
that 
state(A,V
A
) and state(B,V
B
) hold and that 
state(A,V
A
) and state(B,V
B
) are jointly responsible 
for state(C,V
C
), consists of the (presupposed) data 
that state(C,V
C
) holds, where the warrant is a zero 
product synergy relation X
0
({state(A,V
A
), 
state(B,V
B
)}, state(C,V
C
)). 
6 Example 
This section gives an example of discourse genera-
tion for the case in section 3. An outline created by 
application of the discourse grammar to the do-
main model in Figure 1 would contain, in addition 
to basic information about the case not requiring an 
argument, several claims requiring further support 
to be provided by the argument generator.  
    First, the claim that it was believed, before test-
ing, that the child’s hearing loss could be due to 
having two mutated alleles of GJB2 would be sup-
ported by an argument constructed using Strategy 
1. The data of the argument is the presupposition 
that the child has hearing loss and the additional 
finding that she has no syndromic features. The 
warrant is the positive influence relations (S
+
) link-
ing the variable representing the child’s GJB2 
genotype to each of the two variables representing 
the child’s symptoms. Note that if a reader ques-
tioned this argument, an interactive system could 
provide information on the source of the data or 
epidemiological statistics backing the warrant. 
      Second, the claim that it is currently believed, 
after testing, that it is unlikely that the child’s 
GJB2 genotype has two mutated alleles would be 
supported by an argument constructed using Strat-
egy 2. The data of the argument is that the child’s 
GJB2 test results were negative. The warrant is the 
positive influence relation (S
+
) from the child’s 
GJB2 genotype to the child’s GJB2 test results, 
which predicts that if the child had this mutation, 
then the test results would have been positive. If a 
reader questioned this argument, an interactive sys-
tem could provide information on the source of the 
data or back the warrant by providing information 
about the rate of false negatives.  
     Third, the claim that it is currently believed, 
after testing, that it is possible that the child has 
some other genetic condition that is responsible for 
her hearing loss would be supported by an argu-
ment constructed using Strategy 3. The data of the 
argument is that she has hearing loss and the cur-
rent belief that GJB2 is not likely responsible. The 
warrant is the negative product synergy relation 
(X
-
) between the child’s GJB2 genotype and an-
other genotype to hearing loss. If a reader ques-
tioned this argument, an interactive system could 
provide information on the proportion of cases of 
hearing loss that are due to other genetic conditions 
as backing for the warrant. 
    Fourth, the claim that it is currently believed, 
after testing, that it is possible that the parents are 
carriers (i.e., each has one mutated allele) of the 
unspecified genotype claimed to be responsible for 
the child’s hearing loss would be supported by an 
argument constructed using Strategy 4. The data of 
the argument is the presupposition that the child 
has two mutated alleles of the other genotype. The 
warrant is the zero product synergy relation (X
0
) 
between the two parents’ genotype for this alterna-
tive to GJB2 and the child’s genotype for this same 
alternative. If a reader questioned this argument, an 
interactive system could provide an explanation of 
the warrant, which is based on the theory of Men-
delian inheritance; or it could provide the argument 
for the data, i.e., the belief that the child has two 
mutated alleles of the other genotype. 
    Finally, the claim that it is currently believed, 
after testing, that assuming they are both carriers 
there is a 25% probability that each future child 
that the two parents have together will inherit two 
mutated alleles of the other genotype would be 
supported by an argument constructed by a strat-
egy not shown in section 5. The data is the as-
sumption that the parents are both carriers, and the 
warrant is the same zero product synergy relation 
(X
0
) used in the argument for the fourth claim. If a 
reader questioned this argument, an interactive sys-
tem could provide an explanation of how the prob-
abilities are determined by zero product synergy.  
7 Experiment 
Argument generation was evaluated in the follow-
ing experiment. Five biology graduate students, 
118
 
screened beforehand for writing ability in biology, 
were shown two patient letters. The letters were 
created by the experimenter by paraphrasing the 
output of discourse generation that would be input 
to the realizer. The paraphrases are similar in syn-
tax and lexical style to letters in  the corpus, but the 
genetic disorders covered in the experiment’s let-
ters differ from those covered in the corpus. One 
letter concerns a child confirmed to have cystic 
fibrosis (CF); the other a child whose test results 
for Waardenburg syndrome (WS) were negative. 
The text of letter CF is given in Figure 2. The first 
column contains annotations describing the com-
municative function of the information: C for 
claim, D for data, W for warrant, and B for back-
ing. Each label is subscripted with an integer refer-
ring to the argument. (The row labeled C
2
/D
3
 
functions as both the claim of argument 2 and the 
data of argument 3.) Annotations were not shown 
to the experiment’s participants. Communicative 
function was used to determine presentation order 
within each argument. Letters CF and WS had 23 
and 25 segments, respectively, where a segment is 
defined as a unit fulfilling one of the above func-
tions, or a non-argument-related function. 
    The goal of the experiment was to conduct a 
preliminary evaluation of the acceptability of the 
arguments in terms of content, explicitness, and 
presentation order within arguments. The partici-
pants were asked to revise each letter as needed to 
make it more appropriate for its intended recipi-
ents, the biological parents of a patient. Partici-
pants were told they could reword, reorder, and 
make deletions and additions to a letter. The results 
are summarized in Table 1, which includes the av-
erage number of segments to/from which informa-
tion was added (New) or deleted (Delete), and 
reordered (Reorder). (Rewordings are not tabulated 
since it was not our goal to evaluate wording.) 
New and Delete are measures of acceptability of 
argument content and explicitness. Reorder is a 
measure of acceptability of ordering. On average, 
the number of New, Delete, and Reorder revisions 
were low: less than two per letter, with most revi-
sions in the category of Reorder. This is encourag-
ing since the system to be built for genetic 
counselors should provide acceptable arguments 
requiring a minimum of revision. 
     To provide more details about the results, first, 
the only segments to which participants added in-
formation are warrants. The deletions of data con-
sist of information presumably already known to 
the recipients, e.g. D
6 
in letter CF; other deletions 
are of part or all of a warrant or all of a backing. 
The only deletions of claims consist of information 
duplicated in another part of the letter; there were  
no cases where a claim was deleted even though it  
could be inferred from data and warrant. The reor-
derings were across-argument, which violates con-
ventional topic structure in the genre, or within-
argument. In the latter, half repositioned a claim 
from final position in an argument to a position 
before the warrant or backing; the other half repo-
sitioned the warrant or backing before the data.  
8 Related Work 
Due to space limitations, this section focuses on 
research on generation of normative arguments (as 
opposed to behavior-change and evaluative argu-
ments), and arguments designed for text rather than 
dialogue. Zukerman et al. have presented several 
papers on argument generation from Bayesian 
network domain models (e.g., 2000). The type of 
domain model used in our work differs in two re-
spects. First, it is based on empirical research since 
it is intended to represent the simplified conceptual 
model presented to a lay audience in this genre. 
Second, it uses qualitative probabilistic constraints. 
One difference in argument generation is that our 
system’s argument strategies are based on analysis 
of the corpus. Also, our system creates an inten-
tional-level representation of an argument.     
     Teufel and Moens (2002) present a coding 
scheme for scientific argumentation in research 
articles that is designed for automatic summariza-
tion of human-authored text. Thus, it would not be 
sufficient for generation from a non-linguistic 
knowledge base. Also, it does not make the finer-
grained distinctions of the Toulmin model. 
     Branting et al. (1999) present the architecture of 
a legal document drafting system. In it, a discourse 
grammar applies genre-specific knowledge, while 
a legal reasoning module creates the illocutionary 
structure of legal arguments. Branting et al. argue 
for maintaining a distinct intentional-level repre-
sentation of arguments to support interactive fol-
low-up discussion. We agree, but our design 
further distinguishes domain reasoning from argu-
ment generation.  
    As for work on ordering and explicitness, Reed 
and Long (1997) propose ordering heuristics for 
119
 
arguments of classical deductive logic. Fiedler and 
Horacek (2001) present a model for deciding what 
can be omitted from explanations of mathematical 
proofs. Carenini and Moore (2000) present an ex-
periment to determine how much evidence is opti-
mal in an evaluative argument.  
9 Conclusions 
This paper presents the design of a discourse gen-
erator that plans the content and organization of 
genetic counseling letters containing arguments. A 
preliminary evaluation of the arguments was prom-
ising. The most important contribution of this work 
is the design of a non-domain-specific normative 
argument generator that creates an intentional-level 
representation of an argument. From the corpus, 
we formulated argument strategies that map formal 
properties of qualitative causal probabilistic mod-
els to components of Toulmin’s model. Due to the 
separation of domain, argument and genre-specific 
concerns and the methodology used for acquiring 
the domain model, this approach should be appli-
cable to lay-oriented normative argument genera-
tion in other domains.  
Acknowledgments  
 
This material is based upon work supported by the 
National Science Foundation under CAREER 
Award No. 0132821. 

References 
Baker DL, Eash T, Schuette JL, Uhlmann WR. 2002. 
Guidelines for writing letters to patients. J Genetic 
Counseling, 11(5):399-418. 
Branting LK, Callaway CB, Mott BW, Lester JC. 1999. 
Integrating Discourse and Domain Knowledge for 
Document Drafting. Proc ICAIL-99, 214-220. 
Carenini G, Moore J.  2000. An empirical study of the 
influence of argument conciseness on argument ef-
fectiveness. Proc Ann Meeting of ACL, 150-7. 
Druzdzel MJ, Henrion M. 1993. Efficient reasoning in 
qualitative probabilistic networks. Proc 11th Nat Conf 
on AI, 548-553. 
Fiedler A, Horacek H. 2001. Argumentation in Explana-
tions to Logical Problems.  Computational Models of 
Natural Language Arguments. Proc ICCS 2001. 
Springer LNCS 2073, 969-978. 
Green N. 2005. A Bayesian Network Coding Scheme 
for Annotating Biomedical Information Presented to 
Genetic Counseling Clients. J Biomed Inf, 38: 130-
144. 
Green N. 2006. Representing Normative Arguments in 
Genetic Counseling. AAAI SSS: Argumentation for 
Consumers of Healthcare. 
Green N. To appear. Argumentation in a Causal Prob-
abilistic Humanistic Domain. Int J Intell Sys. 
Green N, Britt T, Jirak K. 2004. Communication of Un-
certainty in Genetic Counseling Patient Education 
Systems. AAAI FSS: Dialog Sys for Health Commun. 
Green N, Britt T, Jirak K, Waizenegger D, Xin X. 2005. 
User Modeling for Tailored Genomic E-health In-
formation. User Modeling 2005 Workshop: Person-
alisation for eHealth, 
Korb K, Nicholson AE. 2004. Bayesian artificial intel-
ligence.  Chapman Hall/CRC, Boca Raton, Florida.  
Reed C, Long D. 1997. Content ordering in the genera-
tion of persuasive discourse.  IJCAI-97, 1022-27. 
Teufel S, Moens M. 2002. Summarizing Scientific Arti-
cles: Experiments with Relevance and Rhetorical 
Status. CL, 28(4):409-445. 
Toulmin SE. 1998. The uses of argument. 9th ed. Cam-
bridge Univ. Press, Cambridge, England. 
Zukerman I, McConacy R,  Korb K. 2000. Using argu-
mentation strategies in automated argument genera-
tion. Proc INLG-2000.
