Argument and explanation

1. Introduction

Arguments and explanations are invaluable elements of our everyday lives. Arguments help us establish support for claims and play a role in changing people’s beliefs about these claims, while explanations provide us with an understanding of the world around us. Owing to their pervasiveness and practical importance in our lives, argument and explanation have became the focus of extensive research within philosophy, psychology and artificial intelligence (AI). However, that research has not seen the degree of mutual integration it deserves.

The concepts of ‘argument’ and ‘explanation’ are closely intertwined and they are multiply interrelated within cognitive science and AI. The aim of the present paper is to provide broad overviews of research areas that, by their content, should be deeply connected, but, presently, remain almost wholly separate. In order to bridge those divides, we highlight the central issues within research on argument and explanation in both cognitive science and AI, respectively. Specifically, we proceed as follows: we (i) first set the stage with a brief discussion of the general concepts of argument and explanation; we then (ii) go through the respective literatures on argumentation in both AI and cognitive science, in particular psychology; we then do the same for (iii) explanation research. Finally, (iv) we bring argument and explanation together in order to suggest ways in which insights from argumentation research might inform explanation research, and vice versa, both within and across AI and cognitive science.

2. Argument versus explanation

What is an argument and what an explanation? We all know what they are intuitively. Beyond that, however, providing more explicit understanding of both notions and their relationship is not entirely trivial—precisely because both concepts seem so closely linked. Both argument and explanation have a common point of departure. First and foremost, both are answers to a why question: why is something the case? And both involve the provision of reasons in response. This parallel is so compelling that argument and explanation, in fact, coincide on some formal accounts.

Chief among these is the classic model of explanation in the philosophy of science, Hempel’s covering law model of explanation (also known as the deductive-nomological model) [1]. On this account, an explanation is a deductive derivation from a general law of nature. For example, we might try to explain why the sun is in a particular place today. This is our ‘explanandum’, the phenomenon to be explained. On Hempel’s account of scientific explanation we avail ourselves of one or more general laws—say, the laws of planetary motion—plus some particular facts—say, the position of the sun at a previous point in time. Putting these together, one shows that the explanandum (the sun’s current position) is derivable as a deductive derivation from that general law and the particular facts. That is, one shows that the phenomenon one is trying to explain follows logically from general law and particular facts.

In short, explanations—on this account—literally are arguments. It is at this point useful to clarify the ways in which the term ‘argument’ is itself multiply ambiguous [2,3]. The first sense of the term ‘argument’ is that of an argument as a ‘reason’. Giving an argument for something is providing a reason for it. Here, the strongest possible reasons are those from which a claim or conclusion follows by necessity.

This leads to the second, closely related, sense. Here, the term ‘argument’ is used not just to refer to the reason but to the unit comprising reason(s) and claim. That is how the term ‘logical argument’ is understood, and the way ‘explanation’ on Hempel’s account is an argument: the explanation is a structured unit comprising one or more premises and a conclusion.

Third, the term ‘argument’ is used to refer not just to information content, but to a social activity. Here, it is not (just) a single premise and conclusion pair or a sequence of inter-connected claims and counter-claims that is in focus, but also the dialogical, social, activity that is giving rise to these claims. For this sense of argument as a dialogical activity, the argumentation literature distinguishes multiple forms, contrasting, for example, a quarrel with a rational debate. The latter involves the exchange of reasons that are aimed at ‘convincing a reasonable critic’ [4]. It is this latter type of exchange that is the focus of this paper.

Explanation, as a term, shares a corresponding ambiguity: it can refer to the reason, and to the social activity of providing that reason in a particular context. That, too, reflects a commonality across the two terms that we return to below. We assume in the following that it will be clear from the context, for both argument and explanation, what specific meanings are intended.

Despite the notable overlap between argument and explanation just outlined, further reflection reveals that, however closely linked, the two are nevertheless distinct [5]. For example, in the context of rational argument, we are typically trying to advance reasons/arguments that seek to change others’ beliefs in an as yet uncertain claim [4,6]. By contrast, explanations can be provided for claims and events we already know to be true [7].

This can be illustrated with inference to the best explanation (IBE) [8–10]. A canonical example of inference to the best explanation is the following: we go into the kitchen and see that our cheese has been nibbled. The best hypothesis in this situation is that there was a mouse in the kitchen [11]. IBE proponents maintain that the very fact that the mouse hypothesis constitutes the best hypothesis vis a vis our nibbled cheese confers additional epistemic support to the mouse hypothesis being true. One may or may not subscribe to this theoretical position, but, clearly, the cheese-eating mouse is an explanation for the nibbled cheese. Equally clearly, the presence of a mouse is not (in this context) an argument for believing that the cheese has been nibbled. We already know that the cheese has been nibbled. Rather, nibbled cheese is an argument for the presence of a mouse and mouse is an explanation for nibbled cheese.

From this simple example, it becomes apparent that arguments and explanations are two different, separable concepts. We can (but need not only) have explanations for events that are certain, but we typically do not (though we sometimes can) consider arguments for things we already believe to be certain. Consideration of the simple mouse example illustrates cases of clear difference. However, there will also be cases where the distinction is blurred.

In summary, there are multiple, close links between the notions of argument and explanation, and the degree of conceptual overlap is such that there may be occasions where it is hard to clearly decide whether one is looking at one or the other, or even occasions where what is being advanced might be both. This strongly suggests that the two notions might be usefully studied together.

The main goal of this paper is to enable more integrated research into argument and explanation in future. Specifically, we are interested in bringing together the study of the two notions in psychology/cognitive science and AI.

3. Argument in artificial intelligence and psychology

(a) Argument in artificial intelligence

We start with a brief overview of argumentation research within AI. This is itself a rather disparate field, where many of the areas to be mentioned have little connection with one another. One of the reasons for that, we suspect, is the historic accident by which much of the early literature on ‘argumentation’ in AI was not actually concerned with everyday natural language argument (in the third sense outlined above, i.e. as a dialectical activity involving the exchange of multiple, inter-related reasons). The goal of early work on argumentation in AI, arguably, did not view this dialectical activity as a target phenomenon that it wanted to understand in its own right (and consequently build systems to execute or, at least, support). Rather it was interested in argumentation as a tool for accomplishing something else. Some of the most foundational work on argumentation in AI, such as Dung style semantics, for example, has roots as a means of trying to elucidate logic programming [12–15].

Subsequently, this field developed a plethora of non-classical logics and argumentation frameworks as tools for dealing with uncertainty, in particular, tools for non-monotonic reasoning [16,17]. Much of this work was conceived, either explicitly or implicitly, as an alternative to using probability theory for coping with uncertainty [18] (and one lesson learned was that alternatives to probability could turn out to be ‘probability in disguise’ [19,20]).

As a result of its tool-based focus, this body of research in AI is often only rather loosely connected with research that has concerned itself more directly with everyday argument, in particular with natural language text.

The following strands seem worth highlighting in this latter context.

(i) The argument interchange format

The first involves the argument interchange format (AIF)—a canonical machine readable format for representing natural language arguments. A prime use for this format has been argument mapping, see figure 1. The sample map in the figure was drawn with the software tools of OVA (for ‘Online Visualization of Argument’) developed by Reed and co-workers [21]. The particular example is from a recent project examining different ways of representing scientific knowledge, in particular where there is scientific disagreement that might be important to communicate to policy makers in order to reflect accurately extant uncertainty [22]. OVA allows one to take a PDF, highlight text in that PDF, read that into a text box that OVA converts into a node in the map, and then, via pull-down menu, select different types of inferential relationships that connect it with other parts of an argument in order to form an overall map of the dialectical exchange. The utility of this is that it facilitates the creation of argument maps (including creation at scale, in multi-contributor projects) by allowing one to aggregate automatically different maps. It also supports navigation of such maps, and the use of a variety of computational processes defined over these.

• Download figure
• Open in new tab
• Download PowerPoint

Finally, anything annotated in OVA may be added to a large, openly accessible, database with many thousands of argument maps based on the AIF format which will continue to grow as long as people are using these types of tools. OVA is one of multiple tools for argument mapping (e.g. [23,24]), and one of many systems drawing on the AIF. Argument mapping remains popular for a wide range of tasks from large-scale computer-aided discourse visualization [25], through to critical reasoning [26].

For the map in figure 1, it was human analysts going through the text, identifying arguments and identifying appropriate argumentative relationships. In recent years, however, much research has gone into trying to automate such activities.

(ii) Argument mining

Automation of these elements is the focus of argument mining research (for reviews, see e.g. [23,27]). The goal here is to take the steps just outlined with respect to figure 1—extraction of natural language arguments and their relations from text, and the subsequent generation of machine-processable representations for computational models of argument—and have these be conducted by machine. As a field, argument mining has developed rapidly from a niche interest into a focal topic in AI [28] that now commands significant resources both in academia and the corporate sector. Argument mining research has itself brought together researchers from multiple areas such as natural language processing (NLP) and knowledge representation and reasoning. Lawrence & Reed [23] highlight three historic routes to argument mining research: sentiment analysis [29], controversy detection [30] and argumentative zoning [31].

Argumentative zoning seeks to take scientific documents and identify relevant argumentative components. This involved a standard computational linguistics process of researcher-developed annotation tools, which were then used to create corpora that serve as training materials for automated classification (e.g. [32]).

The overall goal of summarizing scholarly articles, however, has recently also entered firmly into the sights of transformer-based NLP tools: BERT [33] and the rapidly expanding list of large language models (LLMs).

(iii) Large language models

LLMs are models with hundreds of billions of parameters that estimate the probability distribution over word sequences [33,34]. Crucially, state-of-the-art LLMs are able to provide reasons for their solutions to problems (e.g. [35,36]) and provide evidence for their claims (e.g. [37]). They have also very rapidly become such a focal point of current discussion, not just within the academic literature, that they arguably need no further introduction. At the same time, the recent pace of developments has been such that any evaluations are likely to be superseded at the time of print (e.g. [38]). This makes more detailed analysis of current capabilities rather futile. There are, however, interesting questions about the relationship between AI and cognitive science that are posed by these models, and we refer the reader to two papers in this special issue that pursue these further [39,40].

(iv) Project Debater

LLMs also seem poised to soon challenge the quantum leap provided by IBM’s ‘Project Debater’ [41]. This system can be given a novel claim or proposition and then finds arguments in support of that claim and does so in an interactive debate with an opponent. Project Debater is capable of generating arguments that seemed convincing to the audience of a debating contest against a human debating champion. This very recently represented not just a wholly new level of automation (and performance) in the context of argument, but one that was hard to imagine a mere decade ago at the advent of argument mining research [42]. Project Debater rests on a combination of some of the aforementioned approaches and technologies. One obvious question for the future is the extent to which ‘generalist’ LLMs will be able to match (or exceed) such performance.

(v) Bayesian argumentation in artificial intelligence

Finally, it is worth mentioning a small pocket of research articles that concern themselves with Bayesian argumentation. These include early [43–46] and more recent [47,48] attempts to generate arguments from Bayesian Belief Networks (BBN, on these generally, see [49–51]). This work merits mention here not because it reflects a sizeable community or body of research within AI, but by virtue of constituting one of the comparatively few, potential points of connection between AI and cognitive science: by virtue of its use of the Bayesian framework, this AI research on argumentation links up with research on argumentation within psychology. We turn to that work next.

(b) Psychology of argumentation

By contrast to argumentation research within AI, the psychology of argumentation is a tiny field (for an introduction, see e.g. [2]). This seems at odds with the central role of argumentation across many real-world contexts. Possibly even more surprising is that a significant proportion of that work has not been conducted by psychologists. Much of what one could class as part of the psychology of argumentation was conducted either in education studies or in communication sciences (and we return to some of the reasons for this below).

One focal point within the psychology of argumentation is the body of work that has concerned itself with critical thinking. This research has sought to understand how one can foster critical thinking, and how good people are at evaluating certain types of arguments [52]. Critical thinking research within education studies (and, relatedly, within developmental psychology) has made wide-spread use of the Toulmin framework developed by Stephen Toulmin in the 1950s [53]. The basic components of Toulmin’s framework are illustrated in the argument map shown in figure 2. Specifically, the Toulmin framework introduces a number of very general distinctions in terms of types of relations that obtain between different components of an overall more complex argument. The inferential relationship between a reason and a claim, for example, rests on the ‘warrant’, and that warrant may itself receive further support (backing). In effect, the warrant explicates why the reason is relevant to the claim.

• Download figure
• Open in new tab
• Download PowerPoint

The nonsense content in figure 2 is chosen to make salient the fact that this scheme captures little about content: the fact that something is classed as a reason for a claim is ultimately based on the fact that somebody advanced it as a reason for a claim. There is nothing in the Toulmin framework that tells one whether it is actually a good, sensible or cogent reason and hence one that should change one’s belief in the claim at issue (for discussion of this point, see [54]). This severely limits the scheme’s utility for the chosen purpose of understanding or fostering critical thinking.

In response, one attempt to move from the descriptive perspective of ‘simply given as a reason’ to the normative perspective of ‘constitutes a good reason’ lies in the so-called scheme-based approaches to argumentation [2,55]. These have become popular within the critical thinking literature, in the informal argument literature within philosophy, and within the AI argumentation literature.

The argument mapping software OVA (figure 1) (as described above) offers schemes identified in that research literature as inbuilt components: a user can select an ‘argument from expertise’ or an ‘appeal to popular opinion’ or particular types of causal argument to represent the inferential relationship (in effect, the warrant) between reason and claim. These argument schemes represent defeasible argument types that are putatively good, but that might be overturned by further evidence. In addition to identifying schemes that represent recurring patterns in everyday argument, the scheme-based literature has sought also to identify so-called ‘critical questions’ that assist with evaluation [55]. These questions offer standard considerations that might help identify a particular instance of this scheme as weak or strong, good argument or bad. This, in turn, has prompted an empirical literature examining how people try to reason with these [56]. The fact that these schemes are also used in a variety of computational systems creates a further point of overlap in argumentation research across AI and cognitive science (beyond OVA see, e.g. [57]).

There are two other research traditions with normative, philosophical orientation, that have prompted psychological research. First is research on the procedural rules that govern rational discourse (e.g. and under the header of ‘pragma-dialectics’ [58,59] , or ‘fairness rules’ [60–62]). Second is psychological research on reasoning [63,64], and more recently, Bayesian argumentation [54,65]. This research is explicitly concerned with ways to measure the degree of support that an argument actually conveys for a claim.

As just outlined, and despite its popularity in the critical thinking literature, the Toulmin framework offers no real normative component. The critical questions of the scheme-based literature improves on that, but the normative foundation of those questions themselves very much remains unexplained. It is in order to move beyond that, toward more fine-grained evaluation of argument quality, that the Bayesian framework has been employed. For example, it has been used both to provide a normative treatment of the so-called argument fallacies (examining the extent to which such arguments should be viewed as persuasive), and to then look, descriptively, at how people actually evaluate them relative to this normative standard.¹ We return to the implications of these interwoven normative and descriptive concerns at the end.

Finally, there is research relevant to a psychology of argumentation under the banner of ‘persuasion’ or ‘attitude change’ as studied within social psychology [67]. While some of that research involves reasons that might be classed as aimed at a ‘reasonable critic’ (and thus overlaps with the type of argument considered in this paper), other aspects of this literature pursue concerns that might be more appropriately classed as ‘marketing’. We likewise return to the persuasion literature in the final section of this paper.

4. Explanation in artificial intelligence and psychology

5. Bringing it all together

In this final section, we offer thoughts on bringing all of this research together. It should now be apparent that there are multiple reasons for why one would want to bring together these currently distinct four fields. The first is that the notions of argument and explanation are not only theoretically closely related notions, they also involve closely related practical applications. If one is interested in building an explainable AI, one should be taking an interest in what researchers are already doing with respect to machine generated arguments. This particular practical connection is obvious and is, at least to some extent, already being pursued.

However, the very fact that the four areas surveyed have all evolved as largely separate fields means also that theoretical unification is desirable in as much as unification is a natural concern of science. Furthermore, unification is also likely to be both theoretically and methodologically productive within the individual fields: it seems highly likely that these fields hold important shareable but as yet un-shared knowledge. Sharing that knowledge would allow these fields to meaningfully refocus some of their research agenda.

One example of a presently un-shared perspective that is likely to be productive concerns potential transfer from argumentation to explanation research. The psychology of explanation has surfaced a number of basic findings about the effects of providing an explanation: explanations increase our confidence in a claim; they increase our confidence that an event will occur when asked to explain a possible future event; and they increase our confidence regarding an event in the past for which we are not sure if it happened or not [93–96].

All of these are things that arguments do also and this, again, reflects the functional overlap and similarity between argument and explanation. In fact, the literature mentioned above that examined the extent to which people can distinguish faithfully between arguments and explanation has suggested also that people will use explanations to support a claim where evidence or arguments are sparse or missing [84,85]. Arguments and explanations clearly target some of the same functional space. Hence it is reasonable to expect that things that are functionally relevant for arguments should also play a role for explanations.

Cross-field sharing of perspectives may thus be beneficial inasmuch as there are features of this functional space that have been central (both practically and theoretically) from the perspective of argument, yet have barely begun to come into view in research on explanation.

Coming from the perspective of argumentation, it is salient that an argument is something that a concrete, specific agent (human or other) provides to a concrete other (or group of others), in a specific context. But this of course is also true of explanations [97]. As discussed in §2, both argument and explanation may be construed as activities, in this case as communicative acts.

One central concern in the communication of arguments (i.e. testimony) is the reliability of the source. For one, the above-mentioned social psychological literature on persuasion has spent 30 odd years on this issue. In that literature, the core models of persuasion have been the so-called dual route models such as the Elaboration Likelihood Model (ELM). These models have sought to identify cognitively distinct routes, or pathways, for convincing people [98–100]. One of these is taken to be an analytic route that focuses attention on the content of the argument. The other is a peripheral, heuristic, route that pays attention to characteristics of the source. Much research has gone into trying to understand the contexts in which people resort to one or the other, and what kinds of source characteristics are relevant to heuristic processing.

More recently, argumentation researchers have stressed that when coming at the distinction between source and content from a normative, Bayesian, perspective, both features of the source and the content of an argument will matter, and that these should interact in determining how much beliefs change [2,56,101–103]. Behavioural evidence now suggests that they interact in people’s intuitive, informal, evaluations of arguments too, and do so in ways that are not well-captured by extant social psychological models of persuasion [104,105].

Viewed from the perspective of the psychological literature on argumentation it thus seems surprising that there has been so little research on reliability and explanation, by comparison. This absence can be felt not just in the psychology of explanation, but in AI research also (particularly as the latter has already highlighted trust, i.e. perceptions of source reliability, as a central goal of xAI).

We have been conducting experimental investigations that indicate that source reliability (that is, characteristics of the person providing the explanation) have effects on the impact of the explanation and that the bi-directional dynamics between content and source mirror some of what has been found in argumentation research [106].

It is consequently encouraging that work on explanation in AI has now finally started to take note of the pragmatics of explanation, that is, an understanding of how interpretations of utterances are generated in particular contexts [107–109].

The second set of considerations for re-focusing the research agenda that emerges from trying to learn from the distinct perspectives across our four areas concerns methods. User testing research, specifically in explainable AI, is effectively applied psychology. This means not only that it can benefit, rather obviously, from the experiences of a long tradition of applied psychological research, there are also more specific, topic specific lessons to be learned.

As researchers who have studied argument and explanation within psychology, we think it would be fruitful for behavioural testing conducted on XAI to more strongly emphasize, and focus on, normative considerations. What we mean by normative considerations in this context is that one should be thinking about, and using as a way to structure one’s research, considerations of what constitutes a good argument and a good explanation. ‘Good’, here, is intended not just in the sense of whatever happens to actually convince somebody, but rather what should convince somebody, that is, what should convince a rational actor.

This matters for explainable AI because one ultimately wants people’s trust to be calibrated to the quality of the system [81]. One cannot just want people to believe or trust a system regardless, as doing so may be dangerous.

However, normative considerations are arguably even more important from a methodological perspective. In our view, researchers will need normative frameworks in order to structure the research in such a way that it becomes generalizable. We think this conclusion follows strongly from the history of research both on the psychology of argumentation and the psychology of explanation.

Our above discussion of the psychology of argumentation drew out some of the characteristics of work in that area that give reason to think there are deep substantive reasons for why this topic that is so central to our everyday lives received so little psychological attention. We see as chief among these the limitations of certain tools. Given just the Toulmin framework (figure 2), all one can really say, is that there is a claim, some reasons for it, some kind of support relationship and, possibly, a rebuttal. Beyond those crude distinctions, it provides no tools for identifying further categories or objects of study.

In other words, the scheme is too limited to enable meaningful theoretical predictions or identify types of arguments across which one might seek empirical generalizations. From the perspective of that framework, the only questions one can ‘see’ and hence ask in empirical studies, is whether particular individuals actually offer reasons, and how complex the inter-relationships between those reasons might be. Beyond that, one cannot really distinguish between someone saying 'it’s raining outside because the pavement is wet' as opposed to ‘strawberry ice cream is more popular than pistachio partly because humans prefer the colour’. These are simply two different arguments with nothing in common other than that both involve a claim and a reason; because they have entirely different content there is no way to form any kind of meaningful generalization over them.

The value of the Bayesian framework as a tool for studying argumentation has been that it allows researchers to ask normative questions about argument strength that attach to the specific content of what is being argued about (see also, [54]). This is made possible because probabilities are intensional and are determined by the content of a proposition [49]. Hence one can ask systematic questions about responses to arguments across different content instantiations. For example, one can ask whether people’s argument evaluation is closer to the normative standard when they are confronted with arguments describing scientific scenarios or with arguments involving familiar, everyday events [66]. Theoretically and practically meaningful questions about how people treat arguments of different content and context thus become possible because one can compare those very different arguments to the same normative standard. In the same vein, we consider it to be more than a coincidence that some of the most successful work on explanation has been based on normative considerations drawn from philosophy [88].

This leads us to believe that behavioural studies in the context of explainable AI will not generate cumulative insight on the user testing side without use of structuring frameworks. Research will be limited, we suspect, to collecting particulars without deeper insight: in effect, ‘this system did this specific thing and this is how convinced people were by it’, and then, in a different study ‘we did this specific thing with this completely different system and this is how people responded there’. Without a systematizing framework, this will produce little in the way of general insight or information gain. And this is precisely why both the psychology of argumentation, and the psychology of explanation, have availed themselves of extant normative frameworks.

Adopting tools like the Bayesian framework, for example, to study such questions may provide a theoretical framework that supports general insights. Beyond that, it should be a welcome and exciting prospect, last but not least, because the normative questions raised in the XAI context are themselves interesting; and both argument and explanation involve interesting normative concerns that are unaddressed to date.

6. Conclusion

To conclude, argumentation and explanation are closely related and overlapping, but nevertheless conceptually distinct, notions. Both argumentation and explanation constitute large topics of research in AI and sizable but smaller topics in psychology. We think closer theoretical and practical integration is required both across the argumentation–explanation dimension and the AI-psychology dimension. Such integration will naturally highlight shared constructs such as source reliability, and we suspect others will emerge from those comparisons. Finally, we suggest that without normative considerations to help derive theory to guide experimental work, future research will unlikely meet fully the practical challenges explainable AI is seeking to address, and will remain unlikely to yield robust, generalizable, insight. In short, there is much to gain from closer integration.

Data accessibility

This article has no additional data.

Authors' contributions

U.H.: writing—original draft, writing—review and editing; M.T.: writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

M.T. was supported by a Royal Academy of Engineering Fellowship.