5.1 The Known Knowns: Multimodality

The first trick is a no-brainer. We have a strict coding speed limit on the speech channel. Very well, we’ll use other channels as ancillary devices, for example gesture. All natural languages used in social interaction use multiple channels. For example, I can say ‘He went that way’ indicating leftwards with my hand. Or I can say ‘The boss says redo it’ while rolling my eyes, indicating disaffiliation with the message. Sometimes these different channels work to give the desirable redundancy (I say ‘He turned left’ while gesturing left), but more often they add new information. How many distinct channels or conduits of information are there actually?

Here we should pause to distinguish the channel, for example vocal-auditory vs. gestural-visual, from the medium, for example English vs. Swahili. English can be delivered in the visual channel by writing of course, or it can be finger spelled in sign language, hence the need for the distinction (Reference LyonsLyons 1977: Chapter 3). But that won’t be sufficient. We also need a notion of layering. Layering of one sort is a notion familiar to linguists through for example the distinct levels of phonetics and phonology, where elements of one layer constitute elements of another. But here we need a notion of overlayering, that is, the possibility of a layer carrying an independent (non-constitutive) signal. A contrastive prosody can be overlaid onto the segmental signal to indicate both grammatical (e.g. interrogative) functions and attitudinal ones (Reference LaddLadd 2014). Part of this has been described as paralanguage, for example the use of creaky voice, the relaxed vocal chords associated with self-confidence. Voice quality, amplitude, pitch and timbre can all be used in this way to add essential information to an utterance.

In a similar way we can view the gestural-visual channel as layered. I can gaze at you with narrowed or wide-open eyes, momentarily or in a sustained manner, with blinks or without blinks. A hand gesture to the left can be close to the body or extended, made with a loose hand or an index finger, and so forth. Each layer carries potentially independent meaning. As Reference AbercrombieAbercrombie (1968: 55) put it, ‘We speak with our vocal organs, but we converse with our entire bodies’.

So here is a way that our engineer can get around the coding bottleneck, by multiplying channels and within channels by multiplying layers. Given the number of potential channels and layers (some sketched in Figure 2) one can see that this can easily multiply the bit rate, the transmission speed, of human communication. Ray Birdwhistell, who invented the term kinesics for the study of bodily posture, estimated that at most a third of the content of human communication is verbalized.Footnote ⁶ That measure is suspect, but there is no doubt that the use of multiple channels can amplify and sometimes multiply the content of the spoken channel.

Figure 2

Circumventing the coding bottleneck by multiplying channels and layers. The figure is only illustrative – how many channels are there actually?

The study of all this is not new. The Greek and Roman orators already classified gestures, but systematic study began in the second half of the twentieth century, on the one hand by anthropologists interested in gesture and bodily deportment and on the other hand by phoneticians interested in the layering of the verbal signal. In addition, social psychologists got increasingly interested in non-verbal communication. A long-running project on the ‘Natural history of an interview’ involved many of the crucial pioneers in the non-vocal channels, including Birdwhistell, Hall, Kendon, Condon, Erikson, Bateson and others, who developed the first ways of describing and annotating filmed interaction.Footnote ⁷ These early scholars were mavericks outside of mainstream disciplines, but with videotape and then digitized video, and from about 2000 with the birth of digital video annotators like ELAN,Footnote ⁸ the study of the multi-layered nature of human communication has developed very rapidly. Recent technical progress is beginning to make it possible, using machine learning, to automatically retrieve facial expressions and particular gestures.

Pragmatics has embraced all these developments, under the rubric of ‘multimodality’, and the study of manual gesture in particular is well advanced. There are standard ways of breaking down hand movements, for example, into preparation phases, the stroke or main gesture, its potential hold, and then retraction (Reference Kita, van Gijn, van der Hulst, Wachsmuth and FröhlichKita et al. 1998). Interestingly, gestures seem to be integrated into comprehension at the same speed as the words that accompany them (Reference ÖzyurekÖzyurek 2014). Sign languages are particularly pertinent with regard to overlayering because of the ability to simultaneously layer multiple articulators both according to grammatical or linguistic rules and according to expressive needs – the face for example plays a critical role in addition to the two hands (Reference Loos, German and MeierLoos et al. 2022).

To sum up, much progress has been made in showing how multimodality compensates for the tight bottleneck on speech production. However, even a cursory review of what we already know throws up many unanswered questions, the Dark Matter that we are after.

5.2 The Corresponding Dark Matter: An Orchestra with 150 Instruments

The study of multimodal communication is relatively new, dependant as it has been on the innovation of recording technologies. So there is plenty we do not know, and some of the known unknowns are obvious. First, manual gestures have been highly studied, but all the other channels and layers lag far behind. The face has been particularly neglected, except in sign language where it sometimes plays a grammatical role (marking e.g. conditionals or questions) and where a lexical item may require a particular expression. But the role of the face in spoken interaction remains very underexplored. On the face alone many potentially co-occurring layers are possible: forehead (frontalis muscle), individual eyebrows, mouth, teeth and jaw position, head tilt and movement in three dimensions, and gaze. Reference DarwinDarwin (1872) advanced the idea that facial expressions may have an innate basis, a view taken to extremes by Reference Ekman, Ellsworth and FriesenEkman et al. (1972), but it is well known that communicative facial expressions are also culturally shaped (Reference RussellRussell 1995). Japanese tend to look to the eyes in situations where Americans look to the mouth, and I have worked in a Papuan culture where wrinkling the nose is not as predicted by Ekman associated with disgust but is rather an expression of ‘Wow!’ Most investigations of facial expressions have used still photos, while what the pragmaticist is interested in is fleeting expressions in social interaction, not adequately captured in stills.

So the situation is this: we have no idea how many layers or channels are reliably in use across cultures. We don’t even know what is in use in much-studied European languages. Tipped off by observing meaningful blinks in a Papuan language, we looked at blinks in Dutch and found that in that language long blinks (c. 400 ms) also carry systematic meaning (Reference Hömke, Holler and LevinsonHömke et al. 2018). That was certainly news to the multimodal community. There is every reason to think we simply haven’t discovered most of the kinesic code. Perhaps to compute the number of such layers we can at least start from the degrees of freedom in each of the articulators. Animators and anatomists know the human hand has 27 degrees of freedom (DOFs) or movement potential; double that for both hands, add arm flexions (7 DOFs * 2), shoulder raises (2 DOFs * 2), torso twists (6 DOFs in the spine), head and neck (6 DOFs) – that’s 98 degrees of movement. To that add Ekman’s 44 Facial Action Coding System (FACS) or facial muscle movements, discernibly distinct directions of eye gaze, eye narrowing and so on. That is well over 150 potential signalling devices. Of course, they are not all in play at any one time, but that is the size of the orchestra that every speaker deftly manages, and abstention of movement can be significant too, as in the deadpan face. Every utterance is a complex multimedia performance – more like a Berlioz mass than the solo flute of linguistic descriptions!

Signals in the visual-motoric channel are by no means the only things we do not understand theoretically. The interactional uses of paralinguistic vocal parameters are also quite obscure. For example, Reference LaverLaver (1994: 199) outlines a complex system of phonation types, where the air passing through the larynx can have a turbulence heard as whispery, while low frequency air pulses yield a creaky voice or vocal fry, or the glottis is stretched to produce falsetto. In some languages these are used for phonemic contrasts, but in all languages they are available for signalling. In North America, creaky voice is typical of males speaking to females, but not to males (Reference Wright, Mansfield and PanfiliWright et al. 2019): its low energy suggests confidence and relaxation. Similarly, there are many sub-phonemic vocalizations that can carry significance, the mhms, clicks and sniffs that pepper interactions (Reference DingemanseDingemanse 2020).

There is a more theoretical problem that needs addressing too. If we have a communication system with multiple channels and many layers, so that a speaker is emanating signals on say a dozen channels, how do we know what signals go together, that is how do we integrate all this (say a smile, a tease, a head shake and an eyebrow raise) into a single coherent message? The problem is exacerbated because the bits that demonstrably go together are not necessarily synchronous. For example, it has long been noted that gestures precede their ‘lexical affiliates’ by half a second or more – that is, one might gesture ‘left’ well in advance of saying ‘left’. Similarly, a tease may be accompanied by a smile in the very last syllable, or a headshake long precede the mention of the thing negated. As a result we have what has been called a ‘multimodal binding problem’ as sketched in Figure 3 (Reference Levinson and HagoortHoller & Levinson 2019).

Figure 3

The multimodal binding problem: How do we know which signals on each of these tiers or layers belong together? How do we unite them into a coherent message?

(Photo S.C. Levinson).

Although in interactive language use generally only one speaker speaks at a time (while simultaneously broadcasting on multiple layers non-verbally), the recipient produces a running commentary of (mostly) non-verbal feedback, again on multiple layers. The recipient may lean forward open mouthed, then raise eyebrows, and smile as the other speaks. How and to what extent these responses play a feedback role in the incoming speech stream is currently almost entirely unexplored. We will return to this point later in the text.

Finally, we now have a rich set of ways of recording social interaction utilizing multiple cameras, head-mounted microphones, eye-gaze monitoring, breathing timing, pupil dilation, EEG-measurement and so on. A number of labs have built up impressive corpora with multiple measures, allowing new findings to emerge (see e.g. Reference Kendrick, Holler and LevinsonKendrick et al. 2023). But at present we struggle to use these measures to full advantage, because our current understanding does not permit automatic coding of just the parameters relevant to participants; at the moment it must be largely done painstakingly by hand.

6.1 The Known Knowns: Action Potential

Our engineer has quite a few further tricks up his sleeve in his efforts to get around the tight constraint on coding speed. A second trick is borrowed from a branch of cryptography called steganography. Steganography is the art and science of hiding one message within another. A classic example is the following (possibly apocryphal) message from a spy intercepted during the First World War:

1. (1) Apparently neutral’s protest is thoroughly discounted and ignored. Isman hard hit. Blockade issue affects pretext for embargo on by-products, ejecting suets and vegetable oils.

If one extracts every second letter in each word one obtains the intended message ‘Pershing sails from NY June 1’. The philosopher Grice repeats another (apocryphal) example: The British General Napier, when asked to chase some brigands in Northern India, ended up conquering the whole province of Sindh. He is said to have telegraphed the Latin phrase peccavi ‘I have sinned’ (he had disobeyed his orders), but the English translation is a pun for ‘I have Sindh’. Here both the overt and the embedded message are relevant.

The principle is clear: you can double the content if you can send two messages simultaneously. Now, surprisingly perhaps, that is what we do the whole time. The philosopher JL Reference AustinAustin (1962) pointed out that when we utter an ordinary sentence in conversation we not only say things, we do things. Packaged up with the words is some kind of action. Austin called this an ‘illocutionary act’ carried by the ‘locutionary act’ of saying, and his key examples were things like promising, naming a ship, condemning a convict, or declaring war. In these cases, by an institutional arrangement, someone so empowered can create a new state of affairs just by following a precise procedure and uttering the appropriate words. He went on to point out that whenever we say something, we are in addition doing something, like making an undertaking that such-and-such is the case. By doing the one thing, we may (either by design or inadvertently) end up doing another: for example, if Vladimir casually remarks what a shame it would be if your daughter fell out of a window, you would be well advised to act cautiously. Such an indirect action (or actions, since they can be multiple) Austin called a perlocution (see Table 1).

The American philosopher Reference SearleSearle (1969) systematized Austin’s observations, suggesting that such illocutionary acts or speech acts can be characterized by a set of conditions that they must meet in order to succeed. These felicity conditions can then be used to suggest a universal set of possible speech act classes: representatives (assertions), directives (requests, questions), commissives (threats, promises), expressives (thankings, apologies …) and declarations (blessings, namings, condemnings and the like that rely on special institutional arrangements). Searle’s systematization has been much used, but it has serious flaws – for one thing, it has lost the transactional quality whereby what is achieved by utterances depends on what the other makes of it (Austin called this ‘uptake’).

The idea that words do things as well as say things may seem metaphorical. But utterances really are actions that fit into action sequences. Consider for example standing at a supermarket checkout, where there’s an exchange of verbal actions (announcing the total, thanking the customer) sequenced into a string of non-verbal actions (taking the money, handing back the change). Although the study of speech acts began in philosophy (the later ideas of Wittgenstein were critical here) and were systematized by Searle, it is conversational analysis (and the related interactional linguistics) that has thrown the most light on how an utterance comes to have the ‘force’ or action it actually has (see Reference LevinsonLevinson 2013a; Reference Depperman and HaughDepperman & Haugh 2022). The interesting thing is that in general there is no one-to-one mapping between the form of words and the actions performed – the exceptions are just those things Austin initially focussed on, like being pronounced guilty or being declared married. Otherwise, there is often and perhaps mostly a many-to-many possible relation between utterance form and action performed. So, while ‘Have you ever tried a heat pack?’ might seem built for doing conditional suggesting, in another context the form may just be a question (cf. ‘Have you ever tried hang gliding?’, although even that might be a conditional offer to tell a story). Although there are favourite forms for some classes of action (e.g. the ‘Could you …’, ‘Would you … ’ forms of English requesting), the relationship between form and function is complex (Reference Drew and Couper-KuhlenDrew & Couper-Kuhlen 2014).

Conversation analysts have found that besides the actions we have colloquial names for (things like requests, promises, suggestions, proposals, greetings) there are many kinds of actions that have a subliminal existence: for example pre-closings (the exchange of ‘Well’s before closing a phone call), pre-tellings (‘Did I tell you what happened on Friday?’), repair initiators (‘Huh?’), sequence-closing thirds (‘Ok’), etc. (Reference SchegloffSchegloff 2007). They have identified the kinds of context – mostly sequential contexts of successive actions – together with forms that may predispose the recipient to attribute specific actions to utterances. If some such characterization can be found, and it is usually a loose family resemblance, conversation analysts talk of a recurrent practice, and they continue to collect new practices. It was when Reference SchegloffSchegloff (1996) noticed a new practice that he speculated that our current knowledge was just the tip of an iceberg, and that there might be hundreds of undiagnosed actions in current employment, the rest being Dark Matter, an analogy I’ve purloined for the title of this Element.

Despite the complexities here, it is pretty clear that this is a good trick: we can greatly amplify the informative content of utterances through encoding actions. In principle, a single utterance can unleash a great chain of actions: ‘Have you ever considered dieting?’ is not likely to just be an idle question, but a pseudo-innocent suggestion, and therefore an insult into the bargain. Even a simple repair initiator like ‘What?’, may also be suggesting that the prior action is in inappropriate. Every utterance is potentially double or triple barrelled. However, it is equally clear that how this system works is rather obscure, which brings us quickly to consider the associated Dark Matter.

6.2 The Associated Dark Matter: Finding Actions

How actions are mapped onto words is one of the most important and most puzzling domains of pragmatics. Any analysis of verbal interaction depends on some kind of attribution of action to utterances. Our theoretical understanding is hampered in multiple ways.

First, there is unclarity over whether even within a cultural group there is a finite list of potential actions (see e.g. Reference Enfield and SidnellEnfield & Sidnell 2017). Perhaps the system is entirely flexible, allowing the ad hoc invention of actions, a bit along the lines Wittgenstein suggested in his theory of language games: we can invent a new game in which utterances may play new roles (Reference LevinsonLevinson 1979a). Second, there has been no systematic comparison of speech act types across a good sample of unrelated cultures. Anthropologists have reported cultures where, for example, there are no promises as we understand them (Reference RosaldoRosaldo 1982; Reference DurantiDuranti 2015), since the inner commitment or sincerity is not something the locals put store by. So we have no idea whether there is some kind of universal inventory of speech act types, and if so what it is. The sentence types interrogative, declarative and imperative are mostly discernible across languages, but these tend to have surprisingly wide functions. It is clear that there are culture-specific speech acts – the very institutional cases that Austin initially focussed on (marrying, christening, condemning and the like) are clear cases in point. But it is possible that a core set of functions is universal (Reference TomaselloTomasello 2008 suggested requesting, informing and sharing), and are supplemented by rich sets of culture-specific actions within idiosyncratic ‘language games’.Footnote ⁹

But by far the largest of our puzzles is how participants actually ascribe actions to utterances. One element of the puzzle is that despite the many-to-many possible mappings from utterances to actions, we apparently get it right nearly all the time. There are only a handful of systematic misreadings of action content noticed in the literature, and I have found only a few in years of trolling through transcripts. Moreover, when engaged in conversation, the phenomenology is certainty of ascription, not a continuous wonder at what the other means. A second element of the puzzle is that action ascription is fast and apparently effortless. Brain imaging suggests that speech act recognition can occur very early during the processing of an incoming turn – even within the first syllable or two, which is extraordinary (Reference Gisladottir, Bögels and LevinsonGisladottir et al. 2018). In general, the turn-taking system in conversation is going to force very fast decision making, and early response planning, but since one responds to the prior action, it is clear that everything hangs on fast action ascription.

A real mystery is how this mapping of function or action onto utterances is done so efficiently and fast. The many-to-many mapping and the possible open-endedness of the functions already appears to pose a formidable cognitive problem. Add to that the speed and accuracy and we have a serious scientific puzzle: How does it work? There are two main suggestions in the literature, but they both have their problems. One of them (championed by early AI research, Herb Clark and myself at various times)Footnote ¹⁰ imagines that to understand an utterance one reconstructs the likely communicative goal of the speaker. This would just be a special case of our general way of understanding other humans’ actions: we interpret behaviour on the assumption that it is goal driven. Seeing someone approach a door and put their hand in a pocket, we jump to the conclusion they are fishing for a key to open the door. In the same way, when in receipt of an utterance one asks what is the speaker’s purpose here? If a student asks outside the classroom ‘Is this the lecture on pragmatics?’, one might reply ‘No, you want Room 208’ foreseeing that they intend to go to the pragmatics lecture, so helping them on their way. Sometimes quite opaque intentions can nevertheless be discerned by interlocutors. Consider the following phone call (from Reference LevinsonLevinson 1983: 343), where the initial question is taken by the recipient to be fishing for an offer – a presumption that turns out to be right.

1. (2)

   C:	Hullo I was just ringing up to ask
   	if you were going to Bertrand’s party	← Q, Pre-Request
   		– Fishing for offer
   R:	Yes I thought you might be	← Joke at transparency
   C:	Heh
   R:	Yes would you like a lift?	← Offer
   C:	Oh I’d love one

So one theory would be that we are continually in the business of goal reconstruction – building models of what drives the other agent’s utterances and actions. This has the virtue of accounting for how a single turn at talk could perform multiple actions: a proximate goal can have an ultimate goal behind it. One problem for this theory is how far back do we go? In the phone call above, R guesses that C wants to know if R is going to the party as a precondition for asking for a ride; but does R also start speculating about why C wants to go the party, and if not, what stops the chain of inferencing? Goals are embedded in higher order goals recursively. A second problem is that we have no theory about how this can be done. Seeing someone buy some wine doesn’t tell us what they want it for; perhaps he will drink it, perhaps it is a gift, perhaps he wants to stock it in case of visitors. A logic of practical reasoning will take you from desires to ends that will satisfy them, but not in reverse, from the ends to the desires. A third problem is all this looks like very intensive cognitive reasoning, but mapping actions onto utterances has to be very fast if the response is not to be delayed.

The alternative theory is simpler. We just build up a vast association network between utterance forms, contexts and actions. So saying ‘Can you reach the butter?’ will just by familiarity come to be taken to be a request. When conversational analysts talk about practices, perhaps this is what they have in mind. A central observation in CA is that sequential context in a series of actions is often a powerful heuristic. So checking a precondition for an action is likely to foreshadow that action, hence the quick inference that the caller in example (2) above was going to ask for a ride. But that question is at the very beginning of a telephone call, just like ‘Can you reach the butter?’ might come out of the blue. So initial actions, for example the first parts of adjacency pairs (sequences like questions–answers, requests–compliances, etc.), often do not have much contextual support. Interactional linguists would add that prosody and linguistic nuance might play a central role in identifying action, but still it does not seem at all plausible that utterances are always or even often action-determinate just by virtue of their form.

There are other problems with this theory too. We do know that giant association engines can do marvels, as shown by current AI natural language processors. But they do require vast training data. If we were behaviourist machines like this, associating conversational sequence, context and form with actions, we would expect children (given less exposure) to make lots of action-ascription errors, and I don’t know any evidence for this. We would also expect the associations to have a probabilistic character, and then we ought not to be surprised if we misattribute actions. But that is not the phenomenology at all, as already remarked. Then there is the striking fact that novel or unusual form-function mappings do occur from time to time, and don’t seem to cause much trouble. A couple of examples are as follows:

1. (3)

   A:	I could eat the whole of that cake ← compliment
   B:	Thanks!	(takes prior as compliment)

1. (4)

   A:	I also have a dog
   B:	Oh I’m sorry ← disqualification for rental, after Sacks 1968)

The business of action ascription is the beating heart of language use – the most vital part of pragmatics. Arguably, the whole point of language is just this, to deliver actions: assertions, questions, re-assurances, proposals, etc. The sequence of these is the thread that ties together conversational activity. And our failure to understand how it actually works should be rather humbling.

7.1 The Known Knowns: Utterance-Type Meanings

Our engineer has another trick up his sleeve. If we have prearranged conventions, we can smuggle in some extra meaning. ‘One if by land, two if by sea’ was the signal that Paul Revere organized to be displayed in a Boston church tower to warn whether the British were coming by land or by sea. More elaborately, the BBC arranged to tell the French Resistance that the Normandy landings had begun by broadcasting verses of Verlaine’s. This is an old trick: in the Old Testament, Jonathan arranges to tell David, who is hiding near the archery practice, whether it is safe to show himself according to whether he tells his boy to fetch the arrows from this side or the far side. A more subtle mode of signalling is to arrange that the mere choice of words or the manner of speaking carry the intended signal.

There is a large range of tricks of this sort to amplify what we mean. They are all based on the choice of words or constructions or manner of production, which will, by tacit arrangement, signal specific messages. I will call these utterance-type inferences, since the choice of the general character of the formulation carries the extra meaning. The most celebrated of these are Reference GriceGrice’s (1967/1989) conversational implicatures. The tacit arrangement in this case are Grice’s maxims of cooperative conversation. My version of these maxims (Reference LevinsonLevinson 2000) boils down to three simple, largely iconic heuristics:

1. (1) What isn’t said, isn’t meant (roughly Grice’s Quantity 1 maxim)

2. (2) Simple form suggests normal extension (close to Grice’s Quantity 2 maxim)

3. (3) Abnormal form suggests abnormal extension (Grice’s maxim of Manner).

Given the first of these rules of thumb, if I say ‘I’ve eaten some of the cookies’, I suggest conversationally (conversationally implicate is the technical term) that I have not eaten all of them, because if I had meant ‘all’ I should have said so. These are the well-known Horn scales, and they operate wherever there are scales of communicative strength. So in the Avianca Flight 502 disaster (mentioned in Section 1.0), the failure of the crew to say ‘emergency’ rather than the ‘urgent’ they did say, led air traffic control to think that it was not an emergency.

The second and third of these rules of thumb work like this: If I say ‘She opened the door’ you will assume that she did it in the normal way by turning the handle, pushing the door, etc. If in contrast I say ‘She caused the door to open’, or ‘She managed to open the door’, you will assume (by rule of thumb (3)) that she did it in some other, special and unusual way. I once coined the paradoxical aphorism ‘the less you say the more you mean’ to capture the interaction between maxim (2) that encourages rich stereotypes and maxim (3) that inhibits them (Reference Levinson, Papi and VerschuerenLevinson 1987). So any equivalent way of saying the same thing will not necessarily carry the same inferences. It seems that in all cases we work with a metalinguistic principle of contrast (Reference Clark and MacWhinneyClark 1987). We ask ourselves: why did she say it this way, when instead she could have said it that other way – either by using a stronger item on a scale, or by using a simpler, more compact expression. Such a principle may be presumed to carry over to prosody.

A huge amount of research has been devoted to Grice’s maxims, and there are rival theories – apart from my own (Reference LevinsonLevinson 2000), the most prominent are Horn’s version (Reference HornHorn 1989) and Reference Sperber and WilsonSperber and Wilson’s (1986) relevance theory. They all rely on some kind of metalinguistic vigilance, in order to detect that additional inference is required.

A rather different kind of utterance-type meaning is so-called presupposition. These are inferences that again arise from the choice of words but seem more attached by convention. For example, the verb regret seems to presuppose the factuality of the thing regretted. So ‘Sue regrets yelling at Tom’ implies she did do so, and so does the negation (‘Sue doesn’t regret yelling at Tom’). It turns out that there is a very long list of English words and constructions that trigger inferences of these sorts that are invariant under negation (see e.g. Reference LevinsonLevinson 1983: 181–184). Presuppositions allow us to use more compact forms: instead of saying ‘Sue thought Tom was an idiot and he is one’ one could say ‘Sue realized Tom was an idiot’ with the same implication. Similarly, if I say ‘The Dean has stopped making sexist comments’, I do not need to add that he used to do so – that is presupposed. Once again there are competing theories of presuppositions, some theorists holding that they are part of semantic content, others holding that they are so-called conventional implicatures – forms with specific usage conditions like honorifics for example – and others like Reference AtlasAtlas (2005: Ch. 3) and myself thinking that they may not actually be so distinct from conversational implicatures of the Gricean sort. Relevant here though is that these inferences, like conversational implicatures, are ‘defeasible’, that is, they can be lifted or cancelled without contradiction either by a context they do not fit, or by explicitly querying them, as in ‘The Dean has stopped making sexist comments, if indeed he ever did do so’. In this way they share with conversational implicatures their presumptive but cancellable quality. Regardless of which theory you buy, presuppositions are additional inferences that can be smuggled through the production bottleneck by a careful choice of linguistic expression or construction.

But the pay-off of having presuppositions is amplified by the fact that presuppositions can be used to trigger yet more inferences. So an utterance like ‘He bought a new lawnmower, pulled the cord and the motor started’ presupposes that there is a unique identifiable cord and similarly an identifiable motor, but it also strongly suggests that the cord and the motor are parts of the lawnmower, and pulling the cord started the lawnmower. These so-called ‘bridging inferences’ (Reference Clark, Johnson-Laird and WasonClark 1977) are things we read in to an utterance, not meanings inherent in the sentence. Although triggered by presuppositions, bridging inferences themselves can be attributed to a version of the second simple heuristic (2) mentioned earlier that licenses, from minimal specification, maximally informative or stereotypical interpretations.

These methods of amplifying coded content by virtue of prearranged rules of thumb or pre-packaging of default assumptions are some of the most productive ways of circumventing the coding bottleneck. They are also among the most researched topics in pragmatics. But that does not mean we fully understand them, in fact there are plenty of residual puzzles, to which we now turn.

7.2 The Corresponding Dark Matter: How Many Principles?

Most theories of implicature start off from Grice’s maxims, and attempt to revise or improve them. (An exception may be relevance theory, although it also relies on a principle of relevance; Reference Sperber and WilsonSperber & Wilson 1986.) Grice viewed these as a natural side product of rational cooperation. The idea, as pointed out in the preceding section, is that such tacit rules of use generate unsaid overlays on messages. Most empirical work has been devoted to generalized conversation implicatures – that is, ones that are generated by metalinguistic considerations of what else could have been said. These are mostly the scalar and manner implicatures. The contrast with what else might have been said provides a search heuristic for what is actually implicated (so ‘He caused the car to stop’ makes one ask why didn’t the speaker say ‘He stopped the car’ – presumably because he didn’t do it in the normal manner). Similarly, failure to answer directly a question like ‘Is John in?’, substituting, say, the response ‘His coat is on the hook’ will implicate the recipient does not know the answer for certain. But outside constrained contexts like these we have no real understanding of how implicatures are calculated. The whole of what Grice called ‘particularized’ implicatures (like the suggestion that John is not far away because his coat is on the hook) remain much more puzzling: How does a recipient find from the forest of possibilities just the implicatures intended within just a few hundred milliseconds? Relevance theory – one of the few theories brave enough to tackle particularized implicatures – talks gaily of adding assumptions as required to derive something of relevance. But AI approaches to this problem – which is essentially Peirce’s problem of abduction (Reference Douven and ZaltaDouven 2021) – show that even in highly constrained domains finding an algorithm that will do anything like this is very problematic (Reference Blokpoel, Wareham, Haselager, Toni and van RooijBlokpoel et al. 2018). So here is a huge open question about how we process these inferences.

Grice went on after introducing his maxims to add ‘There are, of course, all sorts of other maxims (aesthetic, social or moral in character), such as “Be polite” … and these may also generate nonconventional implicatures’ (Reference GriceGrice 1989: 28). We might wish he had said more! Are there lots more maxims that we have failed to notice? The crucial test for such a maxim would be that both in its observation and in its violation it should generate specialized inferences.

I think there is plenty of room for suspicion that we are missing principles of language use which are hidden in plain view. Reference Lakoff, Corum, Cedric Smith-Stark and WeiserRobin Lakoff (1973) and Reference LeechGeoffrey Leech (1983) have run with politeness maxims, following Grice’s tip off, while Reference Brown and LevinsonBrown & Levinson (1978/1987) suggested instead that ‘face motivations’ might lead one to veer away from otherwise rational behaviour. One principle that has been suggested requires the acknowledgement of local opinion: ‘When expressing an opinion on a topic that has been previously discussed, a speaker should correctly indicate the cultural standing of that view in the relevant opinion community’ (Reference StraussStrauss 2004: 161), as in: ‘If there are more Spanish people in the country than there are, you know, more Spanish-speaking people than Americans, then-oh that was riddled with political incorrectness’ (op. cit. 178). It’s not quite clear that such a maxim would generate inferences, beyond the inference that if one doesn’t distance oneself from the consensus, one subscribes to it.

But I suspect that we are having a failure of the imagination here. So, let me fly a few kites, half-baked suggestions for further maxims.

8.1 The Known Knowns: How Language Use Goes on Holiday and Does Figures of Eight

In 1917, Marcel Duchamp stuck a urinal on top of a plinth, signed it ‘R. Mott, 1917’ and exhibited it as a work of art. This caused a great deal of excitement – an uproar might be a better description. And that was exactly what he intended. The juxtaposition of a urinal on a plinth in an art gallery suggested a whole slew of possible conclusions, amongst them that the art world stinks, anything is art if you sign it, a manufactured item like a urinal is actually quite artful if you can look at it in a detached way, let’s piss on art, and a whole load more!

Our engineer will borrow this trick. It’s not a precision instrument, it’s more like a shotgun or a blunderbuss, issuing a wide and diffuse spray of suggestions. Or perhaps, more like throwing a pebble in a pond, and seeing where the ripples go. But this sort of chain of suggestions clearly offers a way of once again avoiding that dreadful production bottleneck – we can mean a great deal more than we say, even indefinitely more.

Manny Schegloff once suggestedFootnote ¹² that in ascribing actions to utterances, the first thing you have to decide is: Is this serious or non-serious? ‘Non-serious’ uses of speech include jokes, teases, ironies, rhetorical questions and the like where the recipient has to detect first that it is counterfeit currency, and second what its purpose is. Delayed detection makes a good joke. Churchill was a master of the back-handed compliment: ‘Mr Atlee is a commendably modest man with much to be modest about’ or ‘We can always count on the Americans to do the right thing, after they have exhausted all the other possibilities’.

The fact that we have a genre of non-serious uses of language opens up the possibility of the ‘just joking’ defence. So, when Trump addressing a police academy said ‘When you guys (the police) put someone in the paddy wagon, please do NOT be too nice’, his press secretary had to deny he meant ‘rough up the detainees’. According to Reference SenftSenft (2008), in Trobriand culture any remark can be recast as ‘just joking’ (in biga sopa genre), but as the Trump example illustrates, the defence works best when combined with a figure of speech (here a litotes or negative understatement).

Figures of speech were already classified and studied in ancient rhetoric, and it is hard to say anything very new about them. Grice of course took the position that his maxims are the trip lines, which if flouted generate figurative meanings, but how exactly was left unclear. One modern development has been to point out the ubiquity of figures of speech in ordinary ways of speaking and their relation to analogical thinking; indeed perhaps we just think in a folk psychology peopled by metaphors like time is space, life is a journey, argument is warfare, and so on (Reference Lakoff and JohnsonLakoff & Johnson 1980). Another development has been to show that words are so often coerced into new construals by a juxtaposition of context that the effect is almost imperceptible (e.g. in ‘He began the sculpture’, begin expects an event complement, and the sculpture must now be understood as a process, not a thing; Reference PustejovskyPustejovsky 1995). Some recent relevance theory approaches end up treating metaphor in just this way, as coercion into a construal (and thus as an ‘explicature’ in that jargon; Reference CarstonCarston 2002). Many modern approaches treat metaphors and the like as semantic/grammatical anomalies or categorical misfits (however mild) from which an inference about speaker meaning is derived by cognitive means guided by pragmatic assumptions (see Reference Hills, Zalta and NodelmanHills 2022 for review). Irony on the other hand relies on the mismatch between what is said and what is taken for granted, the disjunction invoking the contrast between the actual situation and the situation in which the ironic statement would be true (so Reference Wilson, Sperber, Horn and WardWilson & Sperber 2004 liken irony to a quotation where the speaker’s attitude – outrage, amusement, ridicule or the like – must be inferred from the contrast between the actual and the described situations).

Our question here is how the tropes can be employed by our engineer to circumvent the production bottleneck, and the plan of course is to use the striking juxtaposition of ideas to trigger an avalanche of suggested notions. In this way, they can act like little haiku:

1. (5)

   This world?

   Moonlit dew

   flicked from a crane’s bill.

   (Eihei Dogen Kigen, loose translation/interpretation by Michael R. Burch).Footnote 13

Most ordinary language is not so spectacular of course, but it is because of this triggering of potentially indefinite inferences that figurative language plays such an important role in cherished forms of literature or song.

8.2 The Corresponding Dark Matter: Finding the Message

Despite the two millennia of thought about figurative uses of language, there are plenty of residual puzzles. For most figures of speech, we have suggestions about how they are processed but no real algorithms, and that for a reason: given the cascade of inferences potentially triggered, the speaker’s target inferences have to be narrowed down using many aspects of the conversational context, including the prior action sequence, the relationship between speaker and addressee, and such factors as the speaker’s reputation and sense of humour. So, as for much in pragmatics, the actual mechanisms of understanding elude us.

A leading question is to what extent the tropes and figures of speech familiar from English and other European languages are particular to the Western tradition, and to what extent the same inventory of devices can be found across the languages of the world. On the view that metaphor, hyperbole and so on are just varieties of loose talk (Reference Wilson, Sperber, Horn and WardWilson & Sperber 2004), then we can expect to find them everywhere, although not necessarily enshrined and developed in a rhetorical or poetical tradition. On the view that cognition itself is indelibly metaphorical (Reference Lakoff and JohnsonLakoff & Johnson 1980), we might even think that the very same metaphors would have cross-cultural currency – which seems certainly to be sometimes false (Reference KovecsesKovecses 2005). Particular figures of speech may of course be constrained by grammatical resources (e.g. litotes of the kind ‘I’ll be not unhappy if he fails’ depend on the availability of lexical negation). As for irony, although there are comparisons between well-known languages, I know of no large-scale comparison of ironic language use. In general, for all the figures of speech, it is unclear to what extent this is a literary phenomenon, that is, whether such uses of language are largely restricted to, or at least most elaborated in, languages with extensive literary traditions (which account for less than about 5 per cent of languages). But it is likely that all cultures make use of figurative language in bardic traditions, religious contexts or song styles – what is unclear is what use is made of the whole palette of possibilities and how pervasive that use is.

It has long been noted in the child language literature that metaphors cause comprehension difficulties in preschool children, although it remains a matter of debate how early children themselves use intended metaphors (Reference PouscoulousPouscoulous 2011; Reference Falkum, Scott, Clark and CarstonFalkum 2019). A late development might argue against the ‘loose language’ (relevance theory) account of metaphor, since loose language itself doesn’t seem to be a problem for the child. The acquisition story is in this way pertinent to ongoing theory development. The acquisition of metaphorical abilities is clearly an active field of research, but again I do not think there is much systematic research in languages other than European ones, and certainly not in unwritten languages. If there are still plenty of residual questions for metaphor, the acquisition of other figures of speech (metonymy, rhetorical questions, understatement, irony, litotes and the like) belong more centrally in our Dark Matter – and again, one can expect findings here to reflect on central theoretical questions.

Potentially related to the acquisition story is the question of individual differences in the use and processing of figurative language. It is well known that people with pragmatic difficulties, including those on the autistic spectrum, have difficulty with figurative language use (but see Reference Kasirer and MashalKasirer & Mashal 2014). Again, this might not be predicted by the ‘loose language’ hypothesis of relevance theory. This suggests that there is likely a part of the general population that also comprehends or uses figurative language with some difficulty. Findings of this sort would be of relevance for educational purposes, and we may hope that this becomes properly investigated.

A final topic that cries out for attention is the study of multimodal signals associated with spoken figurative language. Non-serious uses, like teases, are typically associated with a very late smile. Ironies are said to be accompanied quite often by eyeball rolling (Reference Colston, Athanasiadou and ColstonColston 2020), although this may be a side effect of dissent from the view literally expressed (see Reference CliftClift 2021). Spatial metaphors are very likely associated with gestures, and more abstract metaphors may be too. But we know far too little about the extent to which in spoken interaction figurative language is ‘flagged’ in other non-verbal modalities.

9.1 The Known Knowns: Trading on Common Ground

We come to the last trick we review here, namely the strategy of piggy-backing on the context.

One way of appreciating how important this trick is, is to think about how the proper usage of words requires the world to be a certain way. Consider the following nonsensical exchange from Alice’s Adventures in Wonderland (Chapter VII) where the context doesn’t meet the conditions that the language presumes:Footnote ¹⁴

1. (6)

   ‘Have some wine,’ the March Hare said in an encouraging tone.

   Alice looked all round the table, but there was nothing on it but tea.

   ‘I don’t see any wine,’ she remarked.

   ‘There isn’t any,’ said the March Hare.

Or consider the utterance ‘May we come in?’ (Reference FillmoreFillmore 1971). The appropriate or felicitous use of this requires the context to conform to the following conditions: There must be an enclosure of some sort that includes the addressee but not the speaker, the speaker must have companions also outside, the addressee must have some authority to permit entrance that the speaker lacks, and the proposed motion is towards the addressee. In this sort of way utterances both conform to contexts and project them, allowing us to smuggle in our presumptions through the encoding bottleneck.

A first and obvious way we exploit the context is through deictic reference (as in May we come in?). We can avoid all sorts of descriptive complexities by simply pointing and saying ‘this’. Philosophers do a big song and dance about deictic reference because it relativizes what is said to a particular speaker, recipient, time and place. It makes language a curious kind of coinage that changes its value according to context (puzzling toddlers in the process). But it saves vast coding complexities, and contributes enormously to utterance compactness – ‘I am here’ is such a saving over ‘Steve Levinson is in the village of Bourn at 8:45 a.m. on July 28^th 2023’ (when I actually wrote this sentence). Deixis then rightly has a pride of place in every pragmatics textbook. The related process of anaphora again saves us repeating the descriptive identifiers of the protagonists in our stories.

The second most important way we exploit the context in order to condense our utterances is by presupposing whenever we can. I can say ‘I’m sorry I’m late, the train broke down’, and now you know without me separately stating it that I came by train. Observing how we actually use names in conversation, Reference Sacks, Schegloff and PsathasSacks & Schegloff (1979) formulated the maxim ‘oversuppose and undertell!’. The reason this can work is because there are ways to indicate tentative trial uses, for example by a questioning pitch or ‘try marker’. So we can gently escalate as needed as in ‘It turns out it was Ben? Ben Wallace? The man who lives in the apartment above me?’. The dangers of over-supposing can thus be mitigated by conditionally offering repair in advance. Presupposition, we have seen, in turn can trigger other inferences, so again presupposing rather than saying will greatly squeeze down what we actually have to say.Footnote ¹⁵

A third important way in which the context can be used in the compression business is by the simple expedient of assuming that the recipient will add the current utterance as an additional premise to an existing stock of premises – the contents of background knowledge or prior discourse – and crank out any additional inferences. One of the attractions of relevance theory (or any incremental account of utterance content, for that matter) is the kind of account it gives of the following little interchange:

1. (7)

   A:	‘Fancy a coffee?’
   B:	‘I need to sleep’

where the wish to sleep added to the fact that coffee is a stimulant, and stimulants interfere with sleep, allows us to crank out B’s answer ‘No!’ Unifying the new utterance with the accumulated context offers an ever-enlarging set of premises for deriving inferences.

Interactive language use occurs not only in conversation but also in many institutional contexts. Again, any fixed arrangements can be exploited. I once made the mistake, towards the end of a lecture, of saying ‘Next week, we’ll look in detail at Austin’s theory of illocutionary acts’, only to find that I had thereby unwittingly ended the lecture early as the students packed up their books and started to go! In the same way, we can start a committee meeting by saying ‘Well, we seem to all be here’. Here, we can formulate a maxim of the sort: ‘Mentioning the preconditions for starting an activity, or the preconditions for ending it, is a way of starting, or conversely ending, the activity in question’. This is possible because institutions are organized in rule-bound activities in which language plays a specific bounded role (these are Wittgenstein’s language games). In these specific activities, particular but simple utterances can do very complex things. Consider ‘Checkmate!’, ‘Guilty’, ‘Over’, ‘Order!’ said respectively in a chess game, a court of law, a cricket pitch and the House of Commons (Reference LevinsonLevinson 1979a). Again, we leverage the context to condense what needs to be said to a minimum.

Conversational analysts have emphasized the importance of the conversational sequence, the ordered sequence of actions performed by a string of turns. In trying to grasp what action is being performed by an interlocutor, participants must follow the real estate adage ‘location, location, location!’. Thus ‘Okay’ after a request has a different force than after ‘How are you?’ or after another ‘Okay’ where it presages leave-taking. Interestingly, conversationalists have to track the structure of such sequences sometimes far back, due to the potential for an elaborate embedding structure. The following example shows recursive embedding of adjacency pairs up to four or five levels deep (each level marked below) – deeper than any recursion in spoken language syntax incidentally (Reference LevinsonLevinson 2013a). Consider the following embedded structure in a service telephone call, where the depth of embedding is represented by levels labelled L1–L5:

1. (8) Telephone call to an artists’ supply store (abbreviated from Reference LevinsonLevinson 1983: 305)C: .. I ordered some paint, some vermillion. And I wanted to order some more, the name’s Boyd (L1)R: Yes how many tubes would you like sir? (L2)C: ... What’s the price now with VAT? (L3)R: I’ll just work that out for you (L4)C: Thanks (L4)(10.0 second pause)R: Three pounds nineteen a tube sir (L3)C: Three nineteen is it= (L4)R: Yeah (L4)C: That’s for the large tube? (L5)R: Well yeah it’s the 37 ccs (L5)C: I’ll just eh ring you back I have to work out how many I’ll need (L2)

Conversational sequence provides a rolling context, which allows for massive condensations of content not only through anaphoric devices but also through ellipses. After initiatory actions – first parts of adjacency pairs like requests, offers and the like – responses are typically elliptical. Note in the following German example how the case marker in the response tracks the missing verb (accusative on the definite article): in some basic way the full sentence ‘I would like to drink a white wine’ has been computed then abbreviated to ‘an(ACC) white wine’. That’s the savings ellipsis buys you, so allowing us to slip sideways through the production bottleneck.

1. (9)

   A:	Was möchten Sie trinken?
   B:	Bitte einen Weißwein (instead of Ich möchte bitte einen Weißwein trinken)

In this section we have reviewed perhaps the most important ways in which we can exploit the context to reduce our coding time. Deft leveraging of the context allows just the tip of the iceberg of communicated content to be exposed to the ears and eyes – the rest we can detect below the waves.

9.2 The Black Hole of Context

If there is a black hole in pragmatics it is the notion of context: the more you dig, the deeper you go – there doesn’t seem to be an end to the factors that might be relevant to understanding a talk exchange (see Reference Goodwin and DurantiGoodwin & Duranti 1992 for a selection of views). There’s the setting in which we talk, the relationships between the interlocutors, their entire knowledge of their worlds, the current conversational sequence, the topic under discussion, all the previous conversations they have had, the superimposition and layering of activities (e.g. eating, drinking, priority talk organizing eating vs. other conversation) and the list goes on. There may not be a principled end to the list: as the prior section suggests, we leverage anything we can.

What is however Dark Matter worthy of investigation is the way in which each of these contributing factors is structured. It is that structure that gives affordances for exploitation for our compression purposes. For example, the deictic parameters recognized by a particular language allow abbreviated reference. Western philosophers assume an agent, a time, a location and (if they have thought about demonstratives) a gestural demonstration. But even for English we’ll need more than that – certainly the addressee, spatial arrangements and the discourse context. In Kwak’wala subjects and objects have to be routinely marked for visibility to the speaker; other systems have exotic distinctions in elevation for demonstratives, or honorifics for person reference. In actual fact, the apparatus you need is quite complex, and varies for almost every language. There are some cross-cultural surveys of demonstratives in particular (charting the number of distinctions, Reference DiesselDiessel 1999, or exploring usage in a limited number of languages, Reference Da Milano, Ramat and RomaDa Milano 2007, Reference Levinson, Cutfield, Dunn, Enfield and MeiraLevinson et al. 2018). But there is much more to do in understanding cross-linguistic variation. Here is a domain crying out for the recording of multimodal behaviour across many languages. Until we have much more descriptive studies using video recording, serious generalization will evade us.

Another aspect of context that we know is highly structured is the sequential organization of conversation (Reference SchegloffSchegloff 2007). But again, we still know rather little about the universality or otherwise of these structures; the first cross-cultural comparison of a range of conversational structures covering a dozen languages is recent and suggests much may be invariant (Reference Kendrick, Brown and DingemanseKendrick et al. 2020). Similarly, repair organization (specifically, how one signals comprehension problems) shows striking parallels across languages (Reference Dingemanse, Roberts and BaranovaDingemanse et al. 2015). However, given the expected differences in the action inventories of different cultures, there is likely to be systematic variation in some notable respects. A related, relatively neglected domain is how syntactic structures are specialized for placement in these action sequences. This has been a focus of work in interactional linguistics (Reference Couper-Kuhlen and SeltingCouper-Kuhlen & Selting 2017), but a great deal more could be done here.

The study of many other aspects of context from this structural point of view seem quite neglected. Cross-cultural patterns in reference to persons and places have had some recent attention (Reference Enfield and StiversEnfield & Stivers 2007; Reference Mark, Turk, Burenhult and SteaMark et al. 2011) again revealing principles guiding the degree to which reference has to be descriptively explicit. Social relationships between interlocutors have of course a fundamental effect not only on what can be taken as given, but also on the interpretation of speech acts (Reference HoltgravesHoltgraves 1994) and much else besides, in a way organized by local social structures. Many other aspects of context, like topical organization, remain fairly obscure.

But we come now to the central puzzle of context. We have to swim in a vast sea of attributed beliefs. In order to exploit background knowledge, we must assume mutual access to designated parts of it shared between us and our current interlocutors. But how do we judge what those parts are? There have been learned disputes about even the very form of these shared beliefs – do I believe that you believe that I believe … (ad infinitum) that we see a table before us? Reference ClarkClark (1996: 96ff) offers a simpler resolution in the sharing of a basis for beliefs (e.g. the book in front of us can be presumed common ground because we share the basis for the belief, namely our co-presence with the book). But regardless of that, how do we make these fine judgements about what it is reasonable to presume? What dimensions of social organization make it likely that you and I share knowledge about some domain?

To appreciate the miracle of this consider that in order not to be the party bore I must not tell you anything you already know, and above all, not anything I have already told you. But this requires a vast inventory of whom I have told what to, boggling in its size and complexity. Consider too Reference Sacks and JeffersonSacks’ (1967 (1995: 560ff)) observation that in the case of bad news, the order of the telling must be first to those most affected, a principle observed after every disaster. Again, this reflects a structured social organization, and it is this structure that apparently allows us to predict or recall what would and wouldn’t be news to anyone. Without this ability to guess or recover shared background assumptions, our communications would either be redundant, or they would misfire on mistaken presuppositions. As mentioned, the principle of ‘over-suppose’ backed up with repair cuts us some slack, but that is all.

There is one other very powerful idea about how we handle context. Gumperz developed the concept of contextualization cues (see Reference Levinson, Eerdmans, Prevignano and ThibaultLevinson 2003 for concise exposition). The idea in a nutshell is that an utterance can carry with it, through modulation of lexicon, grammar and prosody, the very contextual frame within which it should be interpreted, a bit like a snail that carries its own house around with it. A simple example of this might be code-switching into a familial dialect or code to indicate a kinship context. This was Gumperz’s answer to Minsky’s ‘frame problem’, the computational problem of finding the right axioms to get the inferences that need updating. It is a powerful idea too little developed, which may help us to understand how we find our way in the black hole of context.Footnote ¹⁶

9.3 Summarizing the Tricks for Circumventing the Production Bottleneck

Just to rapidly take stock of where we are, we’ve reviewed five tricks for escaping what seems like an absolute limit on the speed and efficiency of human communication, namely the production bottleneck. These tricks exploit the residual cognitive capacity to understand much more than what is actually said. The first trick was to multiply channels and modalities, and the main elements of Dark Matter were the neglected channels and layers and the ‘binding problem’ that unified multifarious signals into a coherent message. The second trick was to multiply the content by adding an action content to each utterance, adding illocution to locution. Here the Dark Matter is how we actually do action attribution, this key capacity on which human linguistic interaction rests. The third trick was to have a tacit agreement about how the form of an utterance might be taken to favour particular construals, using generalized conversational implicatures, presuppositions and the further bridging inferences they afford. Here many questions arise about the exact nature of that tacit agreement, and whether there may be many such maxims we simply have failed to spot. In addition, there is the continuing puzzle about where presuppositions come from – are they arbitrary conventions or systematic side-products? The fourth trick was the use of figurative or non-literal language that delivers a diffuse spray of possible inference chains – which do we follow and why? How cognitively complex is it to understand these forms of language use? The fifth trick was to exploit the context to maximal effect, by not saying what can be only minimally specified or presupposed. Here our central puzzle was how we know what others do and do not know, and the puzzle of how we keep track of what we’ve told and not told particular others. For each of our tricks there are major domains that we can see need studying but remain relatively unexplored. And across all domains there is a striking lack of information on the small, indigenous languages that form 95 per cent of language diversity in the world. Table 2 puts this in tabular form, just noting the main outlines of the argument. We turn now to an even bigger puzzle.