Argument schemes for reasoning about trust^†

2.1.The schemes

We have the following argument schemes. Note that all schemes are stated in positive terms in that the schemes apply, they generate arguments for trusting an individual rather than arguments for not trusting, or for distrusting (we briefly discuss schemes for the latter in Section 6).

This argument scheme is similar to ‘Argument from commitment’ (Walton et al., 2008, p. 335), where an individual is argued to be committed to some proposition on the evidence that what they said or did shows that they are so committed. In the case of trust, past evidence is reliable behaviour and commitment is typically required to be repeated in order to reinforce trust.

This form of argument is distinguished from the previous case because of the fact that the experience is indirect. In the case of this restaurant example, I have not tried the food; I am just inferring from what I see that it is probably good. Unlike the DE case, my inference about the food could be wrong – the crowds might reflect the fact that it is a fashionable restaurant that serves mediocre food, that it has a famous DJ who regularly spins there, or that the owners are regularly comping33 indifferent meals for their friends in the hope that it will entice other people to come in.

In some cases (such as the case where B is a chef), A is also able to validate B’s trustworthiness for themself (by eating B’s food) and so augment an argument from expert opinion (based on the CIA diploma) with an argument from direct experience. In other cases, such as the case where B is a doctor, it is not clear that A can perform any direct validation; and instead, A always infers part of their trust in B as a result of seeing the certificate hanging on the wall in B’s office as an emblem of their expertise.

This argument scheme is clearly related to the ‘Argument from expert opinion’ from Walton et al. (2008, p. 15), and this relationship is discussed in more detail below.

Two further examples help to distinguish between the EO and Au trust schemes. An employee of the Metropolitan Transit Authority (MTA),44 for example, might well be trusted on questions about subway operation, on the grounds that their job means that they will have reliable and up-to-date information. This is trust from authority. However, if they no longer have any affiliation with the MTA (e.g. because they change jobs), then trust from authority would not be warranted. Although, if our MTA employee retired after 40 years of experience developing train schedules, then they could still be considered a trustworthy expert on train schedules long after they leave their job. Taking another example, consider a university professor who is invited by a reporter to comment on the subject in which they are regarded as an academic expert. The reason that they are trusted to answer questions about their area of academic research is because of their knowledge, certified, for example, by their PhD. However, if this professor also held an administrative role, such as head of postgraduate studies, and was asked questions about the requirements necessary for graduation, then they would be trusted as an authority, not as an expert.

We distinguish the idea of an argument based on reputation from an argument based on EO or Au because the recommenders are neither experts nor authorities. If A’s sources are experts then, in our view, A would be using the expert opinion scheme; similarly, if A’s sources represented organisations exercising authority, then A would be applying the trust from authority scheme. We also note that an argument from reputation is not the same as propagating trust. If A hears that B has a good reputation, this is a statement about B’s trustworthiness. It, like all the other schemes we consider here, is the establishment of a link between A and B. (Information derived from reputation may be used in propagating trust between individuals, but the use of the reputation scheme does not imply propagation.66)

For example, suppose I observe a customer, Elin, at the cash register in a shop. She has realised that the cashier returned too much change to her in a transaction, and she tells the cashier about it and returns the extra change that she received in error. From this observation, I determine that Elin is of sound moral nature. Thus, the MN trust scheme could be applied to any information I receive or interaction I anticipate from Elin in the future.

Now, while trust from pragmatism depends on A’s goal being served by trusting B, the scheme makes no reference to B’s goals. If the goals of the two agents are aligned, then there is an additional reason for A to trust B.

This form of trust, then, is all about having goals that align. As the last example showed, it is plausible to trust someone even if their beliefs differ significantly from one's own. However, it is also plausible to construct an argument for trusting someone on the grounds that their beliefs are similar to one's own beliefs, as in the next case.

Talking of evidence that can be verified makes Trust from Mutual Belief sound as if it might be related to Trust from Direct Experience, but this is not the case. DE captures trust that is derived from observing some action. MB captures trust that is derived from finding that another individual shares beliefs. They both might apply in some cases – as, for example, in the case when someone tells us their belief about some objectively verifiable fact like ‘I believe it is raining outside’. If we check this, and it is true, then we have direct experience of the individual as a source of information about what is going on outside. Then, because of a shared belief in current precipitation, we might be inclined to trust that same individual's view about how sensible it would be to venture outside (at least in comparison with how much one would trust someone who, unlike us, believes it is not raining).

Note that several of the schemes described here – Trust from Expert Opinion, Trust from Authority, Trust from Moral Nature, and Trust from Social Standing – might be considered to be specific instances of Trust from Majority Behaviour. The reason, after all, that we trust Dr B to advise us on medical matters is that most doctors are trustworthy on medical matters. What distinguishes arguments constructed from the Majority Behaviour scheme from those more specific cases just listed is that the latter are, in our particular cultural milieu, ones that are especially reliable. As we argued in the introduction, part of the issue in developing a representation of any domain is to identify the relevant set of concepts. The fact that we can distinguish these different forms of ‘majority behaviour’ is a good reason to highlight the schemes.

2.2.Discussion

First we should note that these schemes are all concerned with A establishing its own trust in an individual for themself, not deriving trust as a result of reports from others (as mentioned above, establishing schemes for trust propagation is the subject of Section 4). Second, as noted above, all the schemes are stated in the positive, providing reasons for trusting an individual. Here we do not consider schemes for not trusting an individual or for distrusting an individual – as mentioned above, such concerns are briefly considered in Section 6.

The schemes break down into five categories. The first category is one in which A has collected evidence about the trustworthiness of a given individual. This category includes two schemes, direct and indirect experience, relating to whether the evidence that A has collected is direct evidence of B’s trustworthiness, or evidence of something from which trustworthiness can be derived.

The second category includes three schemes – expert opinion, authority, and reputation – where B is considered to be trustworthy on some subject, or in some role, not because A has observed them doing this (directly or indirectly) but because there is some validation of B that can, in theory at least, be verified by A. If B is somehow validated as an expert or an authority, A can reasonably trust them. Of course, there are many cases of supposed experts or authority figures being shown to be frauds, but the fact that the argument scheme can lead to misplaced trust is not an argument against the scheme so much as an argument for careful posing of the critical questions. For example, concern about fraudulent experts might lead one to be vigilant in checking their certification. In the reputation scheme, A bases their decision on a distillation of reported evidence from other individuals. Since those individuals are not typically validated as experts or authorities (someone reviewing a restaurant on Yelp might be a chef, but it is hard to establish whether this is the case), what A relies on in the reputation scheme is the ‘wisdom of the crowd’ (Surowiecki, 2005). In other words, the average of a number of reported experiences is likely close to the experience that A will have if A is a typical member of the population who are providing the reviews.

It is important to distinguish reputation from reasoning based on referral trust. Reputation is established from many individuals and refers to one specific recommendation. Referral trust is trust in an individual's ability to make trustworthy recommendations, and so is established about one individual, by one of the mechanisms discussed here. Referral trust may be established by reputation, that is, A may decide to trust B to make referrals on the basis of good reports of B’s referral ability that are made by D and E. Note also the distinction between the reputation scheme and the majority scheme. A might decide to trust the quality of the food prepared at restaurant B because of good reports about B. That is an instance of the reputation scheme. Alternatively, A may decide to trust the quality of food at B because A already trusts the quality of the food at all the other restaurants in the neighbourhood of B. That is an instance of the majority scheme.

Furthermore concerning this category of schemes, there is clearly a relation with existing well-known schemes from the informal logic literature. Much has been written about schemes for expert opinion and appeal to authority in general (for example, (Walton, 1997; Walton et al., 2008). The general scheme for expert opinion set out in Walton et al. (2008) expresses that an expert in some subject domain asserts some proposition that can be defeasibly accepted as true:

As can be seen from the list above, one of the critical questions accompanying the scheme probes into the matter of the expert's trustworthiness. In attempting to answer this question, our scheme for trust from expert opinion can clearly be deployed. Most plainly, the trust scheme could be used to provide a positive answer to the trustworthiness question of the general scheme above in that the conclusion of the trust scheme confirms that the expert in question has indeed been validated as being a reliable source. For example, if critical question 4 above has been posed against an instantiation of argument from expert opinion concerning some purported medical expert E, then a response might be that E is a doctor who can be trusted as being a reliable source as concluded by the trust from expert opinion scheme. Of course, the conclusion of the trust from expert opinion scheme is itself open to critical questioning, as we discuss later in Section 3. Although this example demonstrates how a specific issue from a more general scheme can readily be employed to probe into the issue and as such suggests some connections between schemes, currently we do not wish to suggest any kind of taxonomic classification of schemes due to the inherent difficulties of adequately covering the many possible ways in which schemes can interact and relate. However, we are open to the possibility of our specialised schemes for trust being used in conjunction with more general ones where issues of trust arise but are not the central concern of the scheme, as is the case of Walton et al.’s scheme for expert opinion.

Returning to our categories of argument schemes, the third contains two schemes, moral nature (MN) and social standing (SS), that are based on A’s observations of B, but observations that are not directly linked to trust. A might use the MN scheme to infer that B, whom A has observed to be very correct in their dealings with others, will act in a trustworthy way. Similarly, A might use the SS scheme to infer that B, who is a pillar of the local community, will not default on a loan. In neither case does A have any information about B’s trustworthiness – that is, B’s past behaviour or creditworthiness – but is prepared to use aspects of B for which A does have information as a proxy for such evidence.

The next category contains four schemes: majority (M), mutual beliefs (MB), mutual goals (MG), and organisation (Org). We can think of the majority scheme as a way of extrapolating any of the previous schemes. If we can show that any of those schemes for deriving trustworthiness apply to a suitably high proportion of a given population for which we do have evidence, we may be happy to infer that some member of the population for which we do not have evidence is also trustworthy. The mutual beliefs, mutual goals, and organisation schemes are grouped with majority because all four of them are about deriving trust in an individual without very detailed knowledge of that individual. The majority scheme does this on the basis of the class the individual falls into, and the organisation scheme does exactly the same, but on the basis of the individual falling into the specific class defined by members of the organisation. The mutual beliefs and mutual goals schemes use knowledge of some aspect of an individual to place them into an appropriate category and infer trust from that.

The final category contains the schemes of pragmatism (Pra) and prudence (Pru). Neither of these makes inferences that have much to do with the individual in whom trust is inferred. Prudence determines trust by assessing the comparative risk of trusting versus not trusting where the reasoning is about the situation not the possibly trustworthy individual. Pragmatism derives trust from a consideration of A’s goals and whether they are better served by trusting B or not, not on any information about whether B is trustworthy in their own right.

Two other schemes we considered, but rejected, as possible schemes are ‘precedent’, where A has trusted B before and decides B can be trusted again because there were no bad consequences from the previous time(s) B was trusted, and ‘default’, where A decides that B can be trusted despite having no evidence, perhaps with something in mind like ‘tit-for-tat’ (Axelrod, 1984)88 or ‘innocent until proven guilty’. We did not list these schemes because we think they are examples of schemes we have already incorporated in our list of schemes rather than new schemes. Precedent is a form of direct experience, and as we shall see, one of the critical questions that applies to direct experience addresses B’s past behaviour. We consider default to be a form of the majority scheme. After all, if it was not the case that an agent believed that the majority of individuals it was going to interact with would be trustworthy, then the default scheme would not be rational. (In other words, the default scheme is only rational for populations for which the majority scheme suggests trusting individuals.)

Figure 1 shows a visual characterisation of our argumentation schemes that is based on a conceptual model of trust showing a trustor (A) and a trustee (B) interacting in an environment and in a society. Pragmatism, based on self-interest, is an attribute of the trustor and moral nature is an attribute of the trustee. Expert opinion links the trustee to the environment and prudence (viewed as suitability of decision-making) links the trustor to the environment. The trustor has direct experience of the trustee and may share mutual beliefs and mutual goals with them. The trustor also has indirect experience of the trustee via the environment and may know the organisation that the trustee belongs to. A trustee has social standing in the society and may be in a position of authority over the trustor. The society includes agents with resemblance to the trustor and trustee, respectively, and supports distinct trust relationships: reputation via the former and majority via the latter.

Figure 1.

Categorising argumentation schemes for trust.

The above formulation establishes a form of conceptual completeness for our schemes. Given the conceptual model of trust, we have schemes that capture all pairwise relationships between the elements of the conceptual model.99

4.1.The schemes

In the trust literature, inference about trust is often formulated in the context of a trust graph – a graph in which nodes are individuals and labelled, directed links between nodes indicate the amount of trust between individuals. We follow this tradition and illustrate each of the forms of inference as a graph as well as a text description.

Figure 2.

A graphical rendition of trust propagation: (a) direct propagation; (b) reciprocity; (c) co-citation; (d) co-implication; and (e) trust coupling. The solid arcs are the base relations, and the dotted arcs are the inferred relations. Arcs that end with dot indicate referral trust, and arcs that end in an arrowhead indicate functional trust.

Trusting takes place in a specific context, and as Jøsang et al. (2006) point out, in this kind of inference, we sometimes mix functional trust and referral trust. It is often the case, for example, that B trusts C to do a specific kind of thing, which is functional trust, and A infers that C can be trusted to do this because A trusts B to make good recommendations, which is referral trust. In the context of the restaurant recommendation example from Section 2, if I trust Bob as a judge of people, and Bob tells me that Camilla is a good judge of restaurants, I might decide to start judging restaurants by what Camilla says about them. Because my trust in Camilla is restricted to her knowledge of restaurants, I will not necessarily take any suggestions she might make about where to get my car serviced. I might, however, take Camilla's recommendations about where to go for a drink, on the grounds that her knowledge about bars is suitably close to her proven area of expertise (knowledge about restaurants). In contrast, if Bob tells me that Deborah knows a good mechanic, then because I know that Bob can be trusted on referrals, I may decide to take any suggestion that Deborah makes about where to get my car serviced (regardless of whether I know anything about Deborah directly).

In addition, as mentioned above, Francone and Castelfranchi (2010) argue that this kind of propagation – which Francone and Castelfranchi (2010) call transitivity1212 – is only reasonable under especially restrictive conditions more restrictive than those discussed in Jøsang et al. (2006). In particular, in our terminology, Francone and Castelfranchi (2010) say that if A trusts B in some context, and B trusts C in the same context, and A believes that B trusts C in that context, then A trusts C. The context in which A trusts B and that in which B trusts C are allowed to differ so long as they are close enough. If I trust Bob on the subject of restaurants and Bob trusts Camilla on the subject of bars, I might decide that these two areas of expertise are close enough that I am willing to trust Camilla on the subject of bars. We will revisit the role of context in our discussion of critical questions.

Reciprocity is suggested as a method of trust inference in Guha et al. (2004) under the name ‘transpose trust’, though it is not found to be a very strong factor in the empirical work in Guha et al. (2004). This is not really surprising if we consider that trust is all about being able to predict the future actions of other agents. An agent that acts somewhat unreliably (thus making it difficult for others to develop trust in it) may have no difficulty learning to trust another agent that acts much more reliably.1313 Context, of course, adds to the reasons that trust does not have to be reciprocal. While I have been building up my trust in Bob's ability to recommend restaurants, I might not have given him any reason to trust me on the same subject (he might think all my recommendations are lousy, I may not have given him any, and in an extreme case – for example, in the restaurant context, the case in which I read his food blog and never comment on it – Bob may not know who I am). Also, if I am asking for recommendations in the first place, then chances are I do not know myself – which is another reason why Bob may not trust me on this topic. Here, context also plays a role. I might ask Bob about restaurants in a city that I have never visited but one that Bob knows well, which would only cause Bob to doubt my knowledge of restaurants in that particular city. Indeed, my query may make him think that I am a foodie, because I care enough to task for a restaurant recommendation to begin with; which may make Bob trust me to recommend restaurants to him in cities that I know well.

Co-citation is another form of trust inference from Guha et al. (2004), and Guha et al. (2004) assemble empirical evidence that this form of reasoning agrees rather well with people's intuition about trust. (To be more precise, Guha et al., 2004, took numerical information about the trust between A and B, C and B, and C and D, and from this computed a numerical estimate of the trust A has in D. This was found to provide a reasonable match for the trust that experimental subjects said A should have in D.)

This pattern of inference suggests that if A and C both trust the same individual, B, then A may feel that C is a reliable source of information about whom to trust (since they agree on B), and so accepts C’s implicit recommendation about D. In terms of the restaurant recommendation example, if I trust Bob to make good restaurant recommendations, and I know that Camilla also trusts Bob, then I might start to think Camilla knows some good sources of restaurant information, so that when she raves about Dave's ability to pick good places to eat, I might be inclined to listen to Dave.

The idea of co-implication, which is a novel scheme that we are proposing here, came about by considering how trust values are computed in co-citation, that is, by considering the way that they appear to be computed in Guha et al. (2004). In co-citation, the propagation of values goes from A to B to C to D, with, as described above, the fact that A and C share a common opinion of B leading to A taking C’s approval of D as a recommendation of D. Co-implication takes one less step, ending not with the inference of A’s trust in D, but rather the explicit inference that A trusts recommendations made by C.

As an example of co-implication, consider that I learn that both Camilla and I trust Bob's views on restaurants, and so I start to believe that Camilla is a good source of referrals because she trusts the same people that I do.

4.2.Discussion

All the propagation schemes listed above, except co-implication, which, to the best of our knowledge is stated here for the first time, can be found described in the trust literature. However, elsewhere these schemes are not discussed as argument schemes but in terms of mechanisms for computing numerical trust values for the relations indicated in Figure 2 by dotted lines. Of course, this does not mean that they are universally accepted. Even direct propagation, which is the most widely discussed mechanism, is understood not to be universally applicable. Rather it only applies in a rather restrictive context which we capture by the use of critical questions (see Section 5).

The difference between these argument schemes and the ones in Section 2 is that the latter were all schemes to identify when trust can be established between agents in the absence of any information about trust, whereas the propagation schemes are about how to combine existing information about trust relations between agents. In other words, considering trust graphs like those in Figure 2, the argument schemes of Section 2 tell us what solid links exist – which of these basic relationships can be established – while the propagation schemes generate additional connections built on top of the basic relations, providing us with the ability to infer new links.

In addition, when we deal with propagation of trust, we have to be careful with the context of the trust. As we have already mentioned, via Jøsang et al. (2006), propagation involves a combination of trust in an individual with regard to some behaviour – functional trust in that individual – and the trust in an individual to make trustworthy recommendations – referral trust. In direct propagation and co-citation, we infer functional trust. In co-implication and trust coupling we infer referral trust. Figure 2 represents functional and referral trust differently to clarify what is being inferred in each case. Figure 2(a) also clarifies why we do not use the term ‘transitivity’ for direct propagation: since the relation between A and B is different than that between B and C, the inference of a relation between A and C is a different process than that permitted by the usual mathematical notion of transitivity.

It is worth noting that the form of inference at the heart of co-citation, co-implication, and trust coupling is analogous to intercausal reasoning in probabilistic networks (Koller & Friedman, 2009). Intercausal reasoning is perhaps best explained using an example like that in Figure 3(b). If A and C are both causes of B, then a priori if either A or C becomes more likely, then so does B. However, if B is known to be the case, then any increase in the probability of A or C cannot increase the probability of B (since that is fixed at 1). Instead, if the probability of A, for example, increases, the probability of C decreases. Thus, to take the classic example (Pearl, 1988), both rain (A) and the operation of the sprinkler (C) make it more likely that the grass is wet (B). However, if I know the grass is wet, then believing that the sprinkler was on makes it less likely that it was rain that made the grass wet.

Figure 3.

Probabilistic networks in which intercausal reasoning may take place.

As described, the relationship between A and C (in Figure 3) is the form of intercausal reasoning known as explaining away. Evidence that makes one cause of B more likely provides an explanation for B and simultaneously makes the other cause of B less likely. This is arguably the most common form of intercausal reasoning, but forms in which evidence for one cause makes the other more likely are also possible (Druzdzel & Henrion, 1993; Wellman & Henrion, 1991).

If the trust networks in Figure 2(c)–(e) are considered as probabilistic networks, then the corresponding forms of inference make sense in probabilistic terms. In co-implication, Figures 2(d) and 3(b), A’s trust in B and C’s trust in B together increase A’s trust in C. In co-citation, Figures 2(c) and 3(a), the same reason as in co-implication is followed by propagation from C to D. Finally, in trust coupling, Figures 2(e) and 3(c), propagation from A to B precedes the intercausal reasoning of co-implication.

7.1.Related work on trust

As discussed in Section 1, varying definitions of trust have been offered in the literature. We are concerned with the issue of reasoning about trust from a specifically computational perspective. It has long been recognised in the multi-agent systems literature, for example, Castelfranchi and Falcone (2000) and Ramchurn, Huynh, and Jennings (2004), that the autonomy afforded to intelligent agents brings with it uncertainty regarding whom can be trusted or not within multi-agent interactions. There are many different ways in which trust can be classified. For example, trust can be directed towards individuals (such as an interaction partner) or an overall system (such as an auction governed by rules) (Erriquez, 2012; Ramchurn et al., 2004). Trust information may derive from direct interaction with others, indirect interactions, e.g. from reputation information available, or from some third party authority (Huynh, Jennings, & Shadbolt, 2006). As can be seen in the schemes and critical questions we have presented in this paper, we cover these different facets of trust.

Other distinctions such as basic trust (modelling a general disposition based on an agent's experiences), general trust (directed at a particular agent but without considering a particular situation), and situational trust (directed towards a particular agent in a particular situation) have also been set out in the literature (Marsh, 1994). Indeed, Marsh (1994) served as one of the first and most prominent accounts of computational modelling of trust and, as such, provides a detailed formal model of trust. In contrast, our schemes are intended to provide patterns of reasoning that can easily be applied in multiple situations and are broad enough to cover the different scenarios in which reasoning about trust is of importance; though as we showed in Section 2, we can account for some of these distinctions in our general characterisation of our schemes.

More recent work has made the link between trust and argument-based representations. Work by several of the authors has suggested that argumentation, with its ability to capture the reasons for its conclusions and the data from which those conclusions were inferred, is a natural formalism for reasoning about trust (Parsons et al., 2010). Further work in this line (Parsons et al., 2011; Tang, Cai, McBurney, Sklar, & Parsons, 2012) has provided logic-based representations to reason about trust, and it is shown how this can be combined with reasoning about beliefs. Although the formal systems presented in Parsons et al. (2011) and Tang et al. (2012) are incomplete, they make progress towards showing how trust can be represented as a form of argumentation and the work has led to a prototype reasoning system (Tang, Cai, Sklar, & Parsons, 2011) which has continued to evolve (Parsons, Sklar, Salvitt, Wall, & Li, 2013).

There is a growing interest in the use of argumentation to handle trust beyond that described above. For example, Villata et al. (2011) use meta-level argumentation to capture the trust an agent has in an information source, while Stranders et al. (2008) use a fuzzy approach, and Matt, Morge, and Toni (2010) suggest how arguments may be combined with statistical data to augment existing trust models. The scheme-based representation we have presented in this paper tackles the issue of trust from a more flexible perspective than any of the previous approaches to trust argumentation that we have mentioned here, since the definition of a set of schemes captures different nuances of the notion of trust that are conflated in other work, teasing out different aspects of trust rather than treating trust as a monolithic concept.

Two further approaches take a more abstract view. In Erriquez (2012) and Erriquez, van der Hoek, and Wooldridge (2011), a framework is set out that takes inspiration from Dung's abstract argumentation frameworks (Dung, 1995) whereby a graph theoretic model is used to capture the distrust relations within a society, and this model is used to formulate notions of mutually trusting coalitions. In Dung's framework, nodes in a graph represent arguments, but in Erriquez (2012) and Erriquez et al. (2011) they denote agents, and in Dung's framework edges denote a binary attack relation, whereas in Erriquez (2012) and Erriquez et al. (2011) they represent a distrust relation between agents that is used in determining which coalitions are acceptable. Similar in scope is Harwood, Clark, and Jacob (2010) which uses argumentation to decide on trustworthiness, building links between agents, labelling them with ‘trust’ or ‘distrust’, and applying something like Dung-style semantics to decide whom to trust. While these accounts of trust are inspired by well-understood representations from computational argumentation, they remain at the abstract level, without delving as deeply into the specifics of arguments about trust as is done in the scheme-based approach that we have presented.

7.2.Related work on argumentation schemes

We now turn to discuss the literature related to our chosen form of representation, argumentation schemes. Argument schemes were born out of the literature on informal logic, and the most influential work in this area, at least as far as work on computational argument is concerned, is the work of Walton et al. (2008), which we have cited extensively above. This book provides a solid introduction to the use of argument schemes and catalogues a large number of them. We consider these schemes to be very general in that they are not fitted to a specific domain. As a result, they have wide applicability, but do not capture all the subtleties or peculiarities of specific domains. As a result, as we have pointed out above, some of the schemes and associated critical questions in this paper can be viewed as sub-classes of the schemes and critical questions from Walton et al. (2008). We envisage that our schemes could be used in conjunction with the existing more general ones as a way to probe into issues of trust that may arise as part of the critical questioning.

There are also a number of works which, like this paper, identify argument schemes for specific domains. For example, Bench-Capon and Prakken (2010) and Wyner and Bench-Capon (2007) consider argument schemes for legal reasoning, Reed and Walton (2005) discuss argument schemes for agent communication, Ouerdane, Maudet, and Tsoukias (2008) consider argument schemes for decision support, and Tolchinsky, Modgil, Cortes, and Sanchez-Marre (2006) look at argument schemes for deliberation in the sense of Walton and Krabbe (1995), that is, the process by which several entities reach a combined plan for action. In this same line of work, Prakken (2005) presents the case for using argument schemes as an alternative to logic as a means of knowledge representation (again focusing on the legal domain).

More recently, researchers have become interested in transforming argument schemes into computational versions to enable them to be used in systems that perform automated reasoning. For example, some of the authors of this paper developed a scheme for practical reasoning (Atkinson et al., 2006) that has been widely used to facilitate reasoning about what to do in a number of different problem scenarios. Schemes concerning witness testimony and expert opinion have also been used in computational argument (see, for example, Gordon, Prakken, & Walton, 2007), and there have been several attempts to capture argument schemes from legal reasoning in various forms of logic (Bex, Prakken, Reed, & Walton, 2003; Prakken, 2010; Verheij, 2003). The reason that such schemes have attracted the attention of researchers in computational argumentation is that their defeasible nature makes them naturally attractive to the non-deductive form of reasoning that many people are interested in capturing in argument-based approaches. Transferring the natural language schemes into a computational account poses challenges that cover a number of considerations.

Firstly, the adequacy of any representation concerning the different elements that the schemes comprise (such as facts, subjective judgements, causal theories, and so on) needs to be demonstrated. This point has been recognised in work such as Wyner, Atkinson, and Bench-Capon (2012) where a functional language for a computational analysis of schemes is set out. Further, it is shown in Wyner et al. (2012) how argumentation schemes expressed in the functional language given can be systematically related to one another, which is an issue that arises when it is desirable to use multiple schemes in any one setting.

The issue of representing the interconnections between schemes has also been tackled through specifications for a World Wide Argument Web (Rahwan, Zablith, & Reed, 2007) that uses Walton's theory of argumentation schemes (Walton et al., 2008). As such, an ontology is presented to enable the representation of networks of arguments on the Semantic Web.

Another consideration when making the schemes computational is the treatment of critical questions. As noted earlier, according to Walton et al. (2008), the role of critical questions is how the defeasibility inherent in argumentation schemes is recognised. The critical questions are intended to be employed as part of a dialogue whereby the answers to the critical questions can provide a reason for rejecting the conclusion of the scheme. In the computational account of the practical reasoning argumentation that is given in Atkinson et al. (2006), the questions are used to identify attacks that can be made on an argument derived from instantiating the scheme whereby the attack arises due to an assertion made in answering a critical question. For example, consider a critical question such as ‘Does the action have the stated consequences?’. When this question can be answered in the negative and an assertion can be made that the action has consequences different to those stated in the instantiation of the scheme, then this can constitute an attack on the argument instantiating the scheme.