Towards a framework for computational persuasion with applications in behaviour change 1

3.1. Rationality of argumentation

The study of argumentation has largely focused on what constitute good arguments in a normative sense. This involves identifying the features of good arguments that would be appropriate for honest rational agents to present or accept. For an excellent review, see [146]. The careful presentation of premises, and the use of logical reasoning, is much espoused for good quality argumentation. In order to make a case in professional life (such as politics, academia, business, journalism, etc), it is seen as an essential ability. This has led to techniques for constructing and deconstructing good arguments (see for example [48,66]).

In parallel with elucidating what constitutes a good argument, the study of argumentation has also identified types of poor or inappropriate argumentation such as argumentation fallacies (as well as [146], see reviews in [58,145]). Some fallacies are described as formal fallacies since they violate rules of logic (e.g. claiming the antecedent of an implication is true because its consequent is true), or violate rules of probability theory (e.g. the gambler’s fallacy), but many fallacies are informal since they constitute what could be described as unacceptable arguments, such as begging the question (which is providing a version of the claim as a premise, and using that to derive the claim), argumentum ad hominem (arguing against the arguer rather than their arguments), and appeal to authority (arguing that an argument is true because of the position of the arguer).

Given that the study of argumentation has developed such a comprehensive understanding of how to differentiate good from bad argumentation, if we are dealing with rational agents when undertaking persuasion, it appears important to use good arguments and argumentation, and avoid poor arguments and argumentation.

Furthermore, the overall quality of the argumentation is important from a psychological point of view (see for example [64,67]). If a persuader wants to convince the persuadee of an argument (a persuasion argument), then this includes acceptability of the persuasion argument (against counterarguments), believing the premises of the persuasion argument, fit of persuasion argument with agenda, goals, preferences, etc, quality of constellation of arguments considered (balance, depth, breadth, understandability, etc). Comprehensibility of persuasive arguments has also been shown to be an important determinant of successful persuasion [45].

Despite the importance of being rational in argumentation, there is a tendency in the computational models of argument community to overemphasize the need to be rational. As we will cover later, the role of emotional arguments have been found from psychological studies to be important in persuasion. Furthermore, even for rational argumentation, some widely accepted principles such as not deploying ad hominem arguments can be over-restrictive when for example the speaker has a poor reputation or is unqualified in the topic that they are speaking on, though use of argument schema with critical questions goes someway to redressing the balance in this (for a review of argument schema, see [148]).

3.2. Persuasion techniques

Psychological studies have identified persuasion techniques that are seen in human interactions in general. For instance, Cialdini has identified the following six principle of influence [34].

The above principles are supported by empirical studies in psychology. Other studies suggest further principles that appear to useful in various domains such as law (e.g. [52]), healthcare (e.g. [133]), and ecommerce (e.g. [105]).

3.3. Argumentation style

Argumentation style concerns who is presenting the arguments, the language used in those arguments, and the way the dialogue is structured. The issues raised touch on broader aspects of the personality of the persuadee, and the context of the persuasion (see for example [132] for further discussion of these issues).

Determining the precise parameters of argumentation style raises some complex issues for effective persuasion in a specific application. Experiential knowledge and psychological studies offer some general guidance but in practice the specific parameters need to be determined for a specific application.

3.4. Framing of arguments

The framing style of an argument concerns the exact phrasing of the content which can influence how it is perceived. It is increasingly acknowledged in healthcare professions that interactions with patients should be effective communication that is respectful and non-judgmental (see for example [106]). Such a stance would be a natural point for automated persuasion systems. However, this still leaves much latitude in how arguments can be framed.

Consider for example experiments by Tversky and Kahneman where participants are asked to imagine preparations for the outbreak of an unusual disease that is expected to kill 600 people, and to consider two alternative programmes to combat the disease [141]. In the first framing of the study (below), the majority of participants prefer Programme A.

In contrast, in the second framing of the programmes (below), the majority of participants prefer Program D. So even though, the versions above and below are providing the same information, participants can change their preferences based on the framing. The explanation is that in the context of gain (above) people tend to be averse to risk and in the context of loss (below) they try to minimize the loss.

In a user study on the persuasiveness of healthy eating messages [138], positively framed messages (e.g. Most people believe that eating a healthy breakfast contributes to a longer lifespan) were shown to be more efficacious than negatively framed messages (e.g. Most people believe that eating an unhealthy breakfast contributes to a shorter lifespan). Furthermore, Cialdini’s principles of persuasion [34] were considered (i.e. reciprocation, commitment, consensus, liking, authority, and scarcity), and it was found that arguments that appeal to authority (e.g. Studies conducted by health experts have shown that eating a healthy breakfast keeps you energized) were the most persuasive.

3.5. Personality of persuadee

Another key dimension that can affect the success of persuasion is the personality of the persuadee. Obviously, this is not in the control of the persuadee, but if the persuader knows about the personality of the persuadee, the persuader can make a better choice of strategy to use with the persuadee. Consider for example persuading someone to vote in the national election: If the person “follows the crowd”, then telling them that the majority of the population voted in the last election is more likely to get them out to vote, whereas if the person “follows rules rigorously”, then telling them that it is their duty to vote is more likely to get them out to vote. Mistaking the personality trait can have a negative effect on the chances of successful persuasion. Psychology has developed numerous methods for characterizing personality. An important example is the model based on the OCEAN personality traits which are Openness to experience, Conscientiousness, Extroversion, Agreeableness, and Neuroticism [53].

There are many other ways that knowing more about the persuadee can be used in determining an appropriate strategy for persuasion. Attitudes are important determiners in persuasion, and the psychology of attitudes may provide important insights in a persuasion strategy (for a review see [97]). Various other aspects of psychology are routinely harnessed in commerce, particularly ecommerce, such as the psychology of colour, language, graphics, pricing, and negotiation (for a review see [105]). Cultural information about a persuadee, such as nationality, can give important information about the kinds of interaction that a persuadee might respond positively to (see for example [65]). Other kinds of traits, such as political, ethical, or religious, may be important determiners in some contexts for persuasion. Furthermore, detailed information about multi-dimensional private traits can be predicted from digital records of the user (e.g. from Facebook likes) [88].

3.6. Emotion invoked by arguments

Presenting emotional arguments can be important. Emotional arguments are predominantly deployed by the persuader in order to influence the persuadee via emotional devices. For example,

Note, emotional arguments contrast with evidential/logical arguments, such as for example, the following argument.

In certain situations, emotional arguments can be powerful arguments in persuasion. For instance, Lukin et al [95] have shown that with some audiences, emotional arguments are more effective in persuasion than factual arguments. For this, they categorized audiences according to the OCEAN personality traits (i.e. openness to experience, extroversion, agreeableness, conscientiousness, and neuroticism), and showed that conscientious, open, and agreeable people are more convinced by emotional arguments.

4.1. What do computational models of argument offer?

Computational persuasion is based on computational models of argument. These models are being developed to reflect aspects of how humans use conflicting information by constructing and analyzing arguments. A number of models have been developed, and some basic principles established. We can group much of this work in four levels as follows (with only examples of relevant citations).

4.1.1. Dialectical level

Dialectics is concerned with determining which arguments “win” in some sense. In abstract argumentation, originally proposed in the seminal work by Dung [42], arguments and counterarguments can be represented by a graph. Each node denotes an argument, and each arc denotes one argument attacking another argument. Dung defined some principled ways to identify extensions of an argument graph. Each extension is a subset of arguments that together act as a coalition against attacks by other arguments. An argument in an extension is, in a sense, acceptable. For a review of Dung’s approach and alternatives, see [8]. Labels can also be assigned to arguments, and Caminada and Gabbay [27] provided a formalization based on three labels (in, out and undecided) that they show is equivalent to Dung’s formalisation.

There have been numerous developments of abstract argumentation that include alternative definitions for extensions such as ranking-based semantics [1,24,120], and introduction of attacks on attacks [7,103], preferences [2], weights on attacks [44], support relations (see for example [30,31,109] and see Fig. 1 for an example of an argument graph with supporting and attacking arguments), and probabilities [43,72,74,82,92,135]. Furthermore, there has been the development of software solvers for determining extensions (see for example [33,137]), and the application of natural language processing techniques for constructing argument graphs from free text (see for example [93]). In addition, there are methods for argument dynamics to ensure that specific arguments hold in the extensions of the argument graph such as epistemic enforcement in abstract argumentation [9,10,38], revision of argument graphs [36,37], and belief revision in argumentation (e.g. [18,29,41,50]).

OPEN IN VIEWER

Fig. 1.

A simple example of a bipolar argument graph for application for persuading a user to undertake more exercise. A dotted arc from a node A to a node B denotes that argument A supports argument B. A solid arc from a node A to a node B denotes that argument A attacks argument B (i.e. A is a counterargument B).

4.2. Shortcomings in the state of the art

So computational models of argument offer a range of formal systems for generating and comparing arguments, and for undertaking this in a dialogue. However there are shortcomings in the state of the art of computational models of argument for application in persuasion. The current state of the literature does not adequately offer the following and hence there are some exciting research challenges to be addressed if we are to deliver computational persuasion.

In order to focus research on addressing these shortcomings, we can consider how computational persuasion can be developed and evaluated in the context of behaviour change applications.

4.3. Studies with participants

In order to have well-understood computational models of argument that correspond to human behaviour, there is a need to ground these models with studies with participants. The studies undertaken so far validate some aspects of these models, but also indicate some shortcomings in being able to model human behaviour.

Studies performed by Rahwan et al [124] and Cerruti et al [32] investigated various forms of reinstatement in argumentation. The users were presented several argument graphs and were asked to explain how acceptable a given argument is in their opinion. The results show that in some cases, the implicit knowledge about domains can substantially affect the given acceptability levels. However, more importantly, the experiments show that the attacked argument’s acceptability is lowered, but does not fall to 0, which is what would be predicted by the usual dialectical semantics for abstract argumentation. Additionally, introducing the defense for this argument raises its acceptability. However, typically it does not reach the value of 1, which is the level the usual dialectical semantics would predict.

In a study of argumentation dialogues, Rosenfeld and Kraus [129] undertook an experiment in order to develop a machine learning-based approach to predict the next move a participant would make in a dialogue. This work was further extended in [130,131]. The machine learning models were trained on data that incorporated the sequences of arguments in a dialogue that the participants accept. Once trained, the models were able to predict the acceptance an unseen case would have.

In another machine learning-based approach, Huang and Lin [68] developed a software agent for participating in dialogues with potential customers with the aim of persuading them to offer a higher price for goods. Dialogues were constructed from an argument graph, and training was done on simulated scenarios. In testing with users, the agent was able to persuade the participants to increase the mean price offer.

There are also studies with participants by Masthoff and co-workers that investigate the efficacy of using arguments as a way of persuading people when compared with other counselling methods indicating that argumentation may have disadvantages if used inappropriately [107], and that rather than a confrontational approach, argumentation that is based on appeal to friends, appeal to group, or appeal to fun, may be more efficacious [142,143].

Emotion in argumentation has also be the subject of a study with participants in a debate where the emotional state was estimated from EEG data and automated facial expression analysis. In this study, Benlamine et al [15] showed for instance that the number and the strength of arguments, attacks and supports exchanged between a participant could be correlated with particular emotions of the participant.

In our framework, we focus on the uncertainty surrounding the user’s awareness of arguments, and the user’s belief in the arguments s/he is aware of. For this, we have harnessed probabilistic argumentation. Two main approaches to probabilistic argumentation are the constellations and the epistemic approaches [74].

The epistemic approach has been extended with a probability distribution over subsets of the set of attacks [116]. This can be used to represent an agent’s belief in each attack. This is potentially useful for handling enthymemes. Since most arguments are presented in natural language, different agents may interpret them differently, and hence some agents may be belief that an attack holds between a pair of arguments whereas other agents might not.

7.2. Studies with participants

We have undertaken studies with participants to evaluate how they deal with arguments arising in a dialogue [115]. We asked each participant for their belief in the arguments at each stage of the dialogue, and whether they saw a negative (i.e. attacking) or positive (i.e. supporting) relationship between the latest argument added in the dialogue and the previous arguments. For this study, we were able to make a number of observations including the following.

We have also developed methods for acquiring crowd-sourced opinions on arguments, and shown how they can be used for predicting opinions on arguments [79]. We evaluated our approach by crowd-sourcing opinions from 50 participants about 30 arguments. This work shows how it is viable to acquire data from a number of contributers to construct classifiers, and that these classifiers can then be deployed to substantially decrease the number of questions that need to be asked of any particular user in a persuasion dialogue.

7.3. Strategic argumentation

Given the potential for probabilistic argumentation to capture key aspects of uncertainty in the user model, I indicate below how strategic argumentation can be developed to harness the user model.

Key possible dimensions for modelling uncertainty are summarized in Table 3. These developments offer a framework with a well-understood theoretical methodology, and implementations that are computationally viable for strategic argumentation.

The key research issues for our project are summarized in Fig. 4, and described below. The aim is to address these issues in order to provide an integrated theory for applications in behaviour change.

OPEN IN VIEWER

Fig. 4.

Key aspects of our framework for computational persuasion. A solid arrow indicates a necessary flow of information whereas a dotted arrow indicates an optional flow of information.

We have put uncertainty in arguments, in particular in belief in arguments, at the core of our framework for computational persuasion, as we believe this is a minimum necessary for persuasive behaviour. However, we believe that other dimensions are highly desirable for a more comprehensive framework for computational persuasion. In particular, some of the dimensions considered in Section 3 are potentially valuable including rationality of arguments (in particular quality of arguments and of argumentation), persuasion techniques, argumentation style, framing of arguments, and emotion of arguments.