Trust, relevance, and arguments

1.Introduction

In a recent paper (Paglieri & Castelfranchi 2012), we noted that trust and relevance, while being separately the focus of much research in philosophy (Baier, 1995; Cohen, 1994; Floridi, 2008; Gambetta, 1988; O'Neill, 2002), pragmatics (Borg, 2005; Grice, 1989; Sperber & Wilson, 1995; Wilson & Sperber, 2004), and computer science (Borlund, 2003; Bremer & Cohnitz, 2004; Castelfranchi & Falcone, 2010; Cooper, 1971; Crestani, Lalmas, Van Rijsbergen, & Campbell, 1998; Ramchurn, Huynh, & Jennings, 2004), have rarely been brought into contact with the study of information dynamics. Yet, we argued in that paper, their integration has much to offer. The same holds true for argumentation: relevance has often been studied in that context (for a comprehensive survey and treatment of the subject, see Walton, 2004), and trust is also often discussed, especially in relation to the evaluation of testimony (on that topic, see in particular Govier, 1997, 1998). However, the need for trusting (or not) the relevance of arguments has not been addressed, nor the rationality (or lack thereof) we manifest in doing so – with very few exceptions, such as Santibañez (2012). There seem to be a widespread tacit consensus that relevance is not a matter of trust, as far as arguments are concerned.

In this paper, we challenge that assumption, to show that, on the contrary, there is an inescapable element of trust built into our judgements of relevance with respect to arguments, and that such trust can have, and often has, rational foundations. This, in turn, has important consequences for the formal and computational treatment of argumentation, as well as some potential technological implications. The paper is organised as follows: in Section 2, we briefly discuss how relevance is to be understood in argumentation, and propose a distinction between two main types of argumentative relevance; then in Section 3 we ask whether these forms of relevance require trust, and of what kind, whereas Section 4 is devoted to understanding what reasons may justify our trust in argumentative relevance; Section 5 outlines the implications of this analysis for formal and computational models of argumentation, as well as potential technological applications; Section 6 concludes the paper, by summarising its main results and pointing to open-ended issues for future work. While throughout the paper we will often engage with the rich literature on trust and relevance, it is not our purpose to offer a comprehensive survey of either topic, and this paper should not be taken as such.

2.Two forms of relevance in argumentation

It is notoriously difficult to arrive at a comprehensive and satisfactory definition of relevance in information dynamics (on that problem, see Bremer & Cohnitz, 2004; Floridi, 2008): elsewhere (Paglieri & Castelfranchi, 2012) we have argued that such a definition will require making reference to the goals of the parties engaged in information exchange, and we also insisted that the connection between information and goals is made explicit by a belief-based model of goal processing, like the one detailed by Castelfranchi and Paglieri (2007). With respect to argumentation, the role of goals in defining relevance has long been recognised, thus there is no need to elaborate on it any further here. An early mention is found in the following passage by Sidgwick: ‘The more we understand that all thought is purposive the more we shall be able to discover the less obvious kinds of irrelevance in argument’ (1914, p. 149). More recently, Walton (2004) has explicitly tied the definition of argumentative relevance to the dialogical goals of the interaction in which the parties are engaged: in turn, such dialogical goals contribute to determine the type of dialogue, and thus the appropriateness of the arguers’ moves in it (as discussed in Walton, 1998). For instance, a deliberation dialogue aims at determining a course of action deemed acceptable by both parties, whereas in eristic confrontation (e.g. a quarrel) the goal is to vent one's feelings, with little concern for any resolution in the underlying conflict of opinions. Due to this sharp difference in dialogical goals, what is relevant in one context may well not be so in the other.

It has also been often noted that relevance, in argumentation as well as in other forms of communication, is not a matter of topical overlapping (see, for instance, Walton, 2004, p. 10). This is easy to see. Imagine we are debating whether or not the moon has an impact on tides: then my claim that ‘Looking at a full moon over the sea at high tide is lovely’ is clearly irrelevant to the discussion, despite the significant thematic overlap. This has prompted some authors to elaborate a more specific notion of relevance. For instance, Walton proposes the following as a working definition of what he calls material relevance (as opposed to relevance in general):

For the purpose of this paper, it is useful to distinguish two types of relevance that are important in argumentation: external relevance (or argument relevance) and internal relevance (or premise relevance). Let us first offer a working definition for these notions, and then see how they fit in the current literature on relevance in argumentation.

While external relevance concerns the relationship between an argument and the dialogical context in which it is used, internal relevance refers to the proposed connection between the premises and the conclusion of an argument – hence the label ‘internal’, to emphasise that it is a property of the internal structure of an argument. Are the premises truly relevant to the conclusion they are supposed to support? This is the sense in which relevance is considered a key criterion for assessing the cogency of an argument (Govier, 1985; Vorobey, 2006), and the meaning of Blair's (1992) notion of premissary relevance. It is also the notion of relevance implied by the characterisation of ignoratio elenchi given by Aristotle (1939) in his Topics:

The rationale of the distinction between internal and external relevance, if not the precise labels, is not novel in argumentation theory, and it even precedes the article by Jacobs and Jackson (1992) mentioned above. Similar categories can be found in the treatment of fallacies of relevance in logic textbooks from the first half of the twentieth century, such as Castell (1935) and Werkmeister (1948): these authors oppose the fallacy of irrelevant thesis (lack of external evidence) to the fallacy of irrelevant evidence (lack of internal relevance). Also, the distinction between red herring and the fallacy of irrelevant reason in Johnson and Blair (1983) closely mirrors the difference between external and internal relevance. Even more significantly, the same distinction is found, under different names, in the Federal Rules of Evidence (2011) of the USA. In particular, rule 401, ‘Test of Relevant Evidence’, reads as follows:

Interestingly, a slight paraphrase of the rule is sufficient to reveal its connection with the distinction between (a) internal and (b) external relevance:

Some authors have focused more on one or the other aspect of argumentative relevance. Govier's analysis of relevance (1985) in terms of probative weight is essentially concerned with internal relevance – in fact, it closely resembles clause (a) above, since it is a type of probative relevance (Walton, 1992; Woods, 2013). In contrast, the various relevance criticisms discussed by Krabbe (1992) mostly pertain to external relevance, and also Walton (2004) is definitely more focused on this type of relevance, when he repeatedly stresses the importance of the dialogical context in determining the relevance of an argument. As he notes:

As for fallacies of relevance, in argumentation theory their treatment spans the divide between internal and external relevance, although it is not always easy to see on which side of that divide each individual fallacy lies. This largely depends on the fact that even the list of those fallacies is somewhat controversial (on this point, see Walton, 2004; Woods, 2013). The standard quartet includes all the ‘ad fallacies’: ad hominem (personal attack), ad populum (appeal to popular opinion), ad misericordiam (appeal to pity), and ad baculum (use of threats). But the straw man, i.e. attacking a point that unfavourably misrepresents the counterpart's position, and the ignoratio elenchi or red herring are also often cited as fallacies of relevance, the latter being often used as a ragbag category, as Hamblin put it (1970, p. 31): whenever a fallacy cannot be classified otherwise, treat it as a red herring.11 While nowadays these forms of reasoning are no longer considered as invariably fallacious, depending on the context, they still tend to be grouped under the heading ‘fallacies of relevance’ – so much so as to inspire Krabbe's (1992) metaphor of ‘the sponge of relevance’, to indicate the widespread habit of absorbing within this notion a variety of different phenomena.

While we do not aim to settle the issue here, it seems appropriate to further classify the four ad fallacies as problems of internal relevance: for instance, whenever the personal characteristics of an individual have no bearing on establishing whether s/he committed a certain action (e.g. murder) or possesses a specific quality (e.g. being a reliable witness), then those characteristics cannot be used as premises for concluding anything on that action or quality, and doing otherwise would be a case of ad hominem fallacy. Similarly for appeals to popular opinion, emotions, or the use of threats: whether or not they are used fallaciously hinges on whether or not they constitute relevant premises for the conclusion they are supposed to prove – that is, whether or not they have internal relevance for that argument. In contrast, straw-manning seems more a failure of external relevance: it refers to the practice of making an unhelpful contribution to the ongoing debate, precisely because the position of the counterpart is being misconstrued. Something similar can be said about the red herring (ignoratio elenchi), at least based on Walton's definition of that fallacy:

3.Trust in relevance: does it matter for argumentation?

While the centrality of relevance for the study of argumentation is well established, not much work has been devoted to discuss whether trust plays any significant role in assessing the relevance of arguments. In fact, the only exception we are aware of is a recent paper by Santibañez (2012), who proposes an evolutionary connection between relevance (in the sense of Sperber & Wilson, 1995) and trust, largely inspired by Tomasello's (2008) views on the evolution of language. Simply put, the idea is that the cooperative expectations on which human communication is grounded, including a presumption of relevance, could not have evolved in the absence of a certain degree of trust among members of the same community. While we do not find this evolutionary hypothesis implausible (on the contrary, it is consistent with the idea of a reciprocal altruism of knowledge proposed by Castelfranchi & Poggi, 1993), we believe the connection between trust and relevance to run deeper than that: in particular, trust in relevance is an essential ingredient of everyday communication right now, whether or not it played a role in the evolution of language. Appreciating that, however, requires an understanding of trust different from the one favoured by Santibañez, when he writes:

In fact, trust is based on reasons even when it does not seem so at first glance. Take the common occurrence of asking for directions from a complete stranger in the streets of a city you are visiting: you have no hard evidence of the competence, character, or motives of your informant, and you are unlikely to ever meet him/her again to sanction any misdirection s/he may offer – yet, you will often trust in his/her indications. Is that not an act of blind trust? Not at all: at the very least, your trust is based on the reasonable expectation that, lacking any motive to cause you harm or be less than cooperative, the informant in question (having already decided to invest some resources in his/her communicative act) will try to provide good information. This is demonstrated by the fact that your choice is not random: for instance, you will probably avoid bothering someone who is in an obvious hurry, or engaged in some other activity – that is, you will try to pick someone who is likely to be well disposed to answer your query. As for his/her competence, this is also factored in your choice, within the limits afforded by your current predicament: again, you will not select someone who is obviously a fellow tourist (e.g. someone taking pictures of the monuments), but rather someone who seem to be familiar with the surroundings (e.g. someone returning from a grocery shop). All of these are subtle indications of the reason-based nature of your act of trust, even when information on the object of that trust is very scant (another notable example being our frequent trust in Wikipedia entries, well analysed in Goodwin (2010)).

Turning back our attention to relevance in argumentation, it is worth asking whether trust, in the sense outlined above, is required in dealing with external relevance. In assessing whether an argument contributes or not to the matter under discussion, do we need to trust in the proponent of that argument? The answer to this question is in the negative, at least typically. In argumentation, external relevance is either self-evident, openly criticised, or mistakenly assumed, but in none of these cases does trust play a major role. Let us start from the most common situation, when external relevance is self-evident. The reasons behind this fact have been already hinted at: while arguing on a matter of joint interest, all interlocutors tend to know what they are talking about – not in the sense of being necessarily expert on the subject matter, but in the sense of having a clear and often complete view of what transpired so far and thus what are the public (manifest) positions of each party. This makes it relatively easy to directly assess whether or not a new argument is relevant to the matter under discussion, and immediately prompt criticism or demands for clarification if it is not obviously so. Thus, there is not much need for trust in relevance, nor an inclination to grant it, since the whole point of arguing about a certain topic is precisely to subject each position to careful scrutiny, instead of accepting entire arguments based on trust.

Notice that the point we are making is not that external relevance is uncontroversial in discussion (on the contrary, it can become the subject of heated debate), just that it is not typically a matter of trust. If the external relevance of an argument is in doubt, it will be challenged, and the challenge will become part of the discussion. This is the second instance we mentioned above. For instance, let us imagine, following Walton (2004, pp. 25–26), that it has been decided, at the early stage of a committee meeting, that the task at hand is to deliberate whether to take action A or B, and then at some point during the discussion a committee member questions this earlier decision, suggesting that the committee should also consider the possibility of refraining from either action for the time being. Is this motion to be ruled as irrelevant, or not? As Walton notes, this will depend on the circumstances of the case, but that is not our point. Whenever external relevance of arguments comes to be debated in dialogue, it still does not involve trust. The statement ‘Trust me, this is relevant’ is not really a suitable answer to a doubt of relevance: at most, it can be accepted provisionally, as a promise that the relevance of that particular argument will soon become apparent. This sometimes occurs in legal proceedings, whenever a judge, following an objection of relevance, allows one of the parties to proceed, with the admonition that they will have to ‘get to the point’ in short order. While similar cases underscore the provisional and procedural nature of relevance judgements (Walton, 2004), they do not constitute genuine instances of trust: in fact, the alleged relevance of the argument for the case is not accepted until it is openly demonstrated, and only a temporary suspension of judgement is granted.

Finally, it is undeniable that some arguments appear to be externally relevant and are accepted as such, even if they are not. Imagine, for instance, that the ethnic majority in a country is prompted by a particularly moving appeal to indignation to take action against a minority, even if there are no grounds to assume that indignation to be justified, and thus appealing to it should be irrelevant in deciding how to act towards that minority. This is a perfectly possible occurrence: after all, we have ample evidence of rhetorically successful fallacies of relevance. But once again their success does not involve trust: in similar cases, people are swayed by their own cognitive and emotional make up, rather than by any overarching trust in the speaker. They are being played, perhaps expertly, but the key to playing them is not trust.33

To sum up, external relevance does not typically involve trust. However, things are different with respect to internal relevance, that is, the relevance of premises for the conclusion they are intended to support. The mere fact of presenting something as a reason for something else presupposes, among other things, a claim on its relevance: when someone argues that ‘p, therefore q’, the (internal) relevance of p to q is assumed – in fact, it is asserted, in the sense that the proponent commits to it. However, quite often such an assumption cannot be easily checked by the counterpart, thus making its acceptance (or not) a matter of trust, or lack thereof. Consider the following example:

Yet we speculate that most arguers, lacking previous knowledge on the exact species of birds to which Tweety belongs, would be ready to accept (1) as a good presumptive argument – that is, they would be willing to trust in the relevance of its premises for the proposed conclusion. Moreover, we argue that they would usually be justified in doing so: that is, the fact that our trust in premise relevance can turn out to be misplaced does not imply that granting such trust is always a reasoning error. Whether or not the arguer is guilty of ‘hasty trust’ depends on several considerations, which we will discuss in the next section. For the moment, the important point is to have established that trust in internal relevance indeed occurs within argumentative exchanges, whether or not relying on it is normatively correct.

Incidentally, the fact that the relevance (or lack thereof) of a premise depends on other premises that are considered in conjunction with it has already been noted in the literature, for instance, by Johnson and Blair: ‘Relevance (unlike truth) does not inhere in each individual premise but rather in an individual premise taken together with certain other items of evidence or information’ (1983, p. 39).

Another interesting example of how trust in relevance matters for argumentation is given by the following enthymeme (taken from Paglieri & Woods, 2011a):

To sum up, trust in internal relevance can affect both argument acceptance, as in (1), and argument reconstruction, as in (2). Moreover, we suggest that this is not a rare occurrence in everyday argumentation. This is obviously an empirical claim, for which appropriate corroboration will have to be sought – something that is not within the scope of the present contribution. However, as prima facie evidence of the relative frequency of trust in premise relevance, consider the following: a need for trust in relevance is created whenever the link between premises and conclusion is neither self-evident nor already known to the interpreter. But arguments where that link is self-evident or already known are typically of little practical value, and thus rarely used, precisely because their inferential contribution is so obvious. Take the mother of all enthymemes, ‘Socrates is a man, therefore Socrates is mortal’: while this is certainly a fine illustration of a truncated syllogism, it is unlikely the argument will ever be used in a real debate, since everyone is already well-acquainted with the relationship between being a man and being mortal, as well as with the fact that Socrates is indeed a man. On the contrary, there is practical interest in using arguments in which the link between premises and conclusion is neither self-evident nor already known to the counterpart, since these are both more informative (if the situation is fully cooperative) and more useful in presenting the opponent with novel challenges (if the context is adversarial). Hence, we speculate that interpreting and assessing arguments in everyday dialogues will often involve trust or distrust in the relevance of their premise for the conclusion.

It is also worth noting that, in interpreting (1), we trust its premises to be not only relevant, but also maximally so – roughly speaking, we trust that they offer all the relevant evidences needed to support as strongly as possible the conclusion, to wit, whether Tweety can fly or not.44 Hence (1) is interpreted by assuming, among other things, that the arguer does not know whether Tweety belongs to a non-flying species of bird or not. Indeed, we would feel cheated, if we were to find out later on that the arguer knew Tweety to be an eagle all along, since then it would have been maximally relevant to argue along the following lines:

Before proceeding any further, it is important to address a potential objection: how is the proposed analysis of trust in relevance any improvement on other, well-known ways of dealing with cooperative dynamics in meaning construction? After all, readers familiar with pragmatics and philosophy of language might be experiencing a strong déjà vu at this point: disambiguating (1) and (3) requires little more than a straightforward application of Grice's maxim of relevance, whereas enthymeme reconstruction in (2) seems to be guided by a relevance-based version of the principle of charity (for critical discussion of the role of charity in enthymeme interpretation, see Lewinski, 2012; Paglieri & Woods, 2011a, 2011b). Both are old friends of any student of argumentation. Not to mention the fact that ‘trust in relevance’ seems oddly interchangeable with ‘presumption of relevance’, and presumptions in turn have a long scholarly tradition in argumentation theory (Godden & Walton, 2007; Kauffeld, 1998, 2003; Macagno & Damele, 2013; Walton, 1996, 2014a). So why should we embark in this complex analysis of trust in relevance, when more familiar analytical tools are available?

We see at least three benefits in this theoretical move. First, trust is a key ingredient in any type of social interaction, whereas presumptions tend to be discussed in relation to specific phenomena (e.g. language use) or contexts of application (e.g. legal presumptions). Even accepting a certain overlapping between presumptions of relevance and trust in relevance, it would still be useful to clarify that accepting the relevance of an argument requires an attitude of trust, as so many other forms of social interchange do. While language is certainly a unique social phenomenon, its uniqueness does not depend on trust: what linguists and argumentation scholars are used to label presumptions are just acts and attitudes of trust, rather than some specific feature of language alone.

Second, analysing judgements of relevance in terms of trust emphasises the fact that they are not the undiscriminated output of some automatic process, as we shall discuss more thoroughly in the next section. While trust in relevance is often granted by default, this does not mean trusting without reason or justification. It suggests instead that reasons and justifications are needed to suspend the default of trust, rather than the opposite. Incidentally, the whole Gricean view of language is amenable to be reinterpreted in terms of default heuristics, both on the side of the speaker (‘Unless you have specific incentives to deceive or misinform, provide relevant and true information’) and on the side of the interpreter (‘In the absence of specific reasons for diffidence and doubt, assume the message to be true and relevant’). As discussed elsewhere (Paglieri & Castelfranchi, 2010b; Paglieri & Woods, 2011a), such default trust attitudes are parsimonious without making us unduly vulnerable to deception.

This relates to the last and most decisive reason for favouring an analysis of relevance in terms of trust, as opposed to a more standard presumptive account. Trust is a stronger notion than presumption; or, to put it differently, trust is a highly specialised form of presumption, so the latter, broader notion cannot accurately describe the former. Trust, before becoming an act and a relationship, is an attitude towards the other, an attitude made of presumptions (or better, evaluations) on the qualities of someone or something else, and of presumptions (or better, expectations) about her/his behaviour. But a true expectation, as well as the trust than can be based on it, is more than a mere presumption, since it requires a positive conviction regarding a future state of affairs; a possibility is not enough (it might be true/relevant), we need at least a positive probability (it is likely to be true/relevant). Such probability has to be high enough for us to decide to trust as an action, that is, to make ourselves vulnerable, to take some risk. This is exactly the type of risk we expose ourselves to whenever we take at face value the qualities of an argument, including its relevance.

In argumentation theory, attempts to analyse the notion of presumption tend to reveal this natural link with the stronger notion of trust, albeit the label itself is rarely, if ever, used. For instance, Kauffeld (2003) describes ordinary presumptions of truthfulness as based on supposing the speaker to be unwilling to risk criticism, resentment, retribution, etc., for failing to be sincere and reasonably knowledgeable on what s/he is talking about. Albeit Kauffeld, to the best of our knowledge, does not discuss presumptions of relevance, similar considerations would easily apply. But seen in this light, presumption is just a technical term to talk about trust:55 in particular, Kauffeld is suggesting that we trust arguers to be motivated to be truthful (and relevant, we hasten to add) by fear of negative repercussions, given current social conventions. If one is to accept this approach, then the notion of trust appears well suited to convey the high stakes involved even in the most basic act of reliance on others as veridical and relevant arguers. In other words, what grounds our linguistic presumptions is the trust we have reason to put in our fellow arguers, either due to general considerations (e.g. social conventions and principles of parsimony) or to specific motives (e.g. past experience with those arguers, knowledge of their incentives and expertise).

4.When is it OK to trust in premise relevance?

Even granting that we frequently rely on trust regarding the internal relevance of an argument, are we ever justified in doing so? Under what circumstances? And what kind of justification is involved – epistemic, pragmatic, or both? Before even beginning to answer these questions, one needs to distinguish between social expectations and rational justifications, since both are involved in trusting relevance. While these two dimensions are intertwined, they are not to be collapsed or confused. In fact, epistemic vigilance (in the sense of Sperber et al., 2010) and social competence are often at odds with each other: even when challenging or subjecting to critical scrutiny a certain argument is rationally justified, it could also be considered as socially inappropriate – and rightly so, depending on circumstances. Imagine a social dinner, with people engaged in a deliberation dialogue (Walton, 1998) on which wine to order, and then John puts forward the following argument from expert opinion: ‘I know my wine, and this 2006 Barolo is sure to be good. We should go for it!’. To which, according to the standard treatment of this argument scheme, it would be perfectly rational for Lucy to reply: ‘Do you really know what you are talking about?’ Obviously, the most likely rejoinder to that critical question would be an angry look, not a well-argued reassurance on John's credentials as a wine expert, based on the fact that Lucy's comment would be taken as aggressive – which it is, given current social norms on conversation.

The independence between rational justification and social competence works both ways, of course: the fact that Lucy's comment breaches social conventions does not make it any more (or less) justified on rational grounds – after all, John may very well be utterly ignorant on wines, in spite of his boastful remarks. But this is exactly our point: rational justification and social competence are not one and the same.

This has one important implication for establishing the rational grounds for judgements of relevance: while considering under what conditions trust in premise relevance is warranted, we must acknowledge two aspects to this question. First, whether trusting in relevance is socially justified. Second, whether doing so is based on sound reasons, be they epistemic or practical. Let us recall the argument in (2), ‘All butterflies are short-lived, therefore the Psittacula krameri is short-lived’. Imagine that a hypothetical receiver trusts in the relevance of its premises, thereby accepting something like ‘The Psittacula krameri is a butterfly’, and let us stipulate that doing so is rationally ill-advised in this situation – for instance, because there were reasons to suspect the speaker to be dishonest, incompetent, or both. Still, trust in relevance remains socially legitimate, insofar as it is demanded by the mere assertion of (2), as part of its pragmatic meaning: that is, understanding (2) includes assuming its stated premise to be relevant, hence whoever asserts (2) also puts forward such a relevance claim. Granted, taking at face value the argument as relevant may very well be a sign of epistemic laziness. But this does not make the proponent of the argument any less guilty of abusing the trust of the counterpart (or attempting to).

A simple analogy will make the point even clearer: imagine that a young girl decides to walk at night in a notoriously dangerous district of her hometown, and as a result ends up being gang-raped. It would be factually correct, albeit very cruel, to say that she brought such a tragedy upon herself by failing to exert even the obvious minimum of caution. In fact, if the girl somehow had avoided being raped and her parents had learned of her escapade, they would have severely scolded her for such a basic lack of judgement, and rightly so. But none of this makes gang rape, attempted or perpetrated, any less a crime, precisely because one is socially entitled to personal safety, regardless of any prudential consideration. The same applies to trust in relevance, and more generally to the role of trust in argumentation: whether or not we have good reasons for trusting in internal relevance, the mere fact that something is proposed as a premise for a conclusion authorises that trust, as far as basic principles of cooperation are concerned.

Again, this works both ways: precisely because the very functioning of communication requires a basic level of cooperation, a speaker is entitled to expect, and even demand, a certain degree of trust from the counterpart, unless there is previous evidence to the contrary.66 This is why, in the dinner example, John would be entitled to resent Lucy's scepticism on his knowledge of wines, or, similarly, we often react poorly to any sign of distrust from a friend or a spouse, no matter how acceptable their doubts may be on rational grounds. ‘You just have to trust me!’ is a frequent refrain in conversation, one that usually is not used to deceive or avoid scrutiny, but rather to stress the kind of social relationship that has to be maintained throughout a civil debate.

It is easy to lose sight of this balancing act between rational questioning and social demands, especially in argumentation theory, where there is an undeniable tendency to favour purely logical and dialectical factors, as opposed to a broader attention to social conventions (on this point, see Gilbert, 1997; Santibañez, 2012). As a case in point, with respect to relevance, Walton proposes that ‘the presumption is that [an argument] is irrelevant unless its proponent can show otherwise’ (2004, p. 178). A moment of consideration will show that this principle, taken literally, is much too demanding to capture reasonable everyday interactions. Consider the following exchange, inspired by a real conversation between a father and his two-year-old son (indeed, a very typical interaction with children around that age):

None of this should be taken as suggesting that internal relevance is always to be trusted, with no questions asked. Even if social conventions place on us a demand for trust in each other's utterances, and critical probing on relevance (as on anything else) has to stop somewhere, trust in internal relevance can also be based on further epistemic foundations, and often is.88 Conversely, each of these reasons for trust can also become a reason for distrust, if it appears clearly defective in the arguer and/or in the argument. Thus, the following also count as epistemic and pragmatic roots for trust in relevance:

To sum up, it seems that trust in internal relevance is not only a necessary ingredient for argumentation, but also a justifiable attitude, on a variety of grounds: compliance with social conventions (including basic principles of communication), the necessary finite nature of sceptical probing, several considerations regarding both the arguer, his/her proposed argument and the context of the dialogue, as well as a tendency towards parsimony in argumentation. Once again, this is not at all intended to suggest that we are always right, morally or rationally, in trusting the relevance of each other's arguments. However, it certainly vindicates the opportunity for further study of the role of trust in argumentative relevance, and suggests that such study need not be confined to descriptive models, but could also encompass normative concerns.

5.Formal models and computational applications for trust in argumentative relevance

As noted in the Introduction, much attention has been paid to trust and argumentation in computer science: more recently, their interaction has also started to be explored in greater detail (for an up to date survey, see Koster, 2014). The most natural way of integrating these two aspects of information dynamics is by looking at how trust in information sources ought to affect the evaluation of their arguments. Somewhat ironically, this would have been traditionally regarded as the basic ad hominem fallacy, since of course the quality of the source is not necessarily a reliable indication of the quality of its message. Similarly, even the presence of a large number of independent sources converging on the same opinion would once have been considered an ad populum fallacy, based on the fact that all those sources could still be wrong. However, as we discussed above, it has long been recognised that features of the source (including convergence with other independent sources) may nonetheless provide robust presumptive reasons to accept his/her claim, in the appropriate context: for a recent application of this approach to arguments from expert opinion, see, for instance, Walton (2014b). More generally, computer science has always found the connection between source and information to be a central one, especially in distributed artificial intelligence, e.g. in multi-agent systems: when computer agents have to interact autonomously with each other, it becomes essential to reliably assess information based on the reputation of its source, as well as to form well-grounded judgements of its trustworthiness in current circumstances.

Similar considerations apply also to the exchange of arguments with or among artificial agents, thus existing models could be extended to deal with relevance as well. Here, we will briefly discuss three approaches suitable for such an extension, only to hint at how it might be achieved: the modelling of relevance as a feature of arguers, alongside competence and sincerity (Paglieri et al., in press), the use of modal logic to capture trust in relevance (Amgoud & Demolombe, 2014), and the application of meta-argumentation techniques (Boella, Gabbay, van der Torre, & Villata, 2009; Villata, Boella, Gabbay, & van der Torre, 2013) to express claims of relevance, as well as challenges to them, within an abstract argumentation framework (Dung, 1995).

Paglieri et al. (in press) focus on a relatively unexplored aspect of information dynamics: how the quality of a message should impact on the subsequent assessment of its source (feedback dynamics). The issue was discussed theoretically by Falcone and Castelfranchi (2004), but formal models tend instead to concentrate on how the assessment of the source influences the evaluation of the message. In contrast, Paglieri et al. emphasise the bilateral relationship between sources and messages. Their work, however, deals with informational messages in general, and analyses only two features of sources: their competence (i.e. whether or not they are likely to have accurate knowledge of a given domain) and their sincerity (i.e. whether or not they are likely to provide honest information, as opposed to lying). Both dimensions are captured using bipolar pairs of graded beliefs (based on Dubois & Prade, 2008), to allow for ignorance on the source features: that is, an agent may be in the dark as to whether source S is competent (sincere) or incompetent (insincere). Their model details how to integrate these dimensions to generate an expectation on the quality of information provided by the source, and how to differentially attribute praise (or blame) to each feature when that information turns out to be better (or worse) than expected.

While the specifics of their approach are beside the point here, relevance appears to be conspicuously absent from the features to be considered in assessing a source – a fact the authors acknowledge at the end of the paper (Paglieri et al., in press). In contrast, we obviously have a right to expect our sources to be not only knowledgeable and honest, but also capable of providing relevant information (Paglieri & Castelfranchi, 2012). The same applies to arguments, both with respect to external relevance (a perfectly valid argument on the literary value of Gothic novels is not much use in deliberating where to spend a family vacation) and with reference to internal relevance (an argument with irrelevant premises is necessarily of poor quality, even if it might seem otherwise in the case of fallacies). Thus, it would be useful to define relevance also as a property of sources: more precisely, as the property of being likely to provide relevant information. Then such property could be modelled with the same formal tools used for competence and sincerity, to study the interaction between these three features in shaping argument evaluation, as well as the feedback they receive when argument quality proves to differ significantly from one's expectations. Among other things, integrating relevance into the picture would allow capturing a special family of dialectical infractions: omissions. When a source is aware of an argument relevant to the matter under discussion, and yet fails to mention it, then that source can be blamed for omitting something, due to a lack of understanding or honesty. In other words, an omission is not simply a failure to mention any valid argument, but rather a failure to validly convey relevant information. Without a proper analysis of relevance, omissions cannot be identified, because one cannot distinguish between the harmless (indeed, beneficial) habit of passing under silence everything that does not matter, and the damaging policy of withholding valuable information, on purpose or by mistake (reticence).

The modal approach to trust in information sources pioneered by Demolombe (2004) distinguishes six properties of the sender that the receiver may trust in, based on the relationship between informing, believing, and truth. Later on (Amgoud & Demolombe, 2014), it has been proposed to use argumentation to allow reasoning about such properties. Four of the properties identified by Demolombe are listed below:

The rationale of integrating relevance in this modal approach is quite straightforward, yet its formal implementation may be less obvious. The original approach (as detailed in Amgoud & Demolombe) includes two modal operators, one for belief and the other for information sharing, as follows:

Finally, an interesting aspect of relevance is that it cannot be directly captured in an abstract argumentation framework (Dung, 1995), one of the most popular methods to formalise argument structures. The reason behind that limitation is twofold: first, the approach focuses on relationships between arguments and not on the internal structure of each argument, making internal relevance impossible to capture; second, the attack relationship is defined between nodes (arguments), whereas a challenge of relevance is essentially an attack against an edge of the argumentation graph, in that it questions the fact that argument p has anything to do with argument q.1212 This is why capturing the notion of relevance in that context requires expanding Dung's framework with meta-argumentation (Boella et al., 2009; Villata et al., 2013), in which binary relationships between arguments, e.g. support and attack, can be explicitly represented and reasoned about at the meta-level. It remains unclear, though, how this might enable to express specific criticisms of relevance, as opposed to more general attacks against the validity of a proposed relationship between two arguments. After all, challenging the claim that ‘p supports q’ or ‘p attacks q’ does not specify the reason behind the challenge: it could be relevance, but it could also be an objection against sufficiency – e.g. p is acknowledged as relevant for q, but insufficient to presumptively support q. Further work on meta-argumentation will be needed, if this method is to be used to model trust in relevance, along the lines pioneered by Villata et al. (2013).

Even if many possibilities are open to the formal modelling of trust in argumentative relevance, one might remain sceptical on the technological implications of this line of research. Do argument-based software applications really need to incorporate trust in relevance? The answer is a resounding ‘Yes!’. Not only because, quite obviously, the circulation of irrelevant chatter, be it among humans or artificial agents, is a capital sin in this age of information overload, and the ability to detect relevance a cardinal virtue (on that topic, it is always instructive to read Brin & Page, 1998, which illustrates the core ideas on relevance of what later became a small company known as Google). But also because users of argumentation technologies have the need for, and thus a justified expectation of, systems able to be maximally relevant, in the sense discussed in Section 3. Imagine a user who must decide whether to take medicine X to treat illness Y, with the help of an argument-based decision support system; there are no arguments against taking medicine X in the knowledge base of the system, while it contains the following three arguments in favour of that course of action:

6.Conclusion and future work

In this paper, we tried to make a case for the study of trust in argumentative relevance: after a brief survey of the literature, we proposed a distinction between external and internal relevance of arguments, and contended that trust is especially central to the latter; then we discussed several rational grounds that may justify the act of trusting in the relevance of an argument, and sketched possible formal developments for the ideas presented here, as well as their import for argumentation technologies. While much more remains to be said on the connection between trust, relevance, and arguments, we hope to have persuaded readers that this is indeed a worthy line of inquiry, thereby luring others to join us in further exploring its ramifications.

Nonetheless, we would like to close this paper with a word of caution, regarding its technological implications. While the relevance of trust for establishing argumentative relevance (pun intended) is hardly in doubt, its actual implementation faces several practical difficulties. In that regard, we do not completely share Walton's optimism on the promises of heuristics (as those discussed in Pearl, 1984) for the automatic assessment of relevance (Walton, 2004, pp. 120–123). In fact, while Walton's idea of using backward and forward chaining of arguments as (part of) a method to establish relevance has merits, it is likely to be often quite impractical, or even impossible to implement – namely, whenever we lack a shared knowledge structure and inference rules with our interlocutors and/or end-users of argumentative technologies. Walton is aware of this difficulty, but seems to see it just as a minor inconvenience, whereas we consider it a major obstacle: