Distinctive features of persuasion and deliberation dialogues

2.1.Scenario 1: persuasion

In the persuasion dialogue, initially Harry wants to go to the Thai Palace, Jane wants to go to La Zingara and George to Nosh.

2.2.Scenario 2: deliberation

2.3.Commentary on the scenarios

In the dialogues, arguments are put forward that take the form of options presented that should be accepted on the basis of having certain features. Whether they have these features is intended to be an objective matter, whereas whether these features make the option desirable is a subjective matter about which agents may rationally agree to disagree.

The two dialogues are very similar. They use the same speech acts, many of the arguments are the same in both dialogues and the underlying views of the three participants do not change between the dialogues. But there are some crucial differences.

In the persuasion dialogue, it is Harry who wants to go the Thai Palace, and so Harry advances all the pro arguments, while George and Jane make objections. Their objections indicate the points on which they wish to be convinced. George objects to the distance, which could be a matter of time or money. All of them want a good meal, and so George abandons Nosh when his reason for thinking the food is good there is rebutted. Jane is concerned that the food be healthy. George is persuaded on grounds of cost, and Jane accepts once her doubts about the healthiness of Thai cuisine have been met.

Whereas in the persuasion scenario, Harry is committed to making a case for the Thai Palace, in the deliberation scenario, no one is committed to any of the proposals, although they may have their preferences. In this dialogue, first several options emerge, neutrally, without any commitment on the part of the person introducing them, and then a number of criteria emerge together with how the options stand in relation to these criteria. While individual preferences may inform the options and criteria put forward by the various agents, they play no explicit role, unlike in persuasion where the preferences of the agent being persuaded must be satisfied if persuasion is to occur. In deliberation, some criteria will emerge as more important than others: above, wine and atmosphere are established as more important than cost. Finally, the options are assessed according to the criteria that have been chosen in the discussion, and the best fit is chosen by mutual agreement.

We will base our dialogues on the performatives given for persuasion dialogues in Prakken (2006). The form of these performatives carries with it the strong presupposition that the knowledge of the participants will be represented in the form of facts and rules (at least some of which are defeasible) and that the participants are able to reason with these. In order to make our example concrete, we will provide an example of the facts and rules sufficient to produce the dialogues in the above scenarios. We will therefore provide knowledge bases for our participants. We will use Prolog, as the most readily available logic programming language, to express these knowledge bases.

3.1.Example scenario

Let us suppose that each agent ag_i has an individual knowledge base KBagi, comprising some facts Fagi and some rules Ragi, which is expressed as a Prolog program:

These rules may relate to matters of fact (e.g. R4 below) or procedures (e.g. R7 below). Here, the individual KBs should differ only according to differences in the completeness of knowledge. Rules may also be heuristics, like R2 below, which some individuals may adopt and others may not adopt. Importantly, rules also express the preferences of an individual, such as HR1, GR1 and JR1 below. These preference rules will normally differ between individuals, as tastes and aspirations are expected to differ. Now consider a persuasion in which proponent P is attempting to persuade some other agent opponent O, that some conclusion c(a) is the case. Let demo(Π,c(a)) be true if and only if c(a) can be shown from Π. Now the initial position is demo(ΠP,c(a)) and not demo(ΠO,c(a)) and the desired result of the dialogue is that demo(ΠO∪Δ,c(a)) where Δ⊆ΠP.

The dialogue is thus supposed to elicit a suitable set of facts and rules from P to enable O to derive c(a) for some a.

In the persuasion scenario, the conclusion that Harry is trying to establish is acceptableRestaurant(thaiPalace) (which is an example of c(a)), on the basis of a number of facts describing the various restaurants. We can begin by representing the attributes of a restaurant. Let these be

Putting these into Prolog, we obtain, for the example scenarios:33

Each of the three participants will have his or her own, individual, notions as to what makes an acceptable restaurant. These are expressed as, for example, the following, where HR1 denotes Harry's rule, GR1 denotes George's rule and JR1 denotes Jane's rule:

We also need an additional rule to define cheaper (where cheaper(X,Y) stands for “Y is cheaper than X”) and, for completeness, will also add one for quicker (where quicker(X,Y) stands for “Y is quicker than X”).

Thus, Harry desires only Thai food, George wants good food, but at the lowest cost possible, and Jane wants good food, but it must be healthy.

We can now use these facts and rules to construct the initial knowledge bases of Harry, Jane and George.

These are sufficient to generate the dialogue given in Section 2.1.

What is crucial in making this dialogue a persuasion dialogue is the presence of the individual rules HR1, GR1 and JR1. During the course of the dialogue, the participants come to know things that they did not know before, both by being told and by being given information that allows them to deduce new information. For example, Harry learns that the total cost of the Thai Palace is $25, when he is informed of the taxi fare. But these are not matters of persuasion, but rather of information sharing. Persuasion is necessary because the participants, quite properly, have their own idea of what makes a restaurant acceptable, and it is their own personal criteria that must be satisfied if they are to be persuaded. During the dialogue, participants commit to criteria representing their “point of view” or position. These differences in criteria are the source of the conflict that necessitates persuasion and, for persuasion to be successful, the proponent must demonstrate that his option satisfies the criteria to which the opponent has committed in the dialogue. Moreover, it is this that drives the utterances: when George and Jane contribute facts or rules, their contributions are guided by their personal criteria. Thus in step 2, George contributes F2 because he is interested in discovering the total cost of the various options. When Jane mentions the alternative in step 6, it is because La Zingara does satisfy her criteria, and she needs to be convinced that the Thai Palace also satisfies them before she can be persuaded that it is acceptable. For this reason, she invites Harry to demonstrate that the Thai Palace food is also healthy.

3.2.Features of a persuasion dialogue

We can now identify some features of a persuasion dialogue (considering persuasion involving preferences rather than concerning matters of facts, when the dialogue is correcting misconceptions or completing information, rather than persuading44):

We may further note that information may be solicited by a direct question as to whether a satisfies one of the conditions of the rule or by a claim that a does not satisfy one of the conditions of the rule or by a statement that some b≠a satisfies the condition. In turn, P should provide objective information to show that a does satisfy the condition demanded by O.

3.3.Features of a deliberation dialogue

We now turn to deliberation. The first point to note is that at the outset of the dialogue, the claim is uninstantiated. The task of the dialogue is to find some a for which c(a).66 Now, at the outset, even though Harry, Jane and George can all find an instantiation of acceptableRestaurant(X), they do not commit to persuading the others. Instead, they make general observations as to what might be important features of restaurants, with a view to finding some agreed rule AR1 which can be used to find an a for which c(a) can be derived.77

In the scenario:

Thus in the course of the dialogue, the participants have constructed a rule:

Note that the agreed rule does not contain Jane's condition relating to health, Harry's relating to cuisine or George's relating to cost. In fact, the private conditions of Jane and George are satisfied, so it seems that Harry has made a concession. In fact, however, no one endorsed his suggestion that the style of cuisine mattered, and his other suggested criteria of wine and distance were used, so he may be considered compensated. This may be particularly so if Harry also had rules in Π_H such as

3.4.Distinguishing the dialogues

In the deliberation dialogue, all the participants use their private rules to inform their contributions, but they are not allowed to insist on them: it is important to establish a rule that is mutually accepted and capable of being satisfied by at least one option. Thus, we can contrast Jane's move in step 3 of the deliberation dialogue with her similar move in step 6 of the persuasion. In the persuasion dialogue, she is indicating that good quality is a condition of her private rule and that no restaurant which fails to satisfy this condition will be acceptable to her. In the deliberation dialogue, she is rather proposing quality as a condition for the agreed rule. In the three-person dialogue under discussion, conditions are brought into the agreed rule by a kind of vote: one person puts them forward and another endorses them: here Harry agrees that quality is important in step 7. In this way, the condition has the support of a majority. As well as establishing the conditions to be satisfied, the dialogue establishes which options satisfy them. Moreover, because the options are not owned in the deliberation as they are in the persuasion, and because the setting is cooperative rather than adversarial, information showing them to be satisfied can be contributed by anyone, not just by the person who introduced the option. Thus, Jane introduces atmosphere in step 5, and George shows that La Zingara satisfies it in step 7. Finally, note that the final discrimination is a new criterion, not in any of the individual preference rules: this is an example of how the preferences may be a product of the reasoning rather than an input to it (see Searle (2003) for a discussion of this point).

5.1.Persuasion

Persuasion arises from disagreement: at least one participant believes that a particular option is acceptable and one or more others do not. How can such a disagreement arise? Some disagreements are inevitable: since agents may look for different criteria, it will often be the case that an option acceptable to one agent will not be acceptable to another. In such a case, persuasion should fail. But as well as these differences in preferences, there will also be differences in information. One agent may well not recognise the best available option through ignorance of some facts about the various options, so that an option may in fact best satisfy its criteria without it being aware that this is so. Note, however, that it is the criteria of the agent being persuaded that matter: the reasons that attract the persuader to the option may differ. Recall the persuasion scenario presented above, in Section 2.

During the course of that dialogue, Jane and George reveal their individual criteria. In P3, George's claim that the Thai Palace is far away is made because cost is one of his criteria in GR1, and his claim that Nosh is of good quality in P5 indicates the other condition in GR1. Similarly, Jane reveals the two criteria in JR1 in P7. Thus, in order to be relevant, claims made by O 88 must be required to establish the conclusion using the individual preference rule of O.99 Similarly, if O seeks a reason for one of P’s claims, O is accepting that that claim is relevant to the satisfaction of O’s individual preference rule. The key point here is that several of the moves have, in this context, implications that go beyond the basic conditions described in the previous section. Also these pragmatic implications differ according to the role of the speaker. Therefore, the opponent will only make claims relating to features that may lead him to prefer a particular option. So, any claim will correspond to a feature (say p) in the body of O’s preference rule.

Thus, we have the additional pre-conditions for O to make the move claim(O,p(a)):

Similarly, the post-conditions now also include this information:

The situation is the same when O poses questions. Again, O should ask only about criteria which he in fact finds desirable. Thus, question(O,p(a)) also gets the additional pre-conditions:

The agent doing the persuasion, P, can make claims about the attributes of the favoured option, but these will only be relevant to O if they form part of the body of O’s preference rule. If the attribute is not a criterion for O, it can be simply ignored. If, however, it is an attribute that forms part of the body of O’s preference rule and so matters to O, then O should accept the claim with concede. In doing so, O becomes committed to this criterion. Thus, concede(O,p(a)), like question(O,p(a)), has the additional pre-condition

and post-condition

Two key points emerge here. One is the asymmetry present in the dialogue. This reflects the asymmetry of the treatment of the criteria in the bodies of the agents’ individual preference rules: it is the criteria held by O that are important, for it is these that must be shown to be satisfied in order for O to be persuaded. Persuasion arises out of incomplete information, and so P is trying to supply information that O does not have but which would satisfy its criteria. Similarly, O is interested in indicating the criteria it requires information about. Thus, O can use claim, question and concede to indicate the criteria in which it is interested, because they will satisfy its preference rule. In contrast, P has no need to indicate its own criteria, but should avoid supplying unwanted information relating to features in which O is not interested. The purpose of the pragmatic implicature of the moves in a persuasion dialogue is to maintain relevance by informing P of O’s preference rule, so that P can be aware of the facts that O needs to adopt the desired preference.

In the next sub-section, we look at the additional conditions that apply to deliberation.

5.2.Deliberation

We now consider deliberation using the second scenario given in Section 2.2. In this dialogue, the participants begin by making claims about certain features of restaurants. When they make these claims, they do not in fact care whether the facts are already known to the other agents. This might seem redundant, but these claims are not primarily intended to impart information, although they do have this effect for agents who happen to be unaware of the information. These claims are meant to indicate criteria that the speaker would be happy to be included in the group criteria, that is, to appear in the body of the agreed preference rule that the deliberation dialogue is intended to construct. Claim therefore shares the pre- and post-conditions of claim as used by O in a persuasion dialogue. We can thus see moves like D2–D4 as effectively proposing criteria, as well as sharing awareness of the attributes of the various restaurants. As with persuasion, criteria can also be introduced through a question move, if the participant wants to propose an attribute for inclusion in the decision rule, but does not know of an option satisfying it.

In D13, George proposes a concrete rule satisfying some of the criteria previously advanced in the dialogue. Here, George is using the since move and so making claims about the criteria which are, or should be, acceptable to the group. Similarly, in D15, Harry uses a since move to add an extra, discriminating, criterion to the agreed rule.

The key move here is since. Since makes a commitment to the group criteria found in the body of the agreed preference rule AR1. In order to propose that a criterion be adopted as part of the agreed rule, it must previously have been made into a candidate by being introduced to the discussion. Moreover, this introduction should have been by some other agent, since otherwise there would be no reason to suppose general support for this criterion. The additional pre-conditions for since(ag_i,g(a),p(a)) (where g(a) is the goal of the deliberation, e.g. in the scenario acceptableRestaurant(a)) in deliberation are therefore

This reflects that the option should be acceptable to the agent providing the reason, but the criterion should have previously been committed to by some other agent. There is also the additional post-condition

Similarly, if the agent uses concede, it recognises that the conceded attribute should be part of the agreed rule. That the conceded fact is already in some other agent's commitment store and not in that of the conceding agent is already part of the general pre-conditions. In deliberation, however, there is an additional post-condition. Thus, concede(ag_i,p(a)) has the additional post-condition

6.1.Discussing matters of fact

If we consider persuasion as described in Prakken (2006), persuasion takes place in the following ways:

At least one of these must occur for persuasion about a factual matter to happen, and in a complicated persuasion dialogue, all three may be required. This requires certain further assumptions about the context: that the beliefs of the participants are individually incomplete or collectively inconsistent. That is, the beliefs of the participants are defective in some way and are corrected or completed through the dialogue. In our persuasion scenario, all three means of persuasion occur: facts about the options are established, rules allowing relevant facts to be deduced are exchanged and misconceptions are corrected (although this is done through exchanging rules, so blocking inferences previously possible using negation as failure).

Note, however, that throughout our persuasion scenario, the agents are discussing matters of fact, rather than subjective preference, and the desired conclusion can be shown only if it can be shown from the union of the knowledge base of O (which contains the required preference rule) with a subset of the knowledge base of P. Again persuasion exhibits an asymmetry, in that the persuader is able to conceal knowledge that would be harmful to its cause. If O favours a criterion that is, unknown to O, satisfied by its currently preferred option, P would be best advised to conceal the information needed to establish this. Note also that P has an additional means of persuasion available. Recall that when O makes a claim, it reveals not only a criterion, but also an acceptable option. P can attempt to show that the currently preferred option does not, in fact, satisfy this criterion, thus advancing the case for its own option by making its rival less acceptable to O. It is of course important that some option be acceptable (even if it is the lesser of two evils), and so if O’s current preference rule turns out to be unsatisfiable, another rule must be used. In practice, O will have a number of different preference rules, ordered by their priority. This is further discussed in Section 6.3.

With respect to deliberation, the situation is different in so far as there is no asymmetry, and the agents should wish the knowledge of all agents to be as complete as possible. Exploring the truth or falsity of matters of fact thus becomes an inquiry dialogue rather than a persuasion. For this purpose, the inquiry dialogue of Black and Hunter (2009) would be suitable, since it effectively allows all and only those inferences that can be produced from the union of the knowledge bases of the agents concerned.

6.2.Shifts between deliberation and persuasion

In our scenarios given in Section 2, we have tried to stay within one dialogue type. Often, however, a dialogue will contain nested dialogues, of the same or different type, and dialogues may shift, licitly or illicitly, between types. The most usual type of dialogue to be embedded within persuasion and deliberation will be information seeking. In both persuasion and deliberation, facts need to be established, and if these are unknown to the participants, the information may need to be sought elsewhere. For example, if no one knew whether Thai cuisine was healthy or not, and Harry was unwilling to concede defeat in the persuasion scenario, he could initiate an information-seeking dialogue with an agent with the requisite dietary knowledge. Harry's move in P6, where he corrects George's misconception about the quality of Nosh, could be seen as an embedded persuasion dialogue, albeit persuasion as to a matter of fact. In a cooperative dialogue, where there is no question of the strategic suppression of information for persuasive purposes, there is little difference between information seeking and persuasion as to matters of fact. The embedding of information seeking (and persuasion about matters of fact) forms a natural and necessary part of both persuasion and deliberation, since the facts must be established before the decision is made.

More interesting are shifts from one type of dialogue to another. Sometimes, a shift may become necessary: if a deliberation fails to achieve a consensus, then a shift to persuasion or, more usually, negotiation may be required to reach a decision. Some shifts, however, particularly if illicit or unnoticed, are a source of fallacies, as is emphasised in Walton and Krabbe (1995). As an example, Walton and Krabbe describe (pp. 109–110) how argumentum ad baculum (appeal to force) can be seen as an illicit shift from a persuasion dialogue (in which threats are inappropriate) to a negotiation dialogue, where threats can have a legitimate role. What would be the consequence of a shift from persuasion to deliberation? Here, we might see P, recognising that his option does not in fact satisfy O’s criteria, attempting to make a shift to a deliberation dialogue. If successful, O can no longer decide on the basis of O’s own preference rule, but rather has to use some consensual rule that may lack the unsatisfiable criteria of O’s rule and perhaps instead contain some criteria from P’s preference rule. If O fails to notice the shift and accepts the shift to deliberation, then O may fail to reject P’s favoured option, as would be quite proper in a persuasion dialogue, where O is entitled to insist that his own criteria be satisfied. Thus, shifting from persuasion to deliberation effectively removes the burden of proof from P and gives both participants equal power in the decision-making, whereas O has the right be the arbiter in a persuasion dialogue.

Also common is a shift from deliberation to persuasion. Here, the shift can arise in two ways. First an agent might adopt the role of persuader. Whereas in a deliberation there should be no ‘ownership’ of options, an agent may lose sight of this and become the champion of a particular option. There might be little harm in this (since the persuader is effectively relinquishing the power to decide to the other participants), but problems can still arise because the commitments made by the agents differ between the dialogue types. Thus, in our scenario, in D7, Jane claims that Nosh lacks atmosphere. In the context of deliberation, this commits her to no more than a willingness to see atmosphere as part of the collective rule. If, however, this had been a persuasion dialogue, she would have been committed to atmosphere as being part of her decision rule and so to rejecting any option without atmosphere. In this way, an illicit shift to persuasion could force Jane to reject an option lacking atmosphere, even though this is not required by her preference rule and even though this had not been adopted as part of the collective decision rule. So, this shift is somewhat two edged. By moving from a participant in a deliberation to the persuader in a persuasion, the agent gives up some decision-making power, but can force unintended commitments on the other participants, if they fail to notice the shift. This may have strategic advantage: moreover, a subsequent shift back to deliberation can then leave the opponents with the unwelcome commitments, but without the extra decision-making power.

A second way in which an agent may try to shift from deliberation to persuasion is through attempting to assume the role of O. The object here is to arrogate the decision-making power for itself. If the shift is successful, it will be able to use its own preference rule as the collective rule and foist a burden of proof on the other agents. In such a situation, the agent should be aware that in making the shift it will become committed to any criteria it has previously introduced into the dialogue.

Thus, shifts from one dialogue type to another are something that participants in a dialogue need to be aware of, and avoid, if they are to avoid having different rules from those they intended applied to their utterances and if the objectives of the dialogue are to remain constant.1111 Shifts, and further examples of analyses of them, are further discussed in Walton (2009).

6.3.Reasons for criteria

Supplying reasons for why an agent adopted the criteria used in its preference rules, using, for example, the since move, might also be used to enrich the possibilities for persuasion. If the agent currently has a rule that it is unable to apply, perhaps because some condition is known to be satisfied, then such information might lead it to adopt a different rule. For example, if an agent explains its preference for Thai food on the grounds of its health-giving properties, other agents may also develop a preference for that cuisine. Equally an agent may abandon criteria if they are shown to be based on an incorrect inference: for example, if it discovers that French food is not so healthy after all. There is, however, an important difference between preference rules and factual rules when P has a rule that O does not currently recognise. Where a rule concerns a matter of fact, the agents should adopt it (we are assuming that all the participants are cooperative and sincere1212), but this is by no means clear when it concerns a reason for desiring something, that is, the individual preference rules. Suppose, for example, an agent suggests that having Indian cuisine is a reason for favouring a restaurant and is then asked to justify this. It might well reply that Indian cuisine should be a criterion because it is spicy. This can be a perfectly acceptable reason for one person and yet unconvincing to another whose digestion is delicate and whose taste lies towards different food. Thus, eliciting reasons for criteria from O through use of why, and then attempting to satisfy them, is more likely to be persuasive than proposing rules that are themselves subjective. This, however, represents a level beyond what we have thus far been discussing: we are now in the realms of reasons for reasons, and this we will leave for further exploration in future work.

In deliberation also, proposal of rules for endorsing criteria may be useful in encouraging others to support their adoption by the group, but will not be decisive, since these rules can legitimately reflect the taste of a single agent. Agents may even support criteria for different reasons: one may like Italian food because it is healthy, while another simply likes the taste. Here, the subjective element means that we cannot derive the criteria of the group from the union of the knowledge bases of the group and the union of the rules for adopting criteria. Instead, the criteria for a condition being adopted as part of the collective preference rules needs to be endorsed by each agent according to its own tastes.

6.4.Preference functions

In the examples given above, the individual preference rules have been straightforward: the restaurant is acceptable if it satisfies some list of criteria. In practice, however, criteria will differ in importance, and so determining acceptability will be a more complicated procedure, which may involve different approaches such as voting methods, as discussed in McBurney et al. (2007), or quantitative approaches to handling preference change (Liu 2009). Relating more to our framework is the potential for the use of several rules, and weighted criteria, which we discuss below.

First, we should distinguish between essential and desirable criteria. If, for example, one has only a limited time to eat (perhaps during the lunch break of a conference), then quick service may be an essential feature of the restaurant and may lead to it being chosen no matter how less desirable than its rivals it is in other respects. The key point here is that an acceptable option must satisfy all essential criteria; desirable criteria can then be used to discriminate between those options that meet these essential criteria. The implications for persuasion are that the essential criteria must be shown to be met or shown not to be essential. If quick service is said to be essential and an agent is attempting to persuade to an option that does not have this feature, the reason must be shown to be inapplicable or at least not essential. For example, if quick service is posited as essential because the conference restarts in an hour, it has to be argued that the rule either is not satisfied (e.g. because the lunch break is in fact two hours) or need not be applied (e.g. because the after-lunch session is of no interest). If such moves are successful, the criterion is no longer essential: it may be abandoned altogether (ok, so we can take as long as we like) or retained as a desirable criterion (ok, but I still like quick service). Some mechanism for making this distinction needs to be provided. In our simple treatment, all criteria are essential. If, however, we added another rule, say GR2, to be evaluated before the current GR1, with an additional criterion in the body, this would represent a desirable attribute. We could add several rules in this way, indicating different combinations of desirable criteria, with their degree of desirability being indicated (in our Prolog programs) by the order in which they occur. There must always, however, be at least one rule, and this will contain the essential criteria, and all the other rules must include these criteria in their bodies.

In deliberation, a similar distinction between essential and desirable attributes can be drawn and several rules proposed. Here, however, if even a single participant can claim (and defend the claim) that a criterion is essential for him, then that criterion has to be adopted as essential by the group. If only one member of the group has to catch a train, then it is essential to go somewhere with quick service or abandon the notion of eating together.

If we allow weights on criteria, then the acceptability of an option may turn not on whether an option satisfies a criterion or not, but on which of the two criteria is preferred. Thus, it may be accepted that two restaurants are acceptable in all respects, except that one is close by and the other has quick service. Now persuasion may be a matter of inducing the persuadee to change his preference order from closeby≺quickService to quickService≺closeby. This in turn suggests other rules, the conclusions of which will be such preference orderings. For example, we could argue that if taxis are readily available, then closeby≺quickService. Like reasons for including and excluding criteria, however, such rules are unlikely to be coercive: an agent is always free to reject them. This then is a matter for genuine persuasion rather than inquiry.1313

Less clear cut is what happens to preference orders in deliberation. Each participating agent will have its own preference order: how can these be combined to give a group order? Of course, arguments can be presented as above to attempt to align them, but there is no guarantee that such arguments will succeed. It is possible to merge the orderings into a partial order consistent with all the individual orders, but this may be insufficient to resolve issues. This then may prove a problem for deliberation, especially if too many criteria are included in those accepted by the group. The question of reaching an agreement on numeric weights rather than an ordering is, we would suggest, too difficult to contemplate, which is why numeric weights do not in practice feature in such discussions. Some discussion of these issues can be found in Bench-Capon and Prakken (2010) and Bench-Capon, Prakken, and Visser (2011), but since our main purpose in this paper is to characterise the different kinds of dialogues rather than explore the nature of preference functions, the single rule expressing essential conditions is adequate for our current purposes.