Entrenchment: a distributed perspective

19. Entrenchment: A distributed perspective Stephen J. Cowley1 Abstract The paper traces memory consolidation, chunking and automatization to, not minds, but agentenvironment interactions. On this view, the self-organising powers of a Central Nervous System (CNS) use entrenching to shape a living system’s routines. While this applies to many species, in humans the results also sustain talking, listening and literacy. Taking a distributed perspective on language, I ask how skills arise as people speak, use linguistic artifacts, institutions and social practices. To focus on what is entrenched, I begin by contrasting everyday uses of ‘information’ with scientific applications of the term. I stress that, in finding an entrenched pattern one measures differences in frequency (e.g. ‘and’ is more frequent than ‘but’) that relate only indirectly to activity or, more precisely, what is measured (e.g. utterings of ‘and’ and ‘but’). Human language unites culture, biology and neuro-behavioral activity. As a result information of many kinds becomes ‘entrenched’. Neuro-computation draws on linguistic aspects of individual and social life as people attune their CNS to select ways of using experience. Unlike other species, humans build language routines on the basis of speech habits. In pursuing the argument, first, I trace entrenchment to pragmatics or concerted social action. Second, I stress that agent-environment interactions enable robots to simulate language-use. Brains, I suggest, are more likely to use sensorimotor-based quasi-robotic functions than stored facts about language-types (‘linguistic forms’). Frequency comparisons can thus link a history of entrenching to what agents actually do. Just as embrained bodies enable coordination, embodied brains enable persons to use the environment to establish functions. People use the world, listen, talk and, thus, mesh language with action and perception. Given social resources – languages, inscriptions and social practices – people use action and pragmatics to become linguistic actors. 1. Introduction The paper approaches entrenchment by linking radical embodied cognitive science (Chemero, 2011) to a distributed view of language (see Cowley, 2007; 2011a; 2014). The concept of entrenchment is used to approach how information, brains, bodies and language interact and, thus, to offer an alternative to input-output models. Using pragmatic evidence I argue that facts about linguistic entrenchment show language to be multi-scalar: frequency effects influence brains, how people coordinate and, in mid-term scales, how they say things. Verbal constraints use information that is statistical, gives rise to habits and can be used to reflect on emotions, reasons and meanings. In short, it enables people to act with words while transforming brains and the sense of self. Entrenchment thus sustains linguistic routines as people interlace actional and pragmatic skills with artifice and cultural resources (including languages). While entrenching appears in all embrained species, its linguistic counterpart both characterises homo sapiens sapiens and, strikingly, can be readily simulated in robots. Infants manifestly develop skills as they self-calibrate by making use of social knowledge to evaluate actions. As they become persons, action and pragmatics thus open up

1

Corresponding author: Stephen Cowley, Centre for Human Interactivity and the COMAC Cluster, Department of Language and Communication, University of Southern Denmark (Slagelse). [email protected]

1

a world of social relationships, experience, and ways of cooperating in self-display that are extended by skilled linguistic action. Radical embodied cognitive science posits that agent-environment interactions ground cognitive capacities. By focusing on how agents – and lineages – coordinate with each other and the world, attention turns to being with the world. As a post-cognitive view, no appeal is made to fixed knowledge of an objective world. Brains need, not representational content, but bodies that drive action and perception. Though offering only brief discussion of embodied cognition, such positions have become influential where biology is taken to ground human cognition. Language is traced to, not how it appears to an observer, but a multi-scalar history. On a distributed perspective (Cowley, 2014), language enacts human action-perception (it is ecological) by linking epigenetically based capacities for social coordination (its dialogical bases) with recurrent historical patterns (its verbal aspect). People need, not a language faculty, but skills in linking linguistic coordination to multiscalar cultural dynamics. Since language reduces to neither an object nor a system of units, Nigel Love (2004) appeals to two orders: while ‘first-order’ linguistic activity enacts wordings (nonce events that people perceive as tokens of phenomenal ‘types’), these do not reduce to replicating instances of identical behaviour (‘use’). Rather, first-order language by living, feeling human bodies also evokes the point of view used to reflect on ‘words’ and ‘meanings’. In one sense (that of the post-Saussurian tradition), second-order constructs (Love, 2004) dominate language. While Love’s approach is epistemological, Lassiter (2015) pursues the ontological consequences of stressing two orders. Echoing Aristotle, language consists in both matter and structure. Analysis of ‘structure’ as form and meaning thus ignores bodily activity (or matter) and falls foul off written language bias (Linell, 2005). In denying that writing captures the essence of language, the distributed perspective traces first-order activity to historical constraints. The first-order has been pursued as utteranceactivity (Cowley, 1994; Spurrett and Cowley, 2004), languaging (Maturana, 1978; Linell, 2009; Thibault, 2011), coordination (Raczaszek-Leonardi & Kelso, 2008) and public language-use (Fowler, 2010) all of which can be traced to sense-saturated coordination or interactivity (Steffensen, 2013). Yet, to understand language in its verbal aspect, one cannot ignore linguistic techniques (Bottineau, 2012) or routines. Techniques and routines draw on skills that enable humans to recall the past, anticipate and, above all, use circumstances to act, feel, think and activate beliefs. The main claim of this paper is that, by rethinking entrenchment, new light can be cast on the basis of linguistic routines. 2. Overview On a radical embodied view, cognition derives from agent-environment interactions. This raises the question of whether memory consolidation, chunking and automatization need computer-like brains to ‘process’ language or whether the relevant capacities arise as bodies attune to an interactional history. Rather than ‘represent’ content, a Central Nervous System may fine-tune its activity according to constraints imposed by the needs of communities, organisms and persons. Since both views link statistical facts to ‘meaningful’ talk and writing, I begin with an overview of scientific uses of ‘information’. In so doing, I stress that, to say that X is entrenched, is to identify a measure of frequency differences (e.g. ‘and’ is more frequent than ‘but’). Crucially, this differs from what is actually measured (e.g. a set of actual uses of ‘and’ and ‘but’). Next, I argue that, to trace the entrenching of linguistic ‘information’, activity must be traced to what statistics measure (coordinated action that enacts individual and social repertoires) rather than to the measures themselves (the statistical frequencies).

2

How entrenchment works can be framed by an input-output model posits that information drives neuro-computation. By contrast the radical embodied perspective stresses that the CNS uses a history of coordinating to entrench routines in human powers. The moot point is therefore whether entrenchment is to be traced to (formal) input and/or perception-enriched linguistic action. The difference is thus akin to asking if language is lexico-grammatical (do brains ‘construct a language’?) or whether skill with verbal patterns derives from bodily coordination. In defending radical embodiment, first, I trace entrenching to sensorimotor and pragmatic roots, and second show that robots can simulate pragmatic language-use. If brains are quasi-robotic, as I claim, measures of entrenchment pick out, not neural structures, but how entrenching is co-opted by a population’s neural re-use. Moreover, given that language is multi-scalar by nature, a conclusion can be drawn. Just as embrained bodies drive coordination, embodied brains can attune persons to selected aspects of a language ecology. In short, people learn to realise values by discovering (and making) ‘affordances’. Entrenchment connects action with perception as people set and meet each other’s expectations. As people mesh listening with talk and action, language experience transforms the life-world. Further, with literacy, people can also link socially derived skills to reading and making inscriptions. Entrenching thus gives access to new resources – inscriptions, practices and institutional norms. Entrenchment identifies, not a mechanism derived from ‘learning’, but a way of maintaining social environments where people make widely divergent use of cultural routines and languages (so-called ‘language-use’). 3. Entrenchment: information, brains and language While prefigured by Hermann Paul and Ferdinand de Saussure, recent work on entrenchment arose in Ronald Langacker’s view of linguistic cognition. Entrenchment came to be understood as describing how individuals and populations sensitise to differences in frequency between the use of linguistic patterns. Entrenchment applies across time-scales by identifying whatever-it-is that underpins statistical measures of linguistic frequency as well as statistical reaction time, speech recurrences, patterns in textual corpora, aspects of language learning and evidence of linguistic change. While its multi-scalar nature can seem astonishing (and I return to this below), like many others, I focus on its implications for human agency. Broadly, I share Schmid’s view: “As a first approximation, entrenchment can be understood as referring to a set of cognitive processes – mainly memory consolidation, chunking and automatization taking place in the minds of individual speakers” (Schmid, this volume). However, in turning to radical embodied cognitive science, I emphasise, not minds (whatever those are taken to be) but, rather, coordination by living bodies. Rejecting a parallel with hardware-software relations, I do not see entrenchment as an index of ‘learning’. This is because, while reliant on continuous neural re-organisation, human bodies use its results to transform their doings. Thus, even if we think that we know what language is (a dubious assumption), we should be wary of inflated claims. For example, it has been said that chunking is a “natural and unavoidable way of perceiving language” (Sinclair & Mauranen, 2006, p. 20). But are perceivers the persons, organisms, brains or some kind of integrated system? Are we to assume that chunking has the same basis in all time-scales? Do the same mechanisms underpin the entrenching found in studying reaction-time, discourse frequency, grammatical constructions and historical change? In a valiant attempt to tackle the issue, Blumenthal-Dramé (2012) turns to the brain. Her logic is that, if ‘language-perception’ is separate from circumstances, thinking, gesture, grammar and tone, chunking must draw on input (i.e., perception of form). Mine is the opposite. If entrenching is inseparable from perception-and-action, as I take it to be, its roots lie in agentenvironment interaction and, thus, in pragmatic judgements. 4. A sketch of recent cognitive science

3

Cognitive science once regarded mind as a representation-manipulating machine (Boden, 2008). The approach enshrined two assumptions: first, information processing applied to both computational information and a supposed human equivalent (e.g., saying or understanding “fire is hot”). Second, processing fell under ‘autonomous’ control and individual-centred competencies like vision, knowledge of chess or grammar ‘generation’. In the 1990s, however, the focus shifted to embodiment and its consequences. Driven by robotics, many asked how living systems selfmaintain by using flexible, adaptive behaviour. As cognitive internalism was challenged, ever more rejected views that reduce mind to ‘mental states’ (see Shapiro, 2010; Stewart, Gapenne & Di Paolo, 2010). For some, cognition – and mind – extend into artifacts and technology (e.g. Clark, 2008) and, for others, lived activity shapes life and sense-making (Froese & Di Paolo, 2011). While basic sense-making can be traced to agent-environment covariances (Hutto & Myin, 2013), others generalise the claim to human language (Harvey, 2015). Performance can be meaningful because, even if based in physical ‘information’ (which lacks inherent meaning in the relevant sense), living systems rely on ‘information’ for a situated organism. In the hey-day of generativism, ‘mind/brains’ were said to perceive, construct and produce language. While appeal to the mereological fallacy (Bennett & Hacker, 2003) warns against conflating brains and people, an alternative demands another view of language and action (Keestra and Cowley, 2009). In appealing to entrenchment, I begin by positing that physical ‘information’ can underpin routines. I argue that, while adequate to describe coordination between robots, more is needed to describe both coordination within organisms and how agent-environment relations bring forth a world for living agents. On this view, linguistic skills must be integrated with other capacities because they contribute to experience. Organs (including the CNS) do not ‘cause’ action; thus, though brain-damage can selectively impair function, it does this because brains sustain living bodies that enable language and other behaviour. It is only when one appeals to input-output systems that brains reduce to ‘producers’ and ‘processors’. Physiologically, the CNS contributes to action as agents also perceive. In wine tasting, say, a person may pick out a hint of plum, detect a sommelier’s uncertainty or, perhaps, recall a long forgotten day. Far from using symbol manipulation, such skills link judgement and language to experience with a world beyond the brain. Individual skills can thus be traced to a history of coordinating: perceiving plum or an unexpected resonance connects an observer to language games (Wittgenstein, 1958). So while global statistical structures in community behaviour contribute much to entrenching, an actor-perceiver also make sense (whatever that means) of what results. To address how construal bears on entrenching, one must therefore consider uses of information in science. Collier (2011) offers a clear, comprehensive account of how information is used of physics and living systems. For physicists, the universe displays what McKay (1969) calls “differences that make distinctions” and, in that sense, it is information. These distinctions link quanta, hierachical arrangements of matter and the statistics of flux or, in Collier’s terms, substantive information. Without distinctions, there could be no prospect of life. Accordingly, substantive information enables living and, by extension, neural activity, people and social organisation. However, unlike in physics, information enables life to form cohesive systems. For Collier (2011) functional information arises as living entities self-sustain, replicate and act. In Bateson’s (1979) phrase, living systems draw on ‘difference that makes a difference’ (p. 99) 2 . Applied in this sense, the term ‘information’ highlights organic mechanisms that transform substantive information. The case can 2

Whereas a distinction is, in principle, measurable, differences are distinctions that play a functional role for a system that exhibits cohesion, a system grounded in self-maintaining cellular processes.

4

be exemplified by how DNA uses substantive information. As is well-known its structured nucleotide sequences have remained largely unchanged for hundreds of millions of years. Nonetheless, in a living cell, they become functional in that they regulate the synthesis of proteins that drives metabolic processes and cellular function. DNA structure has a part in functional information because of how it co-functions with RNA and cellular messengers. By way of clarification, Collier places uses of ‘information’ in a nested hierarchy that shows ways of using the term. While most of the universe uses the differences that make distinctions (micro-physics or ‘it from bit’), this captures neither statistical change nor material structure (i.e., negentropy and hierarchical features). Thus rings 1-3 are purely physical. By contrast, rings 4-6 apply only to systems that manifest biological cohesion based on functional information. For example, ring 4 applies to molecular functions, 5 represents skills in acting and assessing action, and 6 human language. While not crucial to my argument, Collier (2011) classifies the latter applications as intentional and linguistic respectively.

1. It-from-bit 1

2. Negentropy 2

3

4

5

6 6

3. Hierarchial 4. Functional 5. Intentional 6. Linguistic

Figure 1. Six applications of information. The more central the type, the more kinds of ‘information’ come into play.

Since the ‘same’ substantive information (e.g., a protein or a pixel) can play many functional roles, both an organism’s immersion in an environment and any resulting entrenching arises as a living beings uses physical process. Where relevant to whole organisms, measures of substantive information pick out distinctions that map, at best, indirectly, onto ‘uses’. Further, when one pursues how organisms sensitise to differences between measures, entrenching is found in most biological processes. In its scientific senses, therefore, information links physical distinctions to their functional use in living systems. Functional information is not with the same as its measures (also called ‘information’). For example, in this second sense, tree rings give information about age. Though the measured structure is substantive, the measure (e.g., 17 rings), can be expressed as numerical, algorithmic or Shannon information. Yet, since the measure bears indirectly on the world, it ‘explains’ nothing about trees. In parallel, I will argue that, ‘entrenchment measures’ lack intrinsic explanatory value. By way of illustration, I offer a homely example of measures-as-information. Let us suppose that, on leaving her house, Sally turns right or left (houses opposite prevent her going straight ahead). She turns right, let us say, 70% of the time and, in 28% of cases, she goes left (2% of the time she returns inside). Although the measure establishes facts, it shows nothing about Sally or her world and, for the same reason, captures no specific use of information. Now let us consider the ‘entrenchment’ of decisions about turning. For example, we might compare the total data set with when Sally leaves

5

the house between 7 and 8 am on weekdays. Now, she goes right 90% of the time, left on 6%, and, on 4%, returns home. By highlighting an observer’s concerns, the measures reveal trends based on conditional probabilities. Early on weekdays, there is a 20% higher chance that she will choose to turn right and, strikingly, a 100% increase in the frequency of returning home. Since the ‘information’ means nothing to Sally (or her brain), it lacks explanatory power (It does not identify a ‘mechanism’).Yet the measures (the changing frequency effects) are valuable for observers. In making measures – and interpreting results – one picks up aspects of what is measured. Typically, these mix functional and substantive information. In the example, therefore, the distinctions/differences matter to Sally – indirectly. These aspects of ‘information’ offer an observer’s view of what is measured that can be used in predictions. For example, if I am to bet on Sally’s movements, I will predict that, on weekdays, she will turn right. She becomes a moving object in space and the measure refers to substantive information. However, I may ask why she turns right/left: in generating hypothesis, I can appeal to functional information. Does she turn right on the way to the bus stop or the office? Is Sally forgetful in the morning? In such cases, I postulate information that, at time t, may (or may not) matter to Sally. Inevitably, as I appeal to the social and physical world and, as an observer, I rely on language. However, the facts relate only indirectly to Sally’s history and are, in this sense, like a tree ring count. Crucially, they offer no support for an entrenchment ‘mechanism’. No measure can show, for example, if Sally has a regular rendez vous with a lover. Their indirectness arises because, as a person, Sally observes actions that others can also observe (she may want spies to believe that she has a lover). Given an observation history, she exerts a degree of control over her own coordinated action. Naturally enough, this complicates how the frequency attributed to biological functions relate to measures (as types and sub-types) and how information (i.e., the measured) functions for a person/organism. Below I will return to the issue of cognitive control. However, the negative argument is complete. It is that, in establishing entrenched differences (as with, say, habits of turning right), one leaves aside the nature of the information (or the mix of information) that is being measured. Emphatically, one provides no evidence for (or against) any inner mechanism. For this reason, we face what can be called the substantive question: to what extent do brains and people need to represent any given mix of information and to what extent do the measures pick up on finely tuned agent-environment relations? 5. Cognitivist views of entrenchment Cognitive linguists often posit a brain (or mind) that uses environmental structure to ‘discover’ how verbal patterns serve in the alleged construction of form-meaning mappings.3 Language is said to be localised in the head. When Chomsky (1965) revived mentalism by proposing a language faculty, he posited that, like a von Neumann machine, brains use ‘representations’. In the cognitive revolution, ‘input’ was purely formal. In defence of the view, others invoked methodological solipsism that placed language and cognition between the ears. Linguistic behaviour (‘performance’) was said to depend on brains that generate intentional states by processing input to an autonomous individual (who has ‘competence’).4 Where language is reduced to forms and use, brains are taken to deduce (or construct) knowledge about frequencies. Indeed, Chomsky’s 3

For example Jenny Saffran (2008) describes the relevant learning as simple: “To the extent that structure in the environment is patterned, learners with appropriate learning mechanisms can make use of that patterning to discover underlying structures” (2008, p. 180). She fails to note that the mechanisms remain unknown – they are posited because she assumes that (inner) language draws on ‘real’ structures. In fact, people are ‘exposed’ to situated gesturing and its physical effects – the data is not intrinsically verbal. 4 Enactivists offer a complementary view in regarding language as participatory sense-making prompts an autonomous organism to use interaction in coping with perturbances ( De Jaegher and Di Paolo, 2007; for critique, see Cowley and Gahrn-Andersen, 2015).

6

programme began as a challenge to the view that language learning used distributional effects based on measures of substantive information (Harris, 1998). If a brain grants central control, entrenchment-based learning must exploit context. This is because, internal cognitive processes must link input to internally stored meaning. Whereas Skinner (1957) treated formal ‘stimuli’ as the basis of language learning, Chomsky (1965) argued for a Universal Grammar by asserting that they are badly impoverished. Even today cognitivists adopt input-output models that treat entrenchment as intrinsic to context-bound language production and processing. For example, it may link to plans that drive to mechanistic sequences of utterances (see Pickering & Garrod, 2004, 2013). In often-used terms, a brain might ‘extract distributional regularities’ and, having done so, ‘store’ the results (e.g. Arnon, 2015). As in Figure 2, input is stored – and reconstructed – within the organism (the diagram’s inner block). Over time, learning reconfigures earlier input (E) as brains construct representations (R). Not only do these allegedly drive production but, when ‘input’ is processed, the results trigger the identification (and/or construction) of regularities. Linguistic knowledge becomes, in part, a matter of frequency effects based on extract–store-and-use. Whether in entrenched or representational guise, linguistic items can link predictive uses to action/language production. A classic exposition is Tomasello’s (2003) claim that children ‘construct’ a language by linking input, statistical learning, and recognitions/ readings of intentional states (as shown by R in Figure 2).

Figure 2. A cognitivist model. Interaction provides input that gives rise to the recognition and reading of intentional states (R) that allegedly suffice to construct (E) or entrench knowledge (in the brain)

The model can be summarized as follows: 1. Brains extract distributional information from input. 2. Brains identify linguistic patterns – and construct entrenched stores of linguistic ‘data’. 3. By linking linguistic ‘data’ to speech production, brains provide opportunities to repeat a cycle of input driven learning. 4. Over time, entrenched mechanisms self-construct to generate new (predictive) mechanisms that are also used in producing linguistic units (constructions). 5. Constructions enable a person to conform to local practices by linking speaking/hearing to other mechanisms (e.g., systems for intention identification and recognition). The view predicts covariance between frequency measures and learning what the observer measures (e.g., the emergence of constructions). Brains or minds ‘perceive language’ by mapping information

7

to structure that is covertly associated with measures of negentropy in speech.5 Implicitly, brains use statistical information (perhaps, together with intention recognition/reading) to solve Harnad’s (1991) symbol grounding problem (see Jost and Christiansen, this volume). Arnon (2015) applies the view to early stages of how children learn to talk. Using measures of differences between frequencies, she asserts that high frequency items ‘tend to be acquired earlier’ and shape both ‘correct production’ and ‘error patterns’. Such evidence, she argues, must be used in modelling acquisition (challenging the Skinner/Chomsky view that language consists in forms). For Arnon, if speech is formal it is also statistical information (for a child). By conflating the measure with the measured (and the observer with the child’s brain), she can leave aside how ‘information’ might drive entrenchment and behaviour. Rather, brains are taken to use relative frequency to formulate and identify chunks of speech. Without argument, organs (or minds) are taken to self-configure by using substantive (negentropic) information to identify and categorise language (i.e., to use measures that correspond to hers). On her view, it is as if Sally’s actions and beliefs depend on frequency observations about turning left and right. It is as if frequency measures pick out what Sally uses as ‘information’ when she decides where to turn. 5. The radical embodied alternative A radical embodied view begins with agent-environment relations and, specifically, how bodies attune to interactional resources. Phylogenetically as well as ontogenetically, dyads, groups and communities act to sustain linguistically saturated ecologies where their own skills develop. As coordinated activity, language patterns enable individuals to use substantive information in devising skills. The view thus rejects both intrinsic intentionality and central control. Since brains use many types of information, measures of frequency bear indirectly on models of usage and theories of language and learning. Measures of Shannon (or algorithmic) information cannot ‘explain’ whatever-is-measured but, rather, like Sally’s right and left turns, they pick out facts that mix the substantive with the functional. While entrenchment always captures differences that make distinctions (i.e., in principle, they are observable), these may have little bearing on how activity is related to verbal patterns. Rather, any function is taken to depend on their being (or having been) differences that make a difference for an agent. Living systems evolved long before brains came to harness functional information. For example, bacteria show feats of ‘intelligence’. In some species, functional information accrues to sound signalling and, in others, properties of sugar concentrations that afford discrimination. In that these differences constitute a bacteria-world, the results are adaptive. Thus, even without brains, statistics connect function with bidirectional organism/environment coupling. If bacteria adapt, analogous use of physical distinctions is likely to shape all adaptive behaviour. In humans, for example, it is plain that touch and other sensations use bidirectional coupling. Often, the case is exemplified by how a blind man uses a cane to feel out functional information. Humans also combine such bodily coordination with affect and routines. As a cultural species, much of what we do and say draws on a degree of collective control: substantive information draws on historically derived discriminations as functional information is used to enable particular behaviours. And so dyads, groups and communities self-maintain as people call each other to account and, in turn, act accountably. Human subjects track and direct one each other’s attending as they orchestrate movements and affect. As they individuate, organisms self-configure as persons who use historical patterns that 5

Measures of Shannon information are often associated with negentropy. In Collier’s (2010) terms – and as shown by Sally’s tale – conditional probabilities do not pick out functional information. For Saffran (2008), the puzzle is how structure in an environment can be paired with learning mechanisms that use the pattern to ‘discover’ linguistic structures.

8

linguists associate with language (Love’s (2004), second-order’). Drawing on slow cultural processes, persons gain skills in using (first-order) wordings in situated and cultural events. This happens when circumstances prompt people to integrate substantive information with historically derived forms of use; experience draws on culture (omitted from Figure 2). Substantive information triggers feedback that brings forth ways of using social and neural networks to come to terms with circumstances. Far from using fixed forms or meanings, as individuals, people navigate overlapping cognitive ‘cultural eco-systems’ (Hutchins, 2014) by using the structures that sustain social practices. Using material and normative constraints, they sustain coordinated activity as ongoing perception-action is interwoven with wordings that can be heard as having meaning potential (‘meaning potential’). Though finer distinctions can be drawn, Figure 3 generalises the view of distributed control. An organism-environment system can act within a dyad, relationship or group (henceforth a social system). Given experience, parts of such systems gain cohesion and independence as brain-side entrenching enables individuals to select currently available substantive information. By hypothesis, feedback mechanisms aid living systems to coordinate and, in humans, the feedback includes statistical facts about how wordings are used (viz. vocal tract actions, ways of speaking and skills in ’ways of speaking). Individuals gain a degree of control over motivated action that arises in social settings. This drives the entrenching of socially derived skills that link understanding with sensorimotor routines. Entrenching applies in learning to coordinate talk, playing an instrument or mastering ball skills. As brains self-construct, bodies monitor both actions and how their results affect social systems as agents perceive and bring forth differences that social roles. The results are habits or ways of acting – routines – that have become entrenched. Just as utterances are heard as utterances of something, ways of using a ball become purposeful (as they are used as passes, pitches or the like). In complex social worlds, people eventually discover games and, of course, come to believe that languages are based folk constructs such as ‘words’. Such beliefs ground meta-routines whereby activity (even football) falls under the partial control of verbal patterns. Persons gain detachment from the flow of lived behaviour that enables them, for example, to choose which language to speak or, indeed, which ball game to play. In talk learn to use skills to recall or talk about the past, imagine the future, or create models. As they do so, they discover the value of assessing what other people do as they anticipate what is likely to happen (or be thought). The approach is shown in Figure 3 below:

Figure 3. A radical embodied cognitive model. Interaction affects both neural systems and the organism’s doings: as it gains experience of social systems, the organism develops modes of action that entrench knowledge, which can then be observed in the carrying-out of socially derived routines.

9

1. Bodies use perception-action and social systems (not input) to ensure that brains selfconstruct through dealings with a partly social world. 2. Brains self-construct by setting parameters based on bodily use of physical and social discriminations (e.g. individuals come to perceive types such as wordings etc.) 3. The results influence social activity (people and routines co-develop as people monitor their own action; gradually their activity becomes tightly constrained by wordings). 4. Individuals develop (meta-) routines used in monitoring, assessing and accounting for linguistic and actional forms of activity. 5. Entrenching of substantive information grants a degree of socially sanctioned control over individually motivated action. As sensorimotor skills ground understanding (e.g., in playing an instrument, using a ball, or talking), individuals develop entrenched habits. Since action arises under distributed control, statistical correlations describe social routines. Lifelong learning prompts people to do entrenching. Neither brains nor minds need to ‘represent’ linguistic forms because verbal patterns (and other impersonal resources) constrain activity. To say, as Arnon (2015) does, that high frequency items ‘tend to be acquired earlier’ is merely to note that, all things being equal, frequent routines are often shared. Similarly, appeal to ‘error patterns’ shows that, in spite of reliability, overshoot occurs. People live in a world that is saturated with information, and in which frequency effects can therefore influence the activity of a perceiving (and situated) body. Chunks are social moves, that is, they are ways of making substantive information functional within interpersonal coordination. And so when they are automatized, they influence social systems and drive memory consolidation, which aids prediction, renders anticipation possible and grants a person access to ready-made resources for organizing activity – chunks evoke events, situations and ways of speaking. 6. Using language as a test case Cognitivist and radical embodied perspectives build on different methodologies. In computational models, a central controller (or module) uses statistical measures to construct causal structures. Conversely, on a radical embodied view, routines link statistical learning (entrenching) with a history of action-perception that drives selection: entrenchment shapes sensorimotor routines. Rejecting appeal to ‘input’, radicals stress how the CNS uses bodily feedback to set the neural criteria used in habit formation. As a result entrenched and functional information can influence future action. Indeed, this is why entrenchment measures (and, especially, conditional probabilities) can be so insightful.6 They apply to sets of cases (e.g. uses of X as opposed to Y) where habitual use is brought together with cases in which individuals make more deliberate use of routines. For the same reason, however, theorists may be misled into using entrenchment to explain actions based in, say, contingency or more deliberate modes of acting. The predictions of computationalist and radical embodied views differ in that the former privilege morpho-syntax and lexis and the latter stress that wordings are both pragmatic and part of (multiscalar) action. Given such differences, they offer contrasting views on why frequency effects apply differently to various kinds of language-activity. One finds, for example, a continuum between more or less frozen pragmatic facts (cf. the model by Schmid, 2014, reprinted as Figure 4 below). Whereas frozen types (on the top) readily disambiguate settings, those lower down require 6

Ways of controlling action are synergies (see, Anderson et al, 2012) or task specific devices (Bingham, 1988: Golonka, 2015).

10

cognitive work (i.e., they apply to settings of use that are hard to imagine). Perhaps this is why, as Schmid (2014, p. 256) notes, “literally all readers of earlier versions [of his paper]” have been moved to offer rich comments. Below, I pursue the importance of the observation by tracing entrenchment to, not representations, but social systems that constrain the changing routines that characterise speaking and acting.

Figure 4. Types of lexico-grammatical patterns arranged to show frozenness/variability (Schmid, 2014, p. 256)

Like proverbs, routine formulae tend to be rather frozen. However, while formulae are common, proverbs are relatively rare. If that is not surprising, why are both as frozen as ‘transparent phrases’? For, whereas laugh out loud or best before resonate with the new media and advertising, proverbs sound old-fashioned. Nor are puzzles restricted to frozen items. Looking down the table one can ask, say, why ‘collocations’ show flexibility. To pursue such issues, I contrast models of computational control with a radical view where control shifts between interlocking systems (cf. Table 1 below). The issue pivots on whether frequency effects (‘measures’) capture neural/mental representations and/or are indices that can be used to track how individuals are likely to draw on social expectations as they use routines. Table 1: Cognitivist and Radical embodied contrasts Computational control

Radical embodied control

Brains use input to extract distributional information. Bodies use action-perception and social systems (not input) in entrenching – a CNS self-constructs as bodies coordinate with the world. Brains identify linguistic patterns and construct entrenched stores.

Brains use physical and social discriminations to set parameters (e.g., individuals come to perceive types as categories, wordings etc.)

By linking entrenchment to speech production, brains The results influence social activity; people and are able to repeat the cycle. routines co-develop; wordings constrain activity. Entrenched mechanisms self-construct new forms of (predictive) mechanism and ways of producing

Individuals come to monitor, assess and account for routines. Gradually, they mesh linguistic and actional

11

linguistic output.

modes of perception.

Entrenched information enables a person to conform to local practices by replicating conformist ways of speaking/hearing (and, perhaps, innate devices for intention recognition).

The entrenching of substantive information grants socially sanctioned control over action. Sensorimotor skills sustain understanding (e.g., in using a ball, or talking), habit formation and entrenchment.

The computational model expects brains (or minds) to extract distributional information from formal ‘input’. Like software, they (1) identify ‘sames’; (2) record and categorize according to types; (3) record and register each type’s frequency; (4) use category tags and/or frequencies to cross-reference tagged-types. Leaving philosophical issues aside, let us, for argument’s sake, allow that brains use tagged aggregates of input strings as ‘constructions’ that guide speech production and transform perception and understanding by means of correction, including pragmatic correction. 7 To the extent that people accept determinate conventions/norms, such processes would enable striking conformity to social practices. Given that search engines work, roughly, as described (by using digital media), the view has prima facie plausibility. How, then, might it be applied to a range of pragmatic phenomena? On a computational view, pattern recognition picks out both ‘linguistic’ and social information. Using linguistico-social criteria, a brain must recognise a proverb as a proverb in order to restrict its use by ‘freezing’. So how can a brain identify proverbs? On a search engine analogy, it will link frequency information to non-frequency pattern.8But how? By computationalist hypothesis, brains must use substantive information (i.e., physical distinctions). However, since embodiment is ascribed no explanatory, the brain’s use of physical distinctions must rely on sentence-processing to calculate frequencies of different input strings or patterns. But, while brains might use innate (‘semantic/linguistic’) information as in Fodor’s (1975) view of concepts or Chomsky’s I-language (1986) to induce pragmatic types from speech data (using statistics), this conflates pragmatics with grammar. It grants brains automatized knowledge about proverbs, collocations, collostructions etc. However, given contrasts between the grammatical and the pragmatic, this can be seen as a mereological error. On the radical embodied view, people, not brains, identify verbal patterns. During development, they self-construct by using experience of, among other things, language in action. Children do not identify ‘sames’ and then apply them. Quite the contrary. For example, around their first birthdays, English speaking one-year-olds may wave ‘bye-bye’ and Italians say and/or wave ‘ciao’. Children orient to greetings (a ‘second-order’ type of behaviour) – before they learn verbal routines used to initiate, regulate, and otherwise enact such behaviour. The pragmatic nature of the skill arises in the delicacy of performance. Greetings are openings and closings – not, say, talk-initiators triggered after every five-second silence. As in all areas of pragmatics, they require a delicacy in timing, tone, discourse context, multi-modal accompaniment, and so on. For example, while the phrase best before is familiar, most readers would regard it as 7

The philosophical issues are well-known. Unless brains rely on innate categories for syntax, lexis and speech, they cannot classify and tag ‘sames’. In the study of language development, however, it seems that learning to talk depends on, not identifying ‘sames’, but learning by combining communication with a babbling history (and the entrenching of categorical perception). 8 It is circular to explain frequency by freezing (or vice versa). Further, frozen expressions can be unusual (e.g., a stitch in time saves nine) and many frequent expressions are frozen (e.g., good morning, best before). Plainly, something else must contribute to what is captured by appeal to the metaphor of ‘freezing’.

12

jocular where used in, say, in enquiries about health (e.g., best before breakfast). And, while some frozen items are vague (e.g., good day), precision characterises the skilled use of others (a stitch in time). Such delicacy cannot be ‘decoded’ by a brain, just because it is not encoded to begin with. So how is it generated? Even if one posits a language faculty, things seem baffling. Competence clarifies neither delicacy nor linguistic fluency because, by definition, pragmatic regularities link rules with meaning (whatever that is). The problem is a general one that appears, say, with collocations like run a bath and run a race. These produce comparable frequency measures; however, if a super-brain were miraculously able to calculate how frozen such expressions should be, it would still also have to ‘know’ their delicacy (compare run a Facebook page and run a car). To posit an autonomous system (whether theorised as a mental unit or an individual) that manages utterance-types obfuscates how a brain or mind can identify, classify, tag and cross-reference ‘forms’ to balance a degree of freezing with context-flexibility. Thus in relation to, say, utterancetypes such as ‘I don’t know’, the brain now needs to gauge that run a bath fits fewer circumstances. This is bizarre. So, if entrenching shapes mental representation of verbal types and their use, it is staggeringly complex. While there are many views on the topic (see, Jost & Christensen, this volume), the main point is that what brains (or minds) ‘store’ may relate only indirectly to measures of frequency. 9 Far from picking out neural categories, observers may pinpoint the relative probabilities of ways of acting/perceiving. Action and perception arise from shifts in distributed control. Social systems such as dyads, individuals or groups do not need stores of linguistic ‘sames’ or brains that tag, classify and crossreference verbal types. Indeed, to hold otherwise is to confuse measures of differences with what may well be necessary effects and mechanisms. In fact, self-organising brains enable bodies to select or render functional certain kinds of available substantive information. This capacity is as old as the minimal sensing of the first embrained bodies that evolved during the Cambrian; at that time, organisms developed mechanisms that grant sensitivity to only some distinctions that are available in the environment (Keijzer, 2015). Importantly, the capacities evolved in relation to tasks, settings and, later, operant social systems. Behaviour, and the functional information that underlies it, is task specific (see, Bingham, 1988; Golonka, 2015). If applied to linguistic usage, differences in frequency depend on the observer. For this reason, frequency measures pick out, not mechanisms, but widely used linguistic moves. Talk is like swordsmanship or playing a musical instrument in using a history of coordinated encounters between persons who rely on embrained bodies. Over time, using repetition with variation, people (and brains) select – that is, make functional – substantive information that serves strategic purposes. Learning to talk is thus a social performance that depends on how persons coordinate in a social and physical world. The outcome of such coordination is entrenching, and its outcomes allow embodied brains to sustain valued behavioural routines. Entrenching is thus bodily and, given its origins, its products can be co-opted to many purposes. For example, children can sensitise to substantive information in learning to babble or identify and make syllabic patterns (e.g., as described by measures of conditional probability). More complex entrenching arises as persons come to use wordings whose frequency effects can be used to reshape perception and action. For example, a human user of proverb can give an idiosyncratic twist to experience. He or she will rely on individual routines that, with habit, can become entrenched. Indeed, a history of such integration 9

In pursuing whether brains ‘predict’ combinatorial frequencies or probabilities, one faces the counter view that they enable personsin–situations to act in anticipatory ways. This general line of thought derives from Bernstein (1967) and, in modern guise, appears in Baber et al.’s (2014) work on how cutlery affects motor control when eating sweetcorn.

13

is the likely basis for many social and linguistic skills. Learning to talk and, later, to hear oneself talking re-tune the brain. In the second year, children begin to attend to wordings and, thus, learn from taking a ‘language stance’ (Cowley, 2011b). Over time, this transforms perception, social understanding and a child’s action repertoire. Memory consolidation grants individuals both rare and ‘grabby’10 habits as well as exposure to a range of routines. Frozen constructions come about as the habits (e.g., the early bird catches the worm) take on pragmatic associations, as Schmid (2014) notes, and as the routines (e.g., how do you do) are richly and flexibly combined with action granting them perceptual salience. During development, social systems gradually exert tighter constraints on the activities of their human parts and, by so doing, prompt strategic social action that restructures neural networks used in strategic social action. A striking nonlinear change arises in the second year when a child discovers new ways of linking speech with action. For Piaget, this is the ‘symbolic stage’ or, in distributed terms, it is when vocalisations come to be heard as utterances of something. Given neural flexibility and re-organization a child learns to hear wordings (by directing attention or taking a language stance). The view fits Anderson’s (2010) hypothesis of neural re-use and Dahaene’s (2009) view of neural recycling. What Menary (2014) calls ‘enculturation’ arises whenever persons use social systems to develop efficacious routines. Although having a homely basis, routines of these types permit flexibility. For example, frozen patterns can be used to illustrate delicacy. Like a baby’s bottom! Rather than regard freezing as automatization, it is likely to be a side effect of how individuals make strategic use of enculturation. The case can be made in relation to, say, hand-shaking: in English-speaking contexts, movement is often coordinated with saying ‘how do you do’. Making a good impression is likely to co-vary with the relative timing of hand-shaking, eye contact, distance and speech synchrony. In short, recurrent sensorimotor sequences (like the phonological and phonotactic patterns of a language or variety) are chunks that can be varied in careful use of social routines. In practice, of course, speech and action merge in whole-body coordination. Nonetheless, there are differences between mastering expressions and ‘doing things with social affordances’ such as hand-shaking. The pragmatic salience of social affordances thus contrasts with that of many symbolic associations. For now, however, it is enough that, far from being ‘represented’, frequency effects contribute to social routines. The claim suggests that substantive information that relate only very indirectly to linguistic models. In turning to how they draw on entrenching, I now consider artificial systems.

7. Entrenching in artificial agents When people master linguistic patterns, they come to use them with a delicacy that attests to both their cultural and functional value. Claims about pragmatic phenomena that invoke input/output models must therefore be rethought, I argue, since measures of behavioural frequency cannot identify neural structures. By hypothesis, radical embodiment prompts and guides the development of skills for social action. It does so by linking individual knowledge to the social routines that are used in coordinating with others. Roughly, a history of repeated task-driven interactions enables members of social systems to sensitize to common values, or equivalently, to make strategic use of substantive information (and thus verbal patterns). The functions of this sensitization – that is, this 10

For O’Regan (2011), certain sensations (e.g., red or the sound of a bell) are grabby. This seems likely to apply to many aspects of speech (e.g., rhythmic patterns) – especially where built into strategic action routines. This fits Baronchelli et al.’s (2010) finding that robot perception can be honed with reference to human physiology.

14

entrenchment – enable the same information to be re-used in novel contexts with varying effects. In the first place, it is highly unlikely that ‘forms’ shape the workings of self-organising networks. In principle, brains, people and social systems may use interactional events to establish ways of realizing values. Functional information may connect agents, bodies and interactional routines. By hypothesis, then, radical embodiment may prompt skills to develop by linking individual ‘knowledge’ to other agents’ social routines. Indeed, this fits the finding that entrenchment characterises the delicate use of pragmatic phenomena. While artificial agents cannot match this delicacy, their abilities give further reason to believe that linguistic behaviour is founded in how substantive information is rendered functional (for living systems). Although robots cannot use entrenchment, they use an artificial equivalent to entrenching. As a way of addressing the symbol grounding problem (see, Belpaeme et al. 2009), one can pursue how agents learn to map language-type ‘symbols’ onto ‘meanings’. In considering robots, I place ‘symbol’ and ‘meaning’ in quotation marks because I adopt an observer’s perspective and an observer’s semantics (a 3rd person view of symbol grounding leaves aside how systems master language). In seminal work, Luc Steels, Tony Belpaeme and others built robots that use formal types (symbols) to act ‘about’ the world. Steels and Belpaemes’ (2005) robots use coordination games to discriminate invariances that can be seen as colours (e.g. red, blue, green). Each robot develops categories that, given a history of correction, trigger symbol-based routines. Thus, using hardware, a well-structured environment, errors, and statistical learning, robots map symbols onto referents – and each robot develops its own ‘categories’. However, internal structures do not map onto the ‘measures’ that they make (as their neural networks come to be weighted). Indeed, such a design would not work because, strikingly, each robot perceives a different ‘world’. Although using ‘identical’ machines in a single environment, their partly overlapping categories drew on individual interaction history. Categories converge only when a ‘coordination game’ enables robots to build linguistic routines. Like humans, the agents use a perceptual history to perform (normative) action. Belpaeme and Bleys (2005) thus regard colour perception and language as uniting cultural constraints with universal (i.e., physical and pseudo-physiological) ones. Elsewhere, the approach is extended to grammars (Steels & de Beule, 2006) and how the physics of human vision changes robot categorization (Baronchelli et al., 2010). While not called entrenching, in these experiments, robots use substantive information to manage social interactions and act ‘about’ the world. They even form network-style representations that mimic memory consolidation, chunking and the automatic use of ‘form-meaning’ pairs. Robots use substantive information to build routines that are amenable to linguistic description. In treating this as entrenching, I stress two points. First, in coordination games, robot action is pragmatic. Second, machines lack the flexibility that, even in babies, is the mark of entrenchment (e.g., knowing when to say bye-bye). Such features arise because, in using statistics to build repertoires based on arbitrary symbol strings, robots do not, and cannot, recognize (phenomenal) ‘sames’. Rather, they are bound to rely on fixed digital values. For this reason, the models show the complexity of identifying symbol-use with entrenched ‘information’. On the plus side, symbols can be grounded in entrenching or how body-world coordination is used in ‘linguistic’ routines. However, since robots can make no use of functional information, the ‘symbols’ can mean nothing for robots. Artificial agents lack values, do not orient, and cannot use other means to refine acting/perceiving. For the same reason, they have been designed to depend on hardware and central control. This takes us to the downside. Since robot moves are meaningless for the agents themselves, human moves – and the results –are likely to be quite different.

15

Artificial agents can also use how substantive information is organised in a social environment to gain sensitivity to formal (or linguistic) patterns. Using a technique like a video game, one can model changes in a phonetic knowledge-base through a history of interactions. Indeed, one can mimic how child phonetic knowledge develops faster than that of adults. Stanford and Kenny (2013) use agent based modelling (ABM) to track how children and adult agents replicate the Northern vowel shift (Labov, 2007). In their model, agents travel between a virtual Chicago and St Louis on a highway where they interact by mimicking one another’s vowel-utterances. Where an interaction identifies phonetically similar vowels, the agent’s algorithm classifies utterances as ‘sames’. Importantly, the system records, not the fact of a match, but a digital version of a partner’s token. In the next encounter, two devices aggregate their records (viz. digitally close matches) in self-presenting and producing (or failing to produce) matches. In effect, control of vowel production is distributed across the population through feedback between interactions and individual exemplarbased memory. Given rich memory systems (i.e., an ability to detect substantive information), the agents’ vowels co-vary in line with large-scale sociolinguistic observations. Over time, the effects of utterance-frequency start to how up as changes in global vowel averages and do so, of course, without any change in mechanism. In contrast to autonomous players of coordination games, these agents show how phonology can self-sustain through agent interactions. Thus, while Labov ascribed children’s pronunciations to something like Chomsky’s (1965) language acquisition device, the model shows that interaction suffices to simulate entrenching. Quite simply, form-based memory (let alone a language acquisition device) is not needed for ‘learning’ about vowel change (for detail, see Cowley, 2016). Rather, an interaction history allows agents to ‘speak’ in ways that prefigure on-going phonetic change. People use pragmatic patterns with a delicacy that attests to their functional basis. Not surprisingly brains, people and social systems learn from agent-world interactions. Yet, the fact does not warrant the view that language enacts a new kind of information. Quite the contrary. In artificial agents, substantive information drives automatic, frequency-based sensitization that has no need to be reorganized and modified by task-oriented feedback. That is, artificial agents used substantive information for entrenching, but not for entrenchment. Given multi-scalar complexity, sensitivity to statistical (substantive) information enables robots to use ‘linguistic’ coordination to perceive just as, of course, ‘symbol use’ alters how they act. Though effects are bidirectional in that they link neural networks and physical stimuli, they use, not replication of ‘sames’ but normative constraints on statistically-based memory. Agents link physical and social distinctions to parameters that shape coordination. While necessarily automatic, they also consolidate memory, master chunks and learn from each other. Crucially, robots use substantive information to as if it was pragmatic. In Stanford and Kenny’s (2013) model, social regularities (‘phonetic knowledge’) use exemplars that are spread across agents – a form of collective memory enacts/tracks sound based on substantive information. Lacking any understanding, the agents use purely physical distinctions (substantive information), to connect internally stored values with social differences in (simulated) pronunciation: like Steels and Belpaemes’ robots, they mimic human-style behaviour. But robots differ from embrained systems that use entrenching to change how they act. Unlike robots, humans grant functional value to patterns and structures as a result of actively perceiving what happens as they coordinate. Human experience can make more or less deliberate use of repetition with variation in developing, say, social distinctions and ways of gauging/honing understanding. As entrenching is supplemented by entrenchment, therefore, people master routines. In pragmatics, by hypothesis, the CNS grants probability role in functional information. Though people use expressions as automatic chunks, they also act with deliberation. The same applies in

16

decision making where 50 years of work show that people do not understand how they come up with ways of ‘going on’. Not only do they use biases and heuristics but they generate insights and rationalisations: for James March (1999), therefore, decision making is interpretation rather than choice. The unlikely outcomes of interpretive acts generate functional information that, among other things, can be used to improve future decision-making.

8. Substantive information, entrenchment and linguistic distribution Viewed as input, utterance acts and texts are patterns used by brains to extract ‘information’. Some invoke a Universal Grammar or an Interaction Instinct to assert that the information is linguistic (or intentional). It is often blithely assumed that linguistic ‘sames’ are automatically recognized and systematized in language use (and, indeed, that this systematicity is somehow reflected in facts about ‘usage’). In theorising language as code-like, entrenching is misleadingly presented as if it generated mechanisms based on assembly of form-meaning pairings. The results, it is said, permit the CNS to use such patterns. The distributed perspective uses a simple deflationary argument to contest code models of this kind (see, Love, 2004; Kravchenko, 2007; Linell, 2005). Claiming that language uses two orders, or is symbiotic (Cowley, 2014), it seeks to explain ‘first-order’ activity in relation to, not wordings, but measurable neurophysiological events. This raises the substantive question: Since entrenchment measures establish facts, how do relations between types bear on the measured (i.e., various applications of ‘information’)? More concretely, how does finding that instances of X are more frequent than those of Y relate to social events during which Xs and Ys are uttered or, for that matter, encountered in textual form? For radical embodied cognitive science, neural activity draws on a history of socio-culturally situated practices (and second-order constructs) that are mediated by whole body coordination. Before turning to entrenchment measures, I turned to scientific views of ‘information’. Although information is always physical (or ‘substantive’), living systems build functional histories by selective orientation to certain information. In applying the argument to language, ‘words’ are found to have delicacy or, more precisely, people always grant wordings a particular sense. Since words are objects of belief, any scientific view must trace language to physical distinctions. One can thus learn from how robots simulate entrenching by using substantive information in a coordination game. They mimic language in bidirectional coupling between symbols and (simulated) perception. There is a second reason for stressing the multi-dimensionality of information. Since we lack model of how brains (or minds) identify linguistic ‘sames’, a better hypothesis is that, as behaviour, language uses physical information based in agent-environment interactions. To avoid the so-called mind-body problem, therefore, one can learn from the powers, and limits, of artificial agents. Stanford and Kenny (2013) show that sound change can arise at a population level in agents that do not represent ‘sames’. Rather, they distribute phonetic memory in space-time. Such observations make it likely that first-order language can be traced to a history of meshing action with perception. Given an agent’s sensitivity to physical distinctions, probabilistic and other aspects of the world can influence entrenching. However, of the cases of artificial agents also suggest that entrenched patterns may appear in populations (as second-order constructs) and not in brains or minds. This is consistent with treating language as part of human interactivity where wordings connect substantive information, bodily movements, and phonetic and visible gestures (e.g., Kinsbourne & Jordan, 2009; Thibault, 2011: Cowley, 2014). Since first-order language is coordination, sense-saturated experience enables people to identify the world’s functional potential . This can explain why measures of entrenchment index linguistic forms and, yet, use substantive

17

information. By implication, then, such measures identify agent routines that characterize whole populations without being realized as mechanisms in individual brains. Indeed, while the robots reconfigure ‘perceptual’ systems, Kenny and Stanford’s (2013) show that rich phonetic memory suffices for managing population level phonological change. While grounded in substantive information, language is irreducible to either statistics or bodily dynamics, just because the functions enabled by entrenching also use a history of living. The irreducibility of life to substantial information appears in the complex nonlinearity of how DNA is able to generate the proteins used in metabolism. Further, just as this organic coding is context sensitive, people show pragmatic delicacy in speech and action. Far from producing forms, humans use sensations to modify unfolding action, which they coordinate as they accomplish tasks and realize functions using social skills. Even if based in physical distinctions, language activity leads to the entrenching of functional information that connects bodily coordination with control over (one or more) CNS. As people speak, or look at texts, they orient to wordings – nonce events that evoke verbal types and personal histories of events (of the sort that observers formalise as linguistic and other knowledge). Given the degree to which wordings alter human capacities, it is striking that robots achieve so much without recourse to functional information. This is significant because, even if brains are quasi-robotic in their entrenching and general sensitivity to substantive information, people quite clearly use a functional history to control what they do. Human intransigence and linguistic delicacy both attest to some free will or, at least, ‘free won’t’ (Ramachandran, 1998). Crucially, people can deviate from automatized uses of consolidated memory and routines that drive what we regard as mindless listening, speaking and acting. Entrenching aids individuals in coordinating within social systems (i.e., in dyads, groups, communities), as they regulate valued aspects of interaction by using techniques to connect substantive information with experience of a lived world. Following the model of coordination in robots, it is safe to assume that physical distinctions underpin entrenched modes of action. Where functional for living systems, such distinctions are bidirectionally coupled with perception, affect and sentience. A history of linguistic activity can thus drive anticipation and the necessary habitmaking. By coming to use wordings infants can gain the meta-control of a person who displays attitudes by meshing vocalizations with other activity. Entrenching thus shapes skilled perception while using ongoing activity to gauge what may happen. Indeed, human language may have coevolved with development: infants manifestly self-transform by coming to hear wordings that are used in constructing species-specific skills. The ideas are summarized here: • •

Bodies-in-the-world use movement to select substantive information that is of potential value to an agent in a niche (and, in some species, in a built social world). Where sensation gives rise to the entrenchment of substantive information, this recalibrates perception-action by using feedback to give functional value to patterns that can be reused in agent-environment interactions (alone and within dyads and groups).

Linguistic entrenching attunes parties to the attunement of others (see Trevarthen, 1998). In Collier’s (2011) terms, dyads value some substantive information which thus becomes functional (in various ways). As it attains a value for the parties, infants come to regulate action and pragmatics and, by so doing, change their worlds – or if you like organism-environment relations. Eventually, they learn to take a language stance (Cowley, 2011b) and, by attending to wordings, come to anticipate ways of speaking. By treating first-order activity as verbal (and non-physical), a child develops skills in auto-control. She or he can pretend, anticipate and gain new skills in

18

manipulating others. Eventually, of course, a child may learn to use proverbs. Although based in a history of entrenching substantive information, the child gains skills in hearing and uttering these kinds of wordings. From the start, these have a particular sense for, not just the child, but also his or her companions. Accordingly, dyadic (and other) forms of social coordination fine-tune a child’s actions to the expectations of a social meshwork. By perceiving and evaluating the doings of others as well as the likely effects of self-behaviour, the self-organizing of the CNS grounds routines. While based in substantive information, entrenchment gains astonishing power as an infant becomes a person with a history, a narrator and a moral self. The child becomes encultured and, as in Dehaene’s (2009) account of how reading, retrains a CNS. This, I suggest, grounds the multiscalar applicability of entrenchment: measures reflect on, not just frequency differences but also an evolutionary and cultural history of meshing experience-based behavior with substantive information. Measures thus capture how individuals use distinctions to link sensed functional differences to routines. Indeed, they can pick out regularities pertaining to domains as diverse as, on the one hand, reaction time and individual learning and, on the other, development, history and language-culture co-evolution. It shows how indirectly entrenchment measures bear on phenomena that include memory consolidation, chunking and (de) automatization. 9. Conclusion Measures of entrenchment attest to the pragmatic and actional roots of human language. People use skilled action-perception to grant functional power to what begins as substantive information. By hypothesis, granting linguistic ‘meaning’ to physical differences underpins the flexibility of human behaviour. In sharp contrast to the computationalist view that neural mechanisms represent verbal types, I have traced entrenchment to how people structure and regulate action-perception that, in the case of language, prompts them to develop and modify techniques for understanding and living. Given a pragmatic basis, linguistic routines draw on social practice in shaping skill and judgement. As brains need no language faculty, people not only to speak and understand automatically but, of course, use functional information with delicacy. Entrenched routines link individual and social history such that people can be, at times, outrageous and, at others, insightful. Humans master activity in which wordings play a part: language connects individuals with collective resources – cultural practices (and languages) are distributed across populations. Entrenchment captures neither statistical measures nor entrenching. Once the distinctions are clear, one finds evidence that entrenchment is based in pragmatics and contrasts with how artificial agents mimic entrenching. Indeed, I argue that entrenching applies to all living systems, at all biological levels. Yet, the cellular processes of metabolism or the bidirectional coupling of ‘intelligent’ bacteria use more than substantive information. One should therefore be sceptical of treating culturally embedded primates that use language – living human beings – as input-output systems. No measure of frequency or frequency differences can, I claim, pick out a ‘mechanism’ or language system. To think otherwise is to take the mentalist view that brains (or minds) are (like) Von Neumann devices. Developing this argument, I claim that the more and less ‘frozen’ state of different pragmatic phenomena demonstrate the collective nature of language. The delicacy with which frozen phenomena can be used attests to the living human subject’s flexibility; language is thus partly constitutive of human existence. As (first-order) activity in which wordings play a part (Cowley, 2014), language enacts social and cultural events and can be ‘explained’ by neither processes occurring in individual brains nor code-like (second-order) language systems.

19

Entrenched routines are partly automatic. Even in monologue, people act under distributed control: in ‘saying something’, their wordings connect their own experience with that of the person to whom they are speaking. As activity, a language interface allows persons to share modes of attending while, of course, linking their own expectations with those often encountered within a community. In support of this view, I sketched how artificial agents use impersonal (digital) resources in processes sharing much with entrenching. While strikingly inflexible, such devices mimic and track ‘linguistic’ routines. What is missing, of course, is any means of granting functional value to physical distinctions. With this in mind, I argued that the sensorimotor basis of linguistic activity can ensure that brains select distinctions that, over time, are gradually entrenched. In a straightforward sense, then, explanations of sensorimotor activity are explanations for the makingfunctions of substantive information. In pragmatics, cultural aspects of language (second-order constructs) constrain how people are likely to speak and hear. Given entrenchment’s sensorimotor roots, however, people must actively create their own habits. They can come to use even proverbs with delicacy. Enculturation permits modes of acting that draw on projecting what can be said and done, as well as how it might be understood. Within cultural eco-systems, people learn about, say, engineering, cooking, or football – modes of activity that are partly defined by language and a human body’s limited possibilities for action-perception. Given a history of sensorimotor activity, physical distinctions take on functional within routines and variations on routines. People come to perform as engineers, cooks or footballers as entrenching connects an individual history, language experience and subjective modes of action and perception. By tracing entrenchment to routines, individual use of verbal patterns can draw on a history of sensorimotor activity. Coordinated activity and human action are thus dependent on substantive information, albeit indirectly. By using a history of entrenching, people gain skills that shape the many routine and flexible modes of acting that are, I suggest, the hallmark of a living human soul. Acknowledgements First of all I thank Hans-Joerg Schmid for his kind words of encouragement and valuable editorial advice. Second, I wish to thank John Collier for inspiring me to engage with how physical information can contribute to the domain of the living. Third, many thanks to Christian Mosbæk Johannessen for help with diagrams. Finally, I wish to thank Robert Port for his concise review and Matthew Harvey for his serious challenge to so many of the wordings I had proposed. As I result I have opened a big can of worms by choosing to contrast entrenching with how people act to coconstruct the entrenchment of routines. References Anderson, M.L. (2010). Neural reuse: A fundamental organizational principle of the brain. Behavioral and Brain Sciences, 33(4), 245-266. Anderson, M. L., Richardson, M. J., & Chemero, A. (2012). Eroding the Boundaries of Cognition: Implications of Embodiment1. Topics in cognitive science, 4(4), 717-730. Arnon, I. ( 2015). What can frequency effects tell us about the building blocks and mechanisms of language learning? Journal of Child Language, 42(2), 274-277. Baber, C., Parekh, M., & Cengiz, T. G. (2014). Tool use as distributed cognition: how tools help, hinder and define manual skill. Frontiers in psychology, 5. Baronchelli, A., Gong, T., Puglisi, A., & Loreto, V. (2010). Modeling the emergence of universality in color naming patterns. Proceedings of the National Academy of Sciences, 107(6), 24032407. Bateson, G. (1979). Mind and nature: a necessary unity. New York, NY: Dutton.

20

Belpaeme, T., & Bleys, J. (2005). Explaining universal color categories through a constrained acquisition process. Adaptive Behavior, 13(4), 293-310. Belpaeme, T., Cowley, S. J., & MacDorman, K. F. (Eds.). (2009). Symbol grounding. Amsterdam: John Benjamins Publishing. Bernstein, N.A. (1967). The co-ordination and regulation of movements. Oxford : Pergamon Press. Bennett, M.R. & Hacker, P. M. S. (2003). Philosophical foundations of neuroscience. Oxford: Blackwell Publishing. Bingham, G. P. (1988). Task-specific devices and the perceptual bottleneck. Human Movement Science, 7(2), 225-264. Blumenthal-Dramé, A. (2012). Entrenchment in usage-based theories: what corpus data do and do not reveal about the mind. Berlin etc.: Walter De Gruyter. Boden, M. (2008). Mind as machine: a history of cognitive science. Oxford: Oxford University Press. Bottineau, D. (2012). Remembering voice past: languaging as an embodied interactive cognitive technique. In Conference on Interdisciplinarity in Cognitive Science Research (pp. 194219). Moscow: Russian State University for the Humanities. Chemero, A. (2011). Radical embodied cognitive science. Cambridge, MA: MIT Press. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT press. Chomsky, N. (1986). Knowledge of language: Its nature, origin, and use. New York: Praeger. Clark, A. (2008). Supersizing the mind: embodiment, action, and cognitive extension. Oxford: Oxford University Press. Collier, J. (2011). Kinds of information in scientific use. TripleC 9(2): 295-304. Cowley, S.J. (1994). The place of prosody in Italian conversations. Unpublished PhD thesis, University of Cambridge. Cowley, S.J. 2007. The cognitive dynamics of distributed language. Language Sciences, 29(5), 575583. Cowley, S.J. (ed.) (2011a). Distributed language. Amsterdam: Benjamins. Cowley, S. J. (2011b). Taking a language stance. Ecological Psychology, 23(3), 185-209 Cowley, S.J. (2014). Linguistic embodiment and verbal constraints: human cognition and the scales of time. Frontiers in Cognitive Science, doi: 10.3389/fpsyg.2014.01085 Cowley, S.J. (2016). Cognition beyond the body: Using ABM to explore cultural ecosystems. In D. Secchi and M. Neumann (Eds.) Agent-based simulation of organizational behavior, pp. 4360. London: Springer. Cowley, S.J. & Gahrn-Andersen, R. (2015). Deflating autonomy: human interactivity in the emerging social world. Intellectica, 63: 49-63. Dehaene, S. (2009). Reading in the brain: the new science of how we read. Harmondsworth: Penguin. De Jaegher, H., & Di Paolo, E. (2007). Participatory sense-making. Phenomenology and the Cognitive Sciences, 6(4), 485-507. Fodor, J. A. (1975). The language of thought. Cambridge, MA: Harvard University Press. Fowler, C. A. (2010). Embodied, embedded language use. Ecological Psychology, 22(4), 286-303. Froese, T., Paolo, E. (2011). The enactive approach: theoretical sketches from cell to society. Pragmatics & Cognition, 19(1), 1-36. Golonka, S. (2015). Laws and conventions in language-related behaviors. Ecological Psychology, 27(3), 236-250. Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1), 335346. Harris, Z. (1998). Language and information. New York, NY: Columbia University Press

21

Harvey, M. I. (2015). Content in languaging: why radical enactivism is incompatible with representational theories of language. Language Sciences, 48, 90-129. Hutchins, E. (2014). The cultural ecosystem of human cognition. Philosophical Psychology, 27(1), 34-49. Hutto, D. D., & Myin, E. (2013). Radicalizing Enactivism: Basic Minds without Content. Cambridge, MA: MIT Press. Jost and Christensen (this volume) Keestra, M., & Cowley, S. J. (2009). Foundationalism and neuroscience; silence and language. Language Sciences, 31(4), 531-552. Keijzer, F. (2015). Moving and sensing without input and output: early nervous systems and the origins of the animal sensorimotor organization. Biology & Philosophy, 1-21. Kinsbourne, M., & Jordan, J. S. (2009). Embodied anticipation: A neurodevelopmental interpretation. Discourse Processes, 46(2-3), 103-126. Kravchenko, A. V. (2007). Essential properties of language, or, why language is not a code. Language Sciences, 29(5), 650-671. Labov, W. (2007). Transmission and diffusion. Language, (83), 344-387. Lassiter, C. (2015). Aristotle and distributed language: capacity, matter, structure, and languaging. Language Sciences, 53A: 8-20. Linell, P. (2004). The written language bias in linguistics: Its nature, origins and transformations. London: Routledge. Linell, P. (2009). Rethinking language, mind, and world dialogically: Interactional and contextual theories of human sense-making. Charlotte NC: Information Age Publishing. Love, N. (2004). Cognition and the language myth. Language Sciences, 26(6), 525-544. MacKay, D. M. (1969). Information, mechanism and meaning. Cambridge, MA: MIT Press. March, J. (1999). The pursuit of organizational intelligence. Oxford: Blackwell Publishers. Maturana, H. (1978.) Biology of language: the epistemology of reality. In G. Miller and E. Lenneberg (eds.) Psychology and the biology of language and thought, pp.28-62. New York: Academic Press. Menary, R. (2014). Neural plasticity, neuronal recycling and niche construction. Mind & Language, 29(3), 286-303. O’Regan, J. K. (2011). Why red doesn't sound like a bell: understanding the feel of consciousness. Oxford: Oxford University Press. Pickering, M. J., & Garrod, S. (2004). Toward a mechanistic psychology of dialogue. Behavioral and Brain Sciences, 27(2), 169-190. Pickering, M. J., & Garrod, S. (2013). An integrated theory of language production and comprehension. Behavioral and Brain Sciences, 36(4), 329-347. Rączaszek-Leonardi, J., & Kelso, J. S. (2008). Reconciling symbolic and dynamic aspects of language: Toward a dynamic psycholinguistics. New Ideas in Psychology, 26(2), 193-207. Ramachandran, V. (1998) Quoted in New Scientist, September 5th, p. 35. Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926-1928. Saffran, J. R. (2009). What is statistical learning, and what statistical learning is not. In S. Johnson (ed.), Neuroconstructivism: the new science of cognitive development (pp. 180-195). Oxford: Oxford University Press. Schmid, H. J. (2014) Lexico-grammatical patterns, pragmatic associations and discourse frequency. In T. Herbst, H.J. Schmid & S. Faulhaber (Eds.), Constructions – collocations – patterns (pp. 239-293). Berlin: De Gruyter.

22

Schmid, H.J. (this volume) Introduction to Entrenchment, memory and automaticity. The psychology of linguistic knowledge and language learning. Shapiro, L. (2010). Embodied cognition. London: Routledge. Sinclair, J. M., & Mauranen, A. (2006). Linear unit grammar: integrating speech and writing. Amsterdam: John Benjamins Publishing. Skinner, B. F. (1957). Verbal behavior. BF Skinner Foundation. Spurrett, D., & Cowley, S. J. (2004). How to do things without words: Infants, utterance-activity and distributed cognition. Language Sciences, 26(5), 443-466. Stanford, J. N., & Kenny, L. A. (2013). Revisiting transmission and diffusion: An agent-based model of vowel chain shifts across large communities. Language Variation and Change, 25(2), 119-153. Steels, L. (2008). The symbol grounding problem has been solved. so what’s next? In De Vega, M., Glenberg, A. M., & Graesser, A. C. (2008). Symbols and embodiment: Debates on meaning and cognition, pp.223-244. Oxford: Oxford University Press. Steels, L., & Belpaeme, T. (2005). Coordinating perceptually grounded categories through language: A case study for colour. Behavioral and Brain Sciences, 28(4), 469-488. Steels, L., & De Beule, J. (2006). Unify and merge in fluid construction grammar. In Symbol grounding and beyond (pp. 197-223). Berlin - Heidelberg: Springer. Steffensen, S. V. (2013). Human interactivity: problem-solving, solution-probing and verbal patterns in the wild. In S.J. Cowley & F. Vallee-Tourangeau (eds.) Cognition Beyond the Brain (pp. 195-221). London: Springer. Stewart, J. R., Gapenne, O., & Di Paolo, E. A. (Eds.). (2010). Enaction: toward a new paradigm for cognitive science. Cambridge, MA: MIT Press. Thibault, P. J. (2011). First-order languaging dynamics and second-order language: the distributed language view. Ecological Psychology, 23(3), 210-245. Tomasello, M. (2003). Constructing a language: a usage-based theory of language acquisition. Cambridge, MA: Harvard University Press. Trevarthen, C. (1998). The concept and foundations of infant intersubjectivity. In S. Braaten (ed.) Intersubjective Communication and Emotion in Early Ontogeny (pp. 15-46). Cambridge: Cambridge University Press. Wittgenstein, L.W. (1958). Philosophical Investigations (2nd Edition). Oxford. Blackwell.

23