Download as ps, pdf, or txt
Download as ps, pdf, or txt
You are on page 1of 4

The Conceptual and Linguistic Facets

of Persuasive Arguments
Daniel Marcu 1
Abstract. This paper provides a body of knowledge that dose of the attacking material strong enough to stimulate his
characterizes persuasive arguments, which is thoroughly defenses but not strong enough to overwhelm him" [13, p.37].
grounded in empirical data derived from communication stud- In examining the problems that an NLG system will have to
ies, psychology, and social studies of persuasion. The paper solve in generating an inoculation message, I assume that the
also discusses the limitations of current theories of argumen- system has access to a knowledge base that contains general
tation and systems in accommodating both the conceptual information about smoking, information about the teenagers
and linguistic facets of persuasive arguments. and their beliefs, and information concerning the structure
of sound arguments. To determine the content to be gen-
1 Introduction erated, a traditional NLG system would evaluate rst the
If natural language generation systems that argue convinc- knowledge that a generic teenager has about smoking. The
ingly and adapt to di erent users are to become more exible system would infer that teenagers already know that smoking
and pro cient, they will eventually have to make use of more is an unhealthy habit, so it would select no information to con-
knowledge of how persuasive arguments can be represented, vey about this topic. It would be likely to select information
generated, and understood. In this paper, I provide a body about the attack that the teenager is going to be subjected to
of such knowledge that is well grounded in empirical data and, using its knowledge of structure of arguments, it would
derived from communication studies, psychology, and social also select some data to increase the credibility of its claim. If
studies of persuasion. the system has outstanding inferential capabilities, it would
To support my claims and provide a better understanding also select some information to contrast the present state of
of the heuristics that could be applied for constructing per- a airs (children want to stay non-smokers) with the one that
suasive arguments, I consider a concrete task: the generation the majority of teenagers will experience (most teenagers will
of a text that is designed to convince teenagers not to smoke. start smoking).
I walk through the steps that a natural language generation The planning modules would determine the order into which
(NLG) system would take in accomplishing this task, and, this information is going to be conveyed: an order that is con-
at each stage of the generation process, i.e., argument selec- sistent with a sound argument structure is one in which the
tion, argument organization, and argument realization, I ap- contrast between the current attitude that is incompatible
ply techniques that are likely to increase the persuasiveness of with smoking and the future attack from the peer group, to
the text. Hence, I examine how the text is modi ed incremen- start smoking, is presented rst. The contrast can be made
tally from one that is neutral to one that is persuasive. In the more salient if it is followed by the statistical data, so let us as-
light of the requirements derived from my study of persuasive sume that the planning modules choose to generate this data
arguments, I evaluate the capability of current argumentation afterwards. The realization process would then choose appro-
theories and systems to accommodate the representation and priate syntactic structures and lexical items to represent the
generation of persuasive arguments. information. It is possible that the output for this generation
task would look similar to the one given below.2
2 Generating persuasive arguments (1) No matter how much one wants to stay a non-smoker, the
2.1 General remarks truth is that the pressure to smoke in junior high is greater than
it will be any other time of one's life: only 25% of the young
Consider an NLG system that is posed the task of generat- adults will not pick up a cigarette and let curiosity take over.
About 30% will become experimental smokers. Of those who
ing a message to persuade teenagers to not start smoking. will start smoking, about 90% will end up with a pack and a
Behavioral studies [16] show that children are born and live lighter for the rest of their lives. We know that 3,000 teens start
through the rst decade of their lives with a natural aversion smoking each day, although it is a fact that 90% of them once
for smoking. However, during the transition from primary to thought that smoking was something that they'd never do.
middle grades, this attitude deteriorates [16] and produces in Text (1) presents a sound argument, it is linguistically cor-
some adolescents a growing chance of experimental and reg- rect, and it re ects accurately the knowledge of the system,
ular smoking. One way to prevent this is by presenting argu- but is not likely to convince too many teenagers to refrain
ments that inoculate, i.e., \by exposing the person to a weak from smoking. I discuss now the features that text (1) needs
in order to be convincing, by modifying it incrementally so
1 Department of Computer Science, University of Toronto, Toronto,
Ontario, Canada, M5S 3G4. 2 All the examples used in this paper were generated manually.
c 1996 D. Marcu
Proceedings of the ECAI 96 Workshop Gaps and Bridges:
New Directions in Planning and Natural Language Generation.
Edited by K. Jokinen, M. Maybury and M. Zock.
that it will re ect heuristics prescribed by current studies teenagers that smoking is a detrimental habit, that it costs
in persuasive communication. The nal text (4) is a slightly money, and a ects peoples' breath since they are likely to be
modi ed version of a text used in an adult-led video, Up in persuaded both by logical and emotional arguments [16]. The
Smoke, which is discussed by Pfau [16, p. 106]. Due to the persuasive e ect can be increased if the involvement of the
nature of the example that I use, some of the heuristics that source is highlighted: we know that smoking : : :. Because it is
I apply may not be appropriate outside the medical domain; a complete argument, the new paragraph (2) can be realized
however, I am not aware of empirical data that supports or either as a pre x or a sux of the initial text (1).
refutes this assumption. (2) I don't know why people waste their breath on these
cigarettes. We know that smoking is a detrimental habit and
2.2 Argument selection in persuasive that it costs money.
argumentation 2.3 Argument organization in persuasive
Receiver factors. Contrary to NLG wisdom, a system ca- argumentation
pable of generating persuasive text will also have to generate The NLG community has long acknowledged that the same se-
information that is known to the audience [12, 9], i.e., the sys- mantic information can be conveyed through a variety of text,
tem will have to reinforce the recipients' (hearers') beliefs that paragraph, and sentence structures. Communication studies
encourage the desired behavior and to argue against the be- emphasize further the role that structure plays, because al-
liefs that inhibit it. This requires a quite sophisticated model though the understanding of a message decreases smoothly
of the user with regard to likes and dislikes. as the same semantic information is presented in a less and
Recent research in the eld of behavioral decision mak- less structured way, the persuasive e ects vanish rapidly [2].
ing [9] shows that persuasion is not a constant function of In persuasive argumentation, one should capture the at-
the beliefs and attitudes, but that it rather varies depending tention of the audience by making the message unusual, by
on the stage of change that the recipient is in. In the stages- presenting the arguments in a novel way [17]. One way to cap-
of-change model for example [6, 9], there are ve stages that ture the attention of the receiver is by starting with a strong
range from precontemplative, the stage in which the recipient remark that establishes a connection between the source and
does not recognize the problem or the need for a change, to the teenagers [12]. Given the information in texts (1) and (2),
maintenance, the stage in which the recipient has maintained one can infer that smoking is a stupid habit. It may be a
a new behavior for more than six months. A recipient may be good idea to capture the attention of the reader by explicitly
stalled at one stage and need a \push" to the next. It is ob- stating this. Such an abrupt start not only increases the de-
vious that the arguments of a message should di er radically gree of involvement of the source and the familiarity between
according to the stage that the recipient is in: the informa- the source and the receiver, but it also constitutes a warning
tion provided to a person who is not aware that smoking is that precedes the persuasive attack that is to follow [12]. The
unhealthy should be di erent from the information provided credibility of the source can be also increased if the statistical
to a person who has been trying to quit smoking for more data is assigned a high-credibility source: instead of we know
than three months. that, I use research tells us.
The receiver's interest a ects the argument selection as Depending on the degree of awareness of the audience,
well: if it is high, the NLG system should select only sound di erent orderings of the arguments to be presented should
arguments; if it is low, the system should select as many argu- be chosen. For an unaware audience, the persuasive message
ments as possible, even if they are unsound, because this will should be composed such that supporting arguments are pre-
have a better e ect on the receiver [12, 2]. A key factor is the sented rst, and refuting arguments second. For an audience
system's ability to reason about the beliefs of the recipient that is aware, persuasion increases if the message refutes rst
and to determine those that support the goal of the message the goal to be achieved and supports it afterwards [12]. In
and those that contradict it. Since it is easier to introduce, our example, when inoculation is performed, the teenagers
reinforce, or build on existing beliefs in a campaign than it are unaware of the pressure that they are going to be ex-
is to try to change existing and entrenched attitudes [19], a posed to: therefore, it is better to support rst the aversion
system should select and construct its arguments around the for smoking that they have, and present them only afterwards
beliefs that support the goal. the arguments that refute this aversion.
Source and message factors. The credibility, attractive- One of the key factors in persuasive argumentation is the
ness, and power of the source of the message play another continuous reinforcement of the beliefs that the receivers hold
important role in argument selection: the more an argument and the emphasis that one should put on cumulative risks
emphasizes them, the more persuasive it is [17]. that people tend to underestimate [9]. One can apply this
Research in communication studies [17] distinguishes be- technique to increase the persuasiveness of text that results
tween the persuasive e ects that logical and emotional ar- from combining paragraphs (1) and (2). For example, both
guments have on a particular audience. For example, logical smoking smells and is bad or No matter how one feels about
appeals are e ective when the audience is sophisticated, when smoking now and no matter how bad one thinks it is, or even
the competence of the source is highlighted, and when the goal if one promises one will never start, 75% of the teenagers will
of the appeal is stimulating cognitive change, while emotional try it at least once (text (3) below) add no new information
appeals are better to be used when the audience is indi er- to text that results from appending paragraphs (1) and (2),
ent and when the goal of the appeal is producing e ective but it reinforces over and over the beliefs of the teenagers.
change [17]. An increase in persuasion can be obtained by explicitly
Several heuristics can be applied by an NLG system to in- including in the text the conclusions that are implicit [12].
crease the persuasive e ect of text (1). The system can remind Explicit conclusions should go either rst or last, but not in
The Conceptual and Linguistic Facets of Persuasive Arguments 2 D. Marcu
the middle. If the conclusion is shattering, as in the case of \stronger" language and \pathetic" words [2], and speci c
telling a patient that they have cancer, the conclusion should terms instead of abstract terms [18, p. 142] increases the de-
go last because after such news, the patient will not be able gree of persuasiveness. In my ongoing example, these recom-
to follow the rest of the message [19]. mendations will make us prefer cancer sticks over cigarettes;
In my ongoing re nement of the inoculation message, I messy over detrimental, and bucks over money.
make explicit the conclusions of each part: I use the metaphoric Increasing the denotative, spatial, and temporal speci city
comparison between smoking and baseball as an intermediate of a text is another way to make an argument more persua-
conclusion; and a direct inference | but something or some- sive [15]. The denotative speci city can be increased by stat-
one convinced them otherwise | as the nal conclusion of the ing the agent, object, and action explicitly, the spatial imme-
warning. This corresponds to the cue of using a climax order- diacy by using demonstratives that are \close" (this, these,
ing [12] and to avoiding the saying of the best things, or the here not that, those, there), and temporal immediacy by us-
conclusions, in the middle of the text. ing the present tense in reference to present events. The ap-
The persuasive technique of framing seems to belong en- plication of the heuristics that are described above yields the
tirely to a micro-organization level of arguments. Despite con- following text:
troversies that exist with respect to psychological explana-
tions of the fear appeals and the relation between fear and (4) Stupid habit! I don't know why people waste their breath
on these things | literally. We used to call them \cancer sticks".
persuasion [17], empirical data suggests that in order to con- I know that you think that smoking is a messy habit, right? It
vince, arguments have to be framed such that the bad con- stinks, it's bad, and it costs bucks. As they say in baseball \three
sequences are emphasized [14]. Hence, although 25% of the strikes and you're out".
young adults will not start smoking and 75% of the young No matter how much you want to stay a non-smoker, the
truth is that the pressure to smoke in junior high is greater than
adults will start smoking have the same semantics, I prefer it will be any other time of your life. Three out of four young
the latter one. adults | and you are one of them | will pick up a cigarette
and let curiosity take over.
(3) Stupid habit! I don't know why people waste their breath on Think about it! No matter how you feel about smoking now,
these cigarettes. We know that smoking is a detrimental habit. no matter how bad you think it is, or even if you swear you'll
It smells, it's bad, and it costs money. As they say in baseball never start, three of every four of you will try it at least once.
\three strikes and you're out". Almost a third of you will become experimental smokers, and
No matter how much one wants to stay a non-smoker, the many of you will end up with a pack and a lighter full time.
truth is that the pressure to smoke in junior high is greater than Researchers tell us that 3,000 teens start smoking each day, and
it will be any other time of one's life: 75% of the young adults you can count on the fact that most of them, like you, once
will pick up a cigarette and let curiosity take over. thought that smoking was something that they'd never do. But
No matter how one feels about smoking now and no mat- something or someone convinced them otherwise.
ter how bad one thinks it is, or even if one promises one will
never start, 75% of the teenagers will try it at least once. About If one now compares texts (1) through (4) it will become ev-
30% will become experimental smokers. Of those who will start ident that persuasive argumentation assumes appropriate de-
smoking, about 90% will end up with a pack and a lighter for cisions with respect to argument selection, organization, and
the rest of their lives. Research tells us that 3,000 teens start realization.
smoking each day, although it is a fact that 90% of them once
thought that smoking was something that they'd never do. But
something or someone convinced them otherwise. 3 Why are representing and generating
persuasive arguments dicult?
2.4 Lexical realization in persuasive Argument selection. The strategies that are employed by
argumentation current NLG systems for content selection (schemata, plan
A general heuristic for increasing persuasiveness is to use an operators, and NLG front-ends) are unable to account for the
intensely worded message if the recipients agree initially with selection of persuasive arguments. One of the reasons is that
the source, but less wordy if they disagree [12]. The large none of these strategies provides a principled way for select-
amount of redundancy and repetition that I have used so far ing arguments that use evidence that is known to the hearer
supports this recommendation, since teenagers agree initially or arguments that argue against the hearer's beliefs. Con-
with the position of the message. sequently, it is natural to ask ourselves whether research in
People have trouble reasoning with probabilities and there argumentation could provide an adequate solution. Unfortu-
is a great disagreement about how di erent probabilistic val- nately, the answer is negative: conceptual structures that are
ues are linked with qualitative adjectives [9]. But as we have proposed by research on argumentation are not adequate for
already seen (section 2.3), the understanding of a message is distinguishing between arguments that reinforce the beliefs of
an important prerequisite for the success of a persuasive com- a hearer and arguments that try to refute them; or between
munication act [17]. Therefore, it comes as no surprise that re- arguments that are logical, emotional, and unsound.
placing quantitative quali ers with qualitative quali ers that Independent of the mathematical abstractions that they
are prescribed by socio-linguistic research is a heuristic that use, conceptual approaches derive their strength from the
can be used for increasing persuasiveness [9]. Socio-linguistic careful characterization of the structures (trees and graphs)
research [4] prescribes for some audiences, for example, the that pertain to sound arguments [18, 3, 5, 11]: their under-
use of virtually always or high probability instead of 99%, and standing is shown to yield certain conclusions independent
unlikely or doubtful instead of 20%. of the propositions that they relate [3] and to alleviate se-
Lexical choice in general plays an important role in persua- mantic ambiguities [8]. One of the weaknesses of the concep-
sive argumentation. One way to increase persuasiveness is by tual structures that have been proposed for representing ar-
eliminating all the quali ers that express uncertainty, such guments previous to this workshop is that they have been
as perhaps, maybe or possibly. The use of metaphors [12], totally disconnected from user models and as a result, NLG
The Conceptual and Linguistic Facets of Persuasive Arguments 3 D. Marcu
systems could not easily accommodate the di erent levels of and hearer; a lack of argumentative theories that can account
involvement and interest, moral pro les, and cognitive stages for the composition of multiple arguments, repeating evidence
that are to be exploited in a persuasive setting. The papers by and conclusions, etc.; and an insucient understanding of the
Zukerman, Korb, and McConachy and Quignard and Baker in role that lexico-grammatical choices have on making an argu-
these proceedings provide, however, a rst cut at these issues. ment more or less convincing.
Argument organization. As far as I know, research in Acknowledgements. I am grateful to Graeme Hirst for his
argumentation has been concerned only with the structure astute comments and for helping me polish the ideas presented
of single arguments; and likewise, NLG systems that provide in this paper. This reasearch was supported by a grant from
explanations and advice [7, 11] don't explore the planning the Natural Sciences and Engineering Research Council of
mechanisms that would account for the generation of texts Canada.
that consist of multiple arguments, i.e., arguments that both
support and refute a given claim. But as I have discussed
in section 2, in order to generate persuasive arguments, one
REFERENCES
[1] J.C. Anscombre and O. Ducrot, L'argumentation dans la
needs also a theory that would describe how arguments can be langue, Pierre Mardaga, Bruxelles, 1983.
put together in a coherent sequence and explain why certain [2] E.P. Bettinghaus and M.J. Cody, Persuasive Communication,
multi-argument structures are more persuasive than others. Holt, Rinehart and Winston, Inc., 1987.
One would also need theories to explain why for the same [3] L. Birnbaum, `Argument molecules: a functional representa-
argument structure one certain framing is more persuasive tion of argument structures', in AAAI{82, pp. 63{65.
than another. None of the techniques employed in NLG and [4] G.D. Bryant and G.R. Norman, `Expressions of probability:
Words and numbers', The New England Journal of Medicine,
no argument structures can account for the persuasive e ect 302, (1980).
that repeating information or reinforcing beliefs has; nor can [5] R. Cohen, `Analyzing the structure of argumentative dis-
they determine when such heuristics be applied. The same course', Computational Linguistics, 13 (1-2), 11{24, (1987).
shortcomings are apparent with respect to the role of explicit [6] C.C. DiClemente and J.O. Prochaska, `Processes and stages
conclusions in argument structures and the nature of the pro- of self-change: Coping and competence in smoking behavioral
cess that takes the decision whether to include them in the change', in Coping and Substance Abuse, eds., S. Shi man
nal text. and T.A. Willis, 319{343, Academic Press, (1985).
Argument realization. My review of persuasive argumen- [7] M. Elhadad, Using Argumentation to Control Lexical Choice:
tation (section 2) has revealed a number of lexical character- A Functional Uni cation Implementation, Ph.D. disserta-
istics that cannot be incorporated easily in current theories of tion, Columbia University, 1992.
[8] S. Guez, `A computational model for arguments understand-
argumentation. On one hand, conceptual schemata [18, 3, 5] ing', in COLING{92, volume 2, pp. 132{137, Helsinki, (1990).
seem to be too abstract for dealing with the di erences be- [9] D.R. Holtgrave, B.J. Tinsley, and L.S. Kay, `Encouraging risk
tween more or less intensely worded arguments, and for ex- reduction. A decision-making approach to message design',
plaining the impact that the use of strong language, speci c in Designing Health Messages. Approaches from Communi-
terms, and denotative, spatial, and temporal speci c terms cation Theory and Public Health Practice, eds., E. Maibach
has on an argument. On the other hand, linguistic accounts and R.L. Parrott, 24{40, Sage Publications, (1995).
were interested more in capturing the relation between lin- [10] E.H. Hovy, Generating Natural Language Under Pragmatic
guistic cues and certain argumentative structures [5, 8] than Constraints, Lawrence Erlbaum Associates, Publishers, 1988.
the actual linguistic realization of the arguments. [11] M.T. Maybury, `Communicative acts for generating natural
An exception to this, is Anscombre and Ducrot's approach [1], language arguments', in AAAI{93, pp. 357{364, (1993).
[12] W.J. McGuire, `The nature of attitudes and attitude change',
which depicts a semantic theory in which some words carry in The Handbook of Social Psychology, eds., G. Lindzey and
besides the usual semantic meaning, an argumentative conno- E. Aronson, volume 3, 136{314, Addison-Wesley, (1968).
tation that orients the argumentative interpretation and the [13] W.J. McGuire, `A vaccine for brainwash', Psychology Today,
continuation of a discourse independently of the informative 3, 36{39, (1970).
content it may carry. Elhadad [7] relies heavily on Ansombre [14] B.E. Meyerowitz and S. Chaiken, `The e ect of message fram-
and Ducrot's theory in building an NLG system that inte- ing on breast self-examination attitudes, intentions, and be-
grates both the conceptual and linguistic facets of argumen- haviors', Journal of Personality and Social Psychology, 52,
tation, but his approach does not seem to be able to account 500{510, (1987).
for all the linguistic and conceptual features that were shown [15] R.L. Parrott, `Motivation to attend health messages. Presen-
to characterize persuasive texts. tation of content and linguistic considerations', in Designing
Health Messages. Approaches from Communication Theory
4 Conclusion and Public Health Practice, eds., E. Maibach and R.L. Par-
In contrast with previous approaches that characterize the rott, 7{23, Sage Publications, (1995).
in uence of pragmatic factors on NLG and the generation [16] M. Pfau, `Designing messages for behavioral inoculation', in
of texts with a ect [10], this paper provides a collection of Designing Health Messages. Approaches from Communica-
features that characterize persuasive arguments that is thor- tion Theory and Public Health Practice, eds., E. Maibach and
oughly grounded in empirical data derived from communica- R.L. Parrott, 99{113, Sage Publications, (1995).
tion studies, psychology, and social studies of persuasion. The [17] J.B. Sti , Persuasive Communication, Guilford Press, 1994.
[18] S. Toulmin, R. Rieke, and A. Janik, An Introduction to Rea-
paper also analyzes the inability of current argumentation soning, Macmillan Publishing Co., Inc., 1979.
theories and systems to accommodate the representation and [19] K. Witte, `The manipulative nature of health communication
generation of persuasive arguments. This inability is shown research. Ethical issues and guidelines', American Behavioral
to stem mainly from a formalization of argumentative struc- Scientist, 38 (2), 285{293, (1994).
tures that has not emphasized enough the role of the speaker
The Conceptual and Linguistic Facets of Persuasive Arguments 4 D. Marcu

You might also like