Affordance Competition in Dialogue: the Case of Syntactic Universals

Affordance competition in dialogue: the case of syntactic universals

Eleni Gregoromichelaki Stergios Chatzikyriakidis University of Duesseldorf University of Gothenburg King’s College London stergios.chatzkyriakidis [email protected]

Arash Eshghi Julian Hough Heriot Watt University Queen Mary University of London [email protected] [email protected]

Christine Howes Ruth Kempson University of Gothenburg King’s College, London [email protected] [email protected]

Jieun Kiaer Matthew Purver Oxford University Queen Mary University of London [email protected] Jozefˇ Stefan Institute, Ljubljana [email protected]

Mehrnoosh Sadrzadeh Graham White University College London Queen Mary University of London [email protected] [email protected]

Abstract (Bruineberg et al., 2019; Paolo et al., 2018). At present, there is a considerable effort going on In this paper, we explore the idea that indepen- to provide accounts of so-called “higher cog- dently developed Dynamic Syntax accounts of dialogue and interaction fit well within the gen- nition” processes, like memory and imagina- eral approach of radical embodied and enac- tion, from an enactive perspective that eschews tive accounts of cognition (REEC). This ap- the necessity of representational content (see, proach enables a rethinking of the grounding e.g., Froese and Izquierdo (2018); Briglia et al. of linguistic universal constraints, specifically (2018); Hutto and Peeters (2018); Werning (2020); tree structure restrictions, as the outcome of Bruineberg et al. (2019); cf. Gallistel and King affordance competition, a general REEC so- (2009)). However, major trends within REEC still ciocognitive mechanism underpinning action remain underdeveloped with respect to the fine- selection. Given this subsumption, we argue such an approach opens up a whole new area grained details of the traditionally considered par of language-related dynamic systems research. excellence representational medium, NL. While broad outlines of an account have been provided, 1 Introduction what is still missing are more specific models of NL-related phenomena standardly conceived as Gregoromichelaki et al. (2019) claim that, in low-level syntactic and semantic issues. Yet in virtue of modelling natural languages (NLs) as ac- the absence of any attempt to incorporate NL actions licensing context-to-context transitions, Dy- tivity in all its complexity within the REEC ac- namic Syntax (DS) fits well within radical embod- count as a general nonmodular view of cognition, ied and enactive accounts of cognition (REEC) its claims remain seriously incomplete. This pa- this view, we take the grammar as an integral part per seeks to do three things: First we set out the of the sociocognitive environment into which hu- general embedding of the DS framework within a mans are enculturated via natural dialogue and REEC perspective, introducing the argument that a interaction. Thus the grammar, as it is being grammar formalism incorporating psycholinguis- developed and constantly revised, is not seen as tic (or “performance”) insights and an externalist qualitatively different from other constraints influ- (inter)action-based view of semantic significance encing the environment-engaging behavioural pat- not only provides a natural basis for a model of terns of individuals and groups in the achievement conversational dialogue but also opens up the pos- of their aims and goals. It is of course possible sibility of achieving an integration of current re- to reify these processes of linguistic engagement search in cognitive neuroscience and linguistic the- with the world and assign them abstract structure orising. Secondly, regarding the finer details, we articulated in representational terms, as in the stan- give a more detailed demonstration of how the DS dard grammatical models in theoretical linguis- framework, with its action-centric dynamic, has tics or models intended to explicitly teach linguis- the potential to capture straightforwardly familiar tic skills. Indeed, folk theories of language and morphosyntactic data for which static frameworks thought (at least in some societies) undeniably pos- had to add auxiliary hypotheses of considerable sess such conceptions of what language is, and complexity. But, on this basis, we can also go we do not doubt that such conceptualisations af- one step further towards explaining how the anal- fect linguistic behaviour. However, we believe that ysis of these phenomena carries over seamlessly such conceptions are neither basic, nor universal when they are embedded in the dynamics of on- (see, e.g. Linell, 2005). Instead, they constitute fur- going conversational dialogue. This is an advan- ther (meta)affordances available within particular tage of the dynamic approach which is not avail- social groups for engaging with the available first able to static frameworks or to individualistic ac- order NL resources and constraining more system- counts both presuming a competence/performance atically the relevant action possibilities. dichotomy. We then address a putative coun- Grammar – which for us includes what are terexample to the REEC general claim of the dis- standardly distinguished as syntax and seman- pensability of standard representational assump- tics/pragmatics – constrains human (inter)action tions and the primacy of action. It concerns a ro- by providing a source of normativity, of what bust structural constraint said to hold of all NLs is right or wrong, of what makes sense or not, (Kempson et al., 2016 a.o). We argue that such a and, hence, what is perceived as rewarding or restriction does not necessitate a representational not, relative to particular social practices. For explanation and can be seen instead as grounded in this reason, grammar is not a construct encapsu- a combination of socio-cognitive constraints and lated within an individual brain or mind. Follow- general properties of dynamical systems (see, e.g., ing Wittgenstein’s well-known arguments against Silberstein and Chemero, 2012) hence fully com- the existence of “private languages” and against mensurate with the view that all aspects of cog- an interpretative conception of rule-following (see, nition are grounded in action without mandatory e.g., Wittgenstein, 1953), we assume that norma- invocation of brain-internal mental representation. tive constraints apply within a public domain of expression (even in cases where we privately re- 2 Rethinking the nature of NL grammars hearse responses to a simulation of such interac- Dynamic Syntax, as set out in Kempson et al. tional challenges in the public domain). (2001); Cann et al. (2005) and much other work Concomitantly, in the REEC affordance liter- since, is a grammar architecture whose core ba- ature, a distinction is made between individual sis is incremental integration of the contribution abilities which are non-representational capaci- of word sequences within a surrounding landscape ties allowing individuals to perceive and pick up of affordances. Affordances, under this perspec- what opportunities and restrictions are available in tive, are relations between possibilities for action their sociomaterial environment and affordances provided by the environment (including the social which are potential interactions with the environ- milieu) and agent abilities available in a ‘form of ment available within practices or ‘forms of life’ life’ (Rietveld and Kiverstein, 2014 a.o.). Within independently of any particular individual agent. Grammars, in our view, operate at the level of ture modelling the landscape of salient affordances regulating actions within practices, both linguis- as a Directed Acyclic Graph (DAG), see Fig 1. tic and non-verbal, and, for this reason, are part DAG nodes (Interaction Control States, ICS) are of the public landscape of affordances available in turn structured states modelling hubs leading to interacting agents in each particular instance further embedded paths of potential interpreta- of engagement with the environment. Whether tions/productions. Interpretation or production re- agents interact solely with the physical environ- lies on the process of actualising selected paths ment or with other agents, their actions (which among the potential provided within local graph not only reveal but actually constitute their concep- structures in the form of trees. tualisations) are enabled and restricted by norma- In formal terms, such path traversals at the in- tive constraints. Such constraints which are non- termediate level of DAG nodes are enacted and representational ways regulating behavioral pat- constrained by means of employing in the con- terns (systems of habits) are imposed by the var- straint language specialised modal operators that ious cultural groups agents inhabit or wish to as- restrict the trajectories of parsing/generation to sociate with. Due to membership in various such incrementally emerging tree-shaped routes. The cultural groups, in each occasion of engagement, emergent licensed trajectories are models of sets agents’ actions in turn “enact” and hence modify of constraints articulated within the logic of finite or enrich the normative constraints available in the trees (LOFT, Blackburn and Meyer-Viol, 1994). current practice. Each node of a tree-structure model is in turn in- Words, both as forms and meanings, and syn- habited by graphs resembling feature structures tactic constructions are established patterns of ac- whose attributes are modelled by modal oper- tions that can be fitted in across various ‘lan- ators bearing linguistically relevant labels like guage games’ to enact the nature of the current type, hTyi, treenode address, hTni, and hFoi activity (‘semantics’) subject to normative judge- for content. In current versions of DS, con- ments emanating from sources outside the acting tent values are Record Types (henceforth RT) agent (despite the fact that sophisticated agents originating in Type Theory with Records (TTR, can simulate internally such externally sourced Cooper and Ginzburg, 2015) under construction judgements). This results in a system constantly in (Eshghi et al., 2013; Hough, 2015; Purver et al., flux (Cooper, 2012) but with enough emergent sta- 2011). RTs’ inherent underspecification fits well bility in each particular occasion to underpin agent with the requirement to model the potential for coordination in the service of various purposes. indefinite enrichment across various dimensions. In accordance with these assumptions, the DS RTs (standing for “concepts”) can in turn be con- syntactic engine, including the lexicon, is artic- ceived as mini-grammars of the DS kind articulat- ulated in terms of goal-driven actions (see also ing affordances for engagement with aspects of the van Benthem, 2011) accomplished either by giv- environment (Gregoromichelaki et al., 2019), or ing rise to expectations of further action opportuni- potential for interaction with others (Eshghi et al., ties, by exploiting contextual resources, or by be- 2017a; Eshghi and Lemon, 2014). In a parallel ef- ing abandoned as unviable in view of more com- fort, Sadrzadeh et al. (2018) show how combining petitive alternatives. Thus words, syntax, and a Vector Space Semantics methodology with DS morphology are all modelled as affordances, op- can model incremental construction of ad hoc con- portunities for (inter-)action, produced and recog- cepts to resolve issues of ambiguity and underspec- nised by interlocutors to perform step-by-step ification, allowing for pervasive variability in NL coordinated mappings from perceivable stimuli interpretation. (phonological strings) to concept-constituting ac- In both these versions of DS, DAGs map out tion patterns (routines, macros) or vice-versa the potential transitions globally available for se- (Gregoromichelaki et al., 2019). lection in a particular context of interaction while, The substance of the DS framework is given by locally, at each DAG node (ICS), the potential a specialised dynamic modal logic (PDL, Proposi- constructions and transitions are constrained by tional Dynamic Logic) whose state-transition lan- LOFT axioms that ground tree axiomatisation guage describes a process of gradual unfolding principles. As DS is a model-theoretic formal- of a diagrammatically laid out relational struc- ism (Pullum and Scholz, 2001), all the inferential activity modelling incremental parsing/production pendencies across more than one agent. is defined at the level of constraints articulated From this viewpoint, in dialogue, what can be through the vocabulary of the relevant logics. The considered as “complete sentences” may emerge transformation of such PDL and LOFT descrip- through a sequence of fragmentary contributions, tions is accomplished via general and lexicon- with each participant adding some fragment to a driven mappings. These model patterns of estab- partial structure. lished and relatively stable sequences of basic actions (macros). Such macros effect the progres- (1) Alex: There. To the left. sive projection of content, i.e., further opportuni- Hugh: Got it . . . ties for action, for emergent trajectories follow- Hugh: but it’s heavy ing the time-linear order of presentation of a lin- Eliot: without the lid? guistic string. The paths traversed by means of Hugh: if you fancy burning yourself. the LOFT modal operators, reflect the constrained In dialogue that is, employing “fragments” and structuring of physical and cognitive actions that switching of roles between speaking and hearing take place during linguistic processing. Some of is widespread because not only the physical con- the outcomes of such actions can be reified as text but also the interlocutors’ actions and attitudes predicate-argument structures by assuming an ex- contribute to the construction of meaning and the ternal, God’s eye, non-modal-logic point of view achievement of (joint) goals. In such circum- that abstracts from the process of inducing and stances, incomplete linguistic dependencies can traversing such structures. This is a useful and of- act as prompts soliciting the interlocutor to pro- ten needed perspective, both for theory construc- vide not only physical objects but also utterance tion and for practical human purposes like teach- continuations: ing, metalinguistic reflection and many others but, in our view, it should not obscure the most ba- (2) A: The the ... sic subpersonal level of processing from which it B: Here. emerges. A: Thanks. And get a a B: mattock. For breaking up clods of earth 2.1 Rethinking syntax as process [BNC (modified)] Bringing this framework to bear on current views within REEC, Gregoromichelaki et al. (2019) ar- (3) Jack: I just returned gue that DS is compatible with a view of NLs as ac- Kathy: from tivities, “languaging”, rather than manipulation of Jack: Finland [Lerner 2004] internal knowledge structures that define arbitrary mappings from sound to propositional symbolic Each such contribution can add unproblematically representations. This is because DS articulates to whatever partial structure has been set out so far, state-transition mechanisms relative to an ever- irrespective of whether or not what precedes it is a evolving interaction context, meshing with nonver- full sentence. The effect is one of a rich potential bal actions and, for this reason, without needing for interactivity between dialogue participants. to only license fixed propositional contents asso- Moreover, fine-grained morphosyntactic and se- ciated with sentence strings. Crucial in this per- mantic constraints, with seemingly no functional spective is incorporating in the formal framework motivation, apply across the exchange of turns and the social normativity expressed by the grammar. sharing of utterances so that the initiation of a turn The native incrementality and action-orientation by one participant constrains the options available of DS, as demonstrated by Kempson et al. (2016) for the next. For example, in the Romance lan- amongst others, is well-suited to implement the guages, Modern Greek, and other language fami- fine-grained modelling of conversational dialogue lies, a well-known morphosyntactic restriction on 1 dynamics. In particular, we are able to capture the clitic clusters is the Person Case Constraint (PCC; wholly fluent manner in which co-participants in Perlmutter, 1970; Bonet i Alsina, 1991). In its a dialogue switch roles, share utterance responsi- most general standard variant, the PCC states that bility, or effect repair, as the formalism permits 1Clitic clusterings are cases in which weakened pronouns the distribution of all syntactic and semantic de- occur in a fixed sequence characteristically immediately pre- ceding or following the verb. “who” “hugged”

who ..., pointermove, ... hugged T2 T5 T7 T13 node [.5] ... make LINK UNFIXED ed-node [.8] make make SUBJECT -node [.3] who T0 make T3 T9 T11 T10 abort L ... INK ed-node [.2]

T4 abort abort T12 ... Figure 1: DAG: Who hugged

?T y(t),Tn(a) ...*Adjunction... ?T y(t),Tn(a), ♦ −→ h↑∗iTn(a), ♦

Figure 2: Introducing an ‘unﬁxed’ node

?T y(t),Tn(a) ?T y(t) ...who... −→ WH : e, ♦ h↑∗iTn(a), ♦ h↑∗iTn(0)

Figure 3: who-induced actions a dative clitic cannot co-occur with a 1st/2nd per- (8) A: Irthe xtes o Giorgos ke son accusative clitic: came yesterday the George and tu... (4) A: Irthe xtes o Giorgos ke himCL−DAT A: came yesterday the George and ‘ A: Yesterday George came and to him ... tis to edose herCL−DAT itCL−ACC gave B: *se edixe?

‘A: Yesterday George came and he gave it to her’ youCL−ACC showed? B: ‘he showed you to him?’ (5) A: *O Giorgos tu se

the George himCL−DAT youCL−ACC edixe. The attempted continuation of the clitic cluster in showed the completion + confirmation request in (8) above ‘A: George showed you to him’ [Greek] is unacceptable because the PCC is violated. This shows that the grammar coordinates simultane- (6) A: *Le me ha dado ously and seamlessly the complementary and mul- himCL−DAT meCL−ACC has given tifunctional actions of multiple agents while they ‘A: S/He has given me to him.’ [Spanish] are involved in interaction without imposing abstract licensing restrictions that concern single in- Notably, such morphosyntactic restrictions also dividuals acting on their own. For example, there apply across turns: is no requirement imposed by the grammar that (7) A: Irthe xtes o Giorgos ke turns need to consist of (overt or covert) complete came yesterday the George and sentences/propositions to be licensed. tis... However, as can be seen here, e.g. in (2) herCL−DAT where the appropriate word is sought, the set of A: ‘Yesterday George came and to her ... affordances in a context are perceived differen- tially by each agent or group of agents depend- B: to edixe? ing on their level of attunement to the relevant itCL−ACC show practices that constitute the source of such affor- B: ‘he showed it to her ?’ dances, their skills, their attentional state, and previous experiences. Available affordances will also be partitioned and probabilistically ranked argument positions eventually so its contribution, depending on the current concerns and purposes a metavariable notated as WH, needs to be held of the agents involved. Agent-relevant affor- in memory until a suitable place is found for it dances are called ‘solicitations’, i.e., a subset of later in the process. Therefore, the identification of the available affordances that stand out with re- the node that accommodates the macro associated spect to particular agents (Rietveld et al., 2018; with who does not specify immediate dominance Bruineberg et al., 2019). We extend this notion relations with respect to the root node, as shown to apply to groups of agents. Such demarcated in the illustration, Fig. 2-3. This is indicated by group solicitations within so-called ‘fields of affor- the dashed line. The underspecified relation is ex- dances’ are modelled in DS by the time-linear un- pressed via the modal operator h↑∗i appearing in folding of the DAG during conversational interac- front of the root node identifier, Tn(a) specifying tion, see Fig. 1 (see also Sato, 2011; Eshghi et al., that the relation between the two nodes is one of 2013; Hough and Purver, 2012).2 (non-immediate) dominance3. Focussing attention at the local level of DAG As can be seen in the topmost path of the DAG, nodes (ICSs), the first step is to consider the eventually the verb hug will be processed and mapping between a string of words, in DS, a its subject requirement will enforce the contribu- sequence of triggers for macros, and the DAG- tion of who to be specified as such. In DS this transformations that it induces. The specialised means that the unfixed node will now become PDL backbone of DS operates by means of pre- fully specified with respect to all dominance re- diction of available open paths of development lations. Once such a DAG path has been suc- constrained by so-called requirements (these are cessfully traversed, the state reached will record indicated by ? appearing in front of any avail- a tree-structured topology of information with no able grammatical action). The constraints thus in- words left to be parsed/generated any more (see duced by the grammar (and other contextual fac- Kempson et al., 2001, 2016; Hough, 2015, for de- tors) define a range of goals for the next steps tails). of processing. Agents attuned to the grammati- 2.1.1 A universal constraint? cal practices available in a particular context can pursue the most relevant goal defined in this con- The constraint on development of such struc- text guided by a process of affordance competition tures which is then proposed for all stages of ev- (Cisek and Thura, 2019). For example, a potential ery DS process is that there be only one copy starting point of such a process can be a node state of any node type at each individual stage of a as displayed in Fig 2. Here we find the indica- particular DAG path. Indeed, it has been ar- tor of the current focus of attention, the pointer gued in major support of the framework that ♦, and a prediction reflecting the expectation that the abandonment of universal constraints on a proposition (?Ty(t)) can be developed. This ex- structure urged by some (Christiansen and Chater, pectation will lead to many alternative predicted 2008; Evans and Levinson, 2009; Bybee, 2010; paths of achieving this development that can be Haspelmath, 2020) can be reversed by the shift to displayed as shown in Fig 1. In Fig 2 we pursue a dynamic system that operates with partial trees one of those paths, the topmost path in the DAG, and their incremental introduction. This is because which is an option made available by the English universal structural constraints can be shown to grammar: a macro called *Adjunction can intro- hold due to the inferential principles of LOFT, duce a radically structurally underspecified (‘un- the constraint articulation basis of describing tree fixed’) node predicted to be needed to accommo- graph transitions in DS. So the proposed restric- date the processing of a content-underspecified el- tion is a general constraint which follows from the ement like who as seen in Fig 3. The latter pro- logic of partial tree construction: each node in a cessed first in a sentence can end up in multiple partial structure at any stage of the development 3 2In order to simplify presentation, the available macros h↑∗i is the modal operator expressing an immediate dom- have been significantly condensed and schematically men- inance relation, h↑∗i is the operator which, using the Kleene tioned through the more central effects they induce; number * operator, expresses the weaker dominance relation. A yet values indicate toy probability rankings for particular paths; more restricted variant is also defined requiring resolution ellipsis (...) indicates that multiple steps have been omitted of the under-specified dominance relation within a single se- as they have been judged as irrelevant to the point we wish to quence of functor relations, hence within a local predicate- make. argument structure process has a unique identity with respect to other (9) Masami-ni prezento-o Kiyomi-ga nodes in that structure; and this restriction ap- MasamiDAT prezentoACC KiyomiNOM plies equally to structurally underspecified nodes katta bought [Japanese] whose precise position in the emergent structure is Kiyomi bought a present for Masami yet to be established. Conceptually, this is a de- sirable property of the system because we assume (10) Masami-ni prezento-o John-ga that the information accumulated on a node is al- MasamiDAT prezentoACC JohnNOM ways underspecified and can be enriched at vari- Kiyomi-ga katta-to omotteiru ous stages of processing, even long after a node KiyomiNOM bought-COMP believed has been initiated. The restriction on a single John thought Kiyomi bought Masami a present uniquely identifying node address ensures that in- This flexibility of ordering and the left periph- formation encountered later in the processing of a ery restriction is characteristic of the verb-final string can have access to some particular already language pattern, by report the most common lan- existing node, even if such a node is radically un- guage pattern among natural languages (Dryer, derspecified in terms of either position or other 2013). But it also occurs in languages with more properties. Consequently, once a node ’unfixed’ flexible word orders like Modern Greek where the relative to particular domain has been introduced verb along with the arguments can occur in any or- no other node can claim a matching underspeci- der. At first sight, it might appear that this freedom fied identification. The effect is a “no-copy” prin- of ordering and piling up of coarguments in the left ciple for any node type. However, as we are go- periphery will be as problematic for DS as it is for ing to see presently sometimes this restriction gets other frameworks (see e.g. Koizumi (2000)). If se- in the way, especially with free word-order lan- quences of noun phrases may occur in any order, guages that can exploit word order for information or, in the embedded case, the left periphery can structuring purposes. In attempting to circumvent host any number of arguments, there would appear the effects of this restriction, linguistic patterns to be, in either case, a violation of the “no-more- might appear that might give the impression of than-one” underspecified relation constraint. functionally unmotivated arbitrariness and taken However, as argued in (Kempson and Kiaer, as evidence for the “autonomy of syntax” and the 2010), within DS, there is a simple explanation of particularity of the “language faculty”. However, this puzzle. The DS system is able to distinguish as we are going to argue such patterns do not pro- degrees of underspecification across all grammat- vide any such evidence. Instead, the confluence of ical notions, from syntactic to semantic domains. various other affordances and constraints explain In the case of node position, fine-grained variants the observed patterns. of ‘unfixed’ relations are defined depending on the 2.2 Syntactic and morphosyntactic puzzles domain within which the underspecification is con- resolved fined. One such variant is the case of a node specified solely as dominated by the root node, so an This dynamic of processing incorporating pars- underspecified relation within that global domain. ing and generation under the same formalism has Another variant that naturally arises due to the lo- been applied to a range of puzzles concerning syn- cality defined by a verb and its arguments is un- tactic and semantic/pragmatic phenomena as well derspecification relative to a more restricted co- as their interaction. For example, in verb-final argumental domain. And naturally there are tran- languages, despite very free ordering of all con- sitive implicative relations among such definitions stituents other than the verb, long-distance depen- with the global less specified transitions being able dencies are subject to the strict constraint: more to be updated into the more locally-constrained than one long-distance “movement” relation is li- ones (Cann et al., 2005). Interaction with other censed only if the items “dislocated” to the left pe- simultaneous grammatical constraints now solves riphery are co-arguments of some embedded verb. the puzzle of free ordering. However, locally such co-arguments retain flexi- The solution is provided by considering the role bility of ordering with respect to each other as in of case morphology. Case in DS has a crucial simple clauses (Pritchett, 1992; Koizumi, 2000): processing and semantic purpose, namely, the introduction and restriction of further possibilities of action, instead of just being seen as a feature- language with no presumption of any allegedly matching reflex of some overarching syntactic universal syntactic structuring that has to be im- structure. What this means for free constituent posed contrary to the observable facts. order in languages such as Japanese and Modern The effects of case are involved again. Dative cl- Greek is that the simple monoclausal sequence in- itics are in general notoriously underspecified. In volves the iterated process of introducing an un- Romance, this underspecified nature of the dative fixed relation followed by an immediate update can be traced back to Latin, where dative mark- step to a specific fixed position provided by pro- ing is characterised as ambiguous (van Hoeke, cessing the case-marking. Once this latter step has 1996) in the sense that, in and of itself, dative been taken, it immediately enables the licensing does not determine a fixed hierarchical position again of a new underspecified relation this time in the structure. Similar diachronic considera- concerning a newly introduced node. Moreover, tions also apply for other languages that show the this process can also apply within the province same PCC effects, e.g. Greek (Chatzikyriakidis, of a currently unfixed node with the more embed- 2010; Chatzikyriakidis and Kempson, 2011). In ded underspecification variant concerning a more DS terms now, given this underspecification of the local domain than the one defined by the overar- contribution of the dative, the plausible assump- ching underspecified node. However, what is not tion to make is that dative clitics are processed on possible in such a case is that the processing of an unfixed node, so as to allow variable interpreta- arguments from various clauses will intermingle. tions of the dative-marked argument depending on This is because the only underspecification that the clausal context. Furthermore, 1st and 2nd per- is available in such circumstances is the more lo- son accusative clitics in Spanish and many other cal variety which can only be updated within a Romance languages are syncretised with the da- local domain. Thus the composite effect of ex- tive, i.e. the same morphological form is used for cluding long-distance dependencies from multiple both the 1st and 2nd person accusative, as well as sources but with local ordering flexibility is then the 1st and 2nd person dative clitics. For this rea- directly predicted through the licensed feeding son, 1st and 2nd person accusative clitics can also relation between the two differently constrained be considered as underspecified and, under DS as- types of underspecified relation. This modelling of sumptions, as processed on an unfixed node too. local word order freedom with immediate update Under these underspecification of form assump- despite the lack of a predicate in verb-final lan- tions and given the “no copy” constraint, we can guages is notably confirmed by experimental work predict that any combination of a dative clitic with establishing the incremental nature of Japanese a 1st or 2nd person accusative clitic will be disal- sentence processing by Witzel and Witzel (2016) lowed. This is because, as we said, two distinct who demonstrate experimentally that both struc- unfixed nodes of the same type cannot be intro- tural and semantic reference judgements are incre- duced in a single domain. As a result, the order- mentally made before the verb is processed.4 ing Dative-1st/2nd person in, e.g., Modern Greek, Moreover, surprising confirmation of the restric- will appear as unacceptable due to incompatible tion on underspecified dependencies and its poten- information coming from the two types of clitics tial evasion comes from its applicability to the su- accumulating on a single node, see (8). On the perficially wholly unrelated morphosyntactic phe- other hand, 3rd person accusative clitics are not nomenon of clitic clustering. As we saw earlier syncretized with the dative and, furthermore, they in (4)-(8) due to the PCC, arbitrary lexical gaps in are always interpreted as direct objects. In DS clustering paradigms are attested. The restriction terms, this means that they are processed within a on unfixed nodes, namely the fact that no more fully specified structure, rather than on an unfixed than one treenode with the same address is pos- node. Thus, the “no copy” constraint does not in- sible, gives us a natural explanation of the PCC terfere with the processing of combinations of a effects because it forces us to look at the observ- dative and a 3rd person accusative clitic, see (7).5 able properties, the morphology, and the interpre- 5There are a number of variants of the PCC, e.g. the weak tation of clitic morphemes within each particular PCC version, which allows combinations of 1st and 2nd per-

4 son pronoun clitics, but this variability has been shown to See also (Kiaer, 2007, to appear) for extensive discussion be afforded by the formal machinery of the system as well and experimental evidence from Korean. (Chatzikyriakidis and Kempson, 2011). The significance of these results concerning 3 NL within the REEC perspective syntactic and morphosyntactic restrictions is that There is a range of approaches within REEC usu- these are observable low-level facts which are usu- ally grounded within dynamical systems models ally taken as sui-generis and arbitrary without any of socio-cognitive phenomena (Chemero, 2009). grounding in human interaction, meaning, or pro- Most of those espouse non-representational ac- cessing, hence apparently warranting complex and counts of perception and action, so-called “lower unavoidable stipulation in distinct components of cognition” or “basic minds”. However, even the grammar. Current individualistic linguistic advocates of radical enactive perspectives stop theories justify this split on the basis of notions short from extending this approach to “higher like competence-performance, modularity (Fodor, order” cognition, especially language (see e.g. 1983), and Marrian computational vs algorithmic Clark, 2016; Hutto and Myin, 2012). Moreover, level distinctions (Marr, 1982; Steedman, 2000; even accounts that aim to develop NL models Kobele, 2012). Within the DS framework, in compatible with dynamical and complex systems contrast, these patterns along with their interac- approaches like connectionism, due to the in- tional effects (see e.g. (8)) are seen to fall out dividualistic perspective they adopt, have sug- in virtue of modelling syntax as the progressive gested occasionally that neural network implemen- shaping of the landscape of affordances for inter- tations strengthen the competence/performance pretation and production in the everyday coordi- distinction and support the assumption of sym- nation of action. Such normative morphosyntac- bolic representations, albeit of an emergent nature tic constraints are explainable within such a gram- (Prince and Smolensky, 1997). mar because they can be seen as historically sed- Against this view, a lot of current REEC imented practices routinising the most frequently research (Rietveld et al., 2018; Bruineberg et al., taken up processing paths (macros), whether as 2019; Paolo et al., 2018) aims to integrate all lev- a parser or a producer. Such routinisations are els of cognitive activity within the enactive ap- not qualitatively distinct from the other grammat- proach. And DS sides with this approach in that ical or lexical constraints given the flexibility af- the constraints defined through the logic under- forded by the grammar (as modelled by the var- lying the framework constitute additional affor- ious paths within the DAG) even within single dances available in the sociomaterial environment instances of interaction. Conceptualised as nor- of human interactions and are not confined within mative constraints, such macros are necessarily individual brain structures (Gregoromichelaki, independent from individual NL users and con- 2018). stitute affordances within the processing environ- DS constraints concern the process of compre- ment (the ‘form of life’). However, in order to be hension/production in a social context as aspects taken up appropriately, the processing agent needs of general perception/action mechanisms imple- to be attuned to their potential, i.e., they need to ap- mented via sensorimotor feedback loops with the pear as ‘solicitations’ to the agents involved. But environment that do not necessarily engage rep- the possibility exists that such affordances are po- resentational constructs (even though reification tentially inadequately grasped either because the of such subprocesses and representational abstrac- agents do not possess the skills (attunement, ‘abil- tion is also considered possible). ities’) required or because the concerns of individ- Under such a view, all grammatical dependen- ual agents and groups prioritise other normative cies are able to function as word-by-word in- considerations in the competition among which af- cremental coordinating devices, i.e., affordances, fordances to pursue. Hence the constant poten- for either interlocutor. They can do that ei- tial for innovation and change as well as flexibil- ther by providing forward momentum in the con- ity and adaptability resulting from novel combi- versation. For example, see (2) earlier for a nations of affordances pursued. Nevertheless, the determiner-noun dependency whose satisfaction availability of such macros in a practice-sharing completes an already initiated statement but also ‘form of life’ allows agent coordination in particu- provides a platform for an enrichment and fur- lar interactional episodes due to their operation as ther specification of the expressed content. Sim- manifestly joint relevant affordances (joint ‘solici- ilarly in (3) a preposition-prepositional comple- tations’). ment dependency functions as a query speech act indicator to induce the interlocutor to pro- pending affordance for specification that has been vide the completion as an answer to the question introduced by the verbless utterance. (Gregoromichelaki et al., 2011, 2013a). Moreover, In contrast, in (12)-(13), the previous utterance, grammatical constraints can disrupt what is ex- which appears to be a fully-specified independent pressible and in what manner by the interlocu- sentence and proposition, is taken over by the inter- tor as in (8). In all these instances, grammati- locutor at the next processing stage. As a result, it cal dependencies can function coordinatively suf- becomes embedded under a verbum dicendi thus ficiently and efficiently irrespective of whether or losing its supposed independence as a “complete not coherent units like the standard representa- thought” and becoming instead part of a report tional constructs of ‘sentences’ or ‘propositions’ that uses the original speaker themselves as the are ever derived or not: ‘animator’ (Goffman, 1981) of the message now subsumed under the authorial control of the sec- (11) Hester: It’s for me. ond speaker but remaining, nevertheless, a joint Mrs Elton, the landlady: And Mr Page? construction. Hence the effectiveness as a hos- Hester: is not my husband. But I would tile move of this construction in English. But to rather you continue to think of me as Mrs. achieve the modelling of this effect it is essen- Page. [The Deep Blue Sea (film)]6 tial that we do not assume that there is a copy (12) Jem: Mary, whatever it is you think you of the original utterance embedded under the re- know you mustn’t speak of it, not if you porting verb and under the second speaker’s voice. want to stay safe. Instead, the second speaker’s utterance employs Mary: says the horse-thief the original speaker’s own words against them, [Jamaica Inn, BBC Transcripts] and that is the whole point of the construction (Gregoromichelaki, 2018) (13) A: SOMEONE is keen. So, instead of propositional knowledge of rules B: says the man who slept here all night and manipulation of representations, the abili- [The A-Word, BBC Transcripts] ties of individual agents for interacting efficiently Under DS assumptions, in (11), we don’t need to with others under the guidance of grammatical invoke pro-drop in English to explain the missing constraints could be characterised as subpersonal subject of Hester’s utterance. Instead, we can im- mechanisms which allow access to a normative plement the intuitive explanation that the subject landscape of affordances. Access to this landscape is provided due to the fact that the utterance is is then mediated by inducing a range of predic- a continuation of Mrs Elton’s interrogative non- tive goals (solicitations) to be fulfilled by either sentential query in the previous turn. For both ut- interlocutor in the very next steps which they will terances, there is no reason to assume that speech be taking. The skills required to take advantage acts are licensed only if interlocutors can derive of solicitations in such a context do not presup- fully specified propositional contents and/or sen- pose any ‘rational’ high-order individualistic in- tential structures embedded under speech act pred- ference, standardly taken to be the basis of all icates. Neither us, as a third party reader, nor Hes- successful human communication (Clark (1996); ter, nor Mrs Elton herself, have particular predi- Sperber and Wilson (1995) and many others). cates in mind when encountering and understand- Instead, the task of selecting appropriate actions ing “And Mr Page?” in (11) since its affordances is taken over by affordance competition, which mesh with the rest of the affordances in our re- operates at the much lower level of sensorimotor spective contexts and allow us to go on interact- contingencies (Cisek and Thura, 2019; Anderson, ing with the utterance or its utterer. And it is pre- 2014; Rietveld et al., 2018). However, this pre- cisely this lack of predicate, not only in the surface supposes that not only does the grammar incor- structure but also in the conceptual structure, that porate processing features like incrementality and allows Hester to make use of the previous utter- predictivity but also that the grammar provides a ance to build her own which is thus fulfilling the shared “workspace” (Kempen, 2014) for both production and comprehension to operate and interact 6 Along with natural data, constructed data from literature, (Gregoromichelaki et al., 2013b). film scripts etc. are particularly relevant in this context as they show that such constructions are not “speech errors” that can In addition, non-linguistic practices need to be easily dispensed with. be integrated with NL actions to contribute framework like DS, there are grounds for seeing their qualitatively identical types of constraints this restriction as a general control property of (Gregoromichelaki, 2018). Evidence that mor- socio-cognitive coordination systems (control in phosyntax directly interacts with embodied situa- the cybernetic sense of ‘regulation’, e.g. Bickhard tional affordances of every day action comes from (2009); Carver and Scheier (2012)) and, therefore, elliptical constructions. For example, in case-rich indeed a general property of physical systems in- languages such as German and Greek, elliptical volving the interaction of multiple simultaneous fragments necessarily display the form suitable to processes. what might have been a complete sentence for- In this connection, Anderson (2014) attributes mulated in response, even when there is no obvi- to individual brain mechanisms (‘abilities’) the ous antecedent of a verbal sentential form– in the function of action control via perception/action German example in (14) the accusative form den feedback loops with the environment without nec- Arzt seamlessly blends with other situational affor- essary representational mediation. Like DS, he dances to constrain future action to the effect that sees “higher” cognitive processes like NL as de- someone should call a doctor:7 velopments of already existing brain capacities, rather than as species-specific cognitive adapta- (14) A and B see a woman lying on the floor: tions. The brain is modelled as a connectionist A to B: Schnell, den Arzt/#der Arzt network, superimposed on the physical brain struc- Quick, the doctorACC /#the doctorNOM ture, inspired by Smolensky (1986)’s architecture but without the individualistic representational in- This shows that what is called an ‘antecedent’ in terpretation originally assigned to this model. The linguistic accounts of ellipsis need not have lin- state transitions and attractor landscapes of this guistic form, and can in fact be an action concur- network control the interaction of the individual rent with the utterance, or even in the future of it and the environment without presuming that the as in commands like (14). Moreover, physical ac- brain constructs a model of the world it inter- tions can combine with linguistic actions to seam- acts with. As in REEC’s view of skills, abili- lessly contribute constituents like subjects and ob- ties, and dispositions, the contribution of the in- jects to sentences (see, e.g., Slama-Cazacu, 1976): dividual brain is seen as complementary to the (15) while sitting on the piano: He doesn’t play equally significant roles of the whole body and [PLAYS TUNE]. But rather he plays –he does the environment. Brain activity on its own is in it better than I do– [PLAYS TUNE - SINGS] no way to be considered as enacting conceptually- [Clark, 2016] transparent reactions to the environment. Some sequences of states can be assigned interpretations as It seems then that NLs are not underpinned by individuated action goals but it is only at the end distinctive mechanisms residing within individual of a sequence of preparatory stages during which minds encapsulated from the demands and contin- multiple partially activated goals coexist and inter- gencies of interaction with the sociomaterial envi- act with environmental inputs that such an abstrac- ronment. tive interpretation becomes possible. Accordingly, the states that the network passes through are not 3.1 Rethinking NL Autonomy assigned representational contents but are individ- The question that then remains is how to think uated by their responsiveness to particular inputs of alleged sui-generis properties of NL morpho- and the behaviours that they enable. This con- syntax that have been adduced as arguments for trol of behavior architecture also dispenses with claims like “the autonomy of syntax” hypothe- a central decision module that processes abstract sis. Within DS, this question is pertinent regard- goals prior to action. Instead, different patterns ing the status of claimed universal NL processing of activation at each state track the gradual con- constraints like the “no-copy” restriction. While vergence of the totality of the control mechanism this remains an issue for much further develop- towards a particular action, that is, they track the ment, we suggest that, within a domain-general “decision” process to pursue a selected subset of 7For a critique of views that purport to reduce all the available affordances that compete for realisa- fragments to a sententialist form of explanation, see tion within an agent’s environment. What is im- Kempson et al. (2018). portant for us here is that, at preliminary stages, cannot be realised by neurons serving overlapping affordance competition in Anderson’s interpreta- action specifications. Similarly, in DS, a node tion is implemented by the fact that neural patterns can only accommodate compatible selections of of activation might reflect simultaneously multiple multiple macros as requirements for further up- partially-activated but incompatible options. This date, for example, a macro associated with mor- is only possible before final convergence to the se- phological marking for gender ‘female’ can com- lected action goal which can only consist of the bine with a macro for ‘3rd person singular’ but not pursuance of multiple but compatible affordances. with a macro for gender ‘masculine’. If inconsis- Architecturally, the potential for the presence of in- tent macros come to appear as goals at the same compatible options arises because various incom- node the result will be anomaly (technically imple- patible affordances can engage the activation of mented in PDL as reaching a state where absurdity, the same network nodes as long as they are not ⊥, holds or, equivalently, that the ABORT action is fully activated. At the final update stage of conver- implemented). The reason is that there will be no gence (i.e. when a “decision” has been reached), possibility of pursuing the action paths indicated the partial specifications of multiple action goals as subgoals by those macros and reaching a con- will be necessarily eliminated if their full acti- sistent state, hence the action control mechanism vation patterns are incompatible, i.e., if they re- provided by the grammar will have failed its pur- quire the same network nodes for their realisation. pose. The nonlinear update function ensures this. Only As a result, in a language like Modern Greek affordances with the strongest activation, having with its otherwise generally free word order we been gradually reinforced through environmental do nevertheless find phenomena of strict order- inputs, will survive. Their prevalence means that ing and epiphenomenally arbitrary positional ex- they will assume the role of controlling action as clusions. The processing reason for these phenom- immediate agent goals while disrupting and in- ena though is that any underspecified clitic mor- hibiting the activation patterns of any competitors. pheme will need to be processed on an unfixed We suggest that the ‘no-copy’ constraint on node whose specifications will be a particular com- LOFT tree-path traversals is of the same type and bination of compatible affordances. Any other aetiology. The graph structure imposed by LOFT similarly underspecified clitic cannot then follow instantiates an action control mechanism within because its morphological specifications will an- the DS conception of NLs as grounded in goal- notate the same node leading to inconsistency. For directed action. Restrictions such as the ‘no-copy this reason, whichever clitic morpheme is initially constraint’ then provide a restricted architectural more strongly favoured by the contextual expres- bottleneck to facilitate action selection in view of sive needs of the interlocutor will win the compe- multiple competing action opportunities in the en- tition among the various candidate combinations vironment. To take a concrete example, we as- and will get to occupy the relevant cluster posi- sume that an ‘unfixed node’ state can accommo- tion thus excluding seemingly unaccountably the date multiple affordances as goals. In the case appearance of other morphemes not only in the of the Greek clitic clusters (see (4)-(8)), during current position but also in sequentially later ones. production, multiple gender, person, and number The only resort remaining for the expression of specifications, compete for realisation at the se- such a defeated and thus excluded combination of quential cluster positions. Each clitic morpheme specifications will then be to appear not as a clitic at a particular cluster position when processed but as a full NP, pronominal or nominal. will trigger a combination of affordances (macros) So, affordance competition is the underlying appearing on a single treenode. But given the cause for action selection, action control, both fa- multiplicity of competing specifications, the even- cilitation and obstruction, with NL surface struc- tual combination of affordance annotations on any tural constraints being subsumed under an overall treenode corresponding to a single cluster posi- behaviour-guidance and general systems-control tion will be the outcome of a convergence after architecture. affordance competition. In this sense, each treen- In closing, we note that, besides capturing non- ode’s specification of action goals is the equiv- determinism and its structural effects in the up- alent of fully-specified activation patterns which dates achieved by morpho-syntactically induced actions, a further integral task in this program is nals follow constrained trajectories modelled as to capture the enormous range of interpretation graph-transitions of restricted shapes, with an in- possibilities which words display, a challenge of- termediate level of tree-structured paths. This ten treated as peripheral for grammar frameworks. structuring accounts for apparent idiosyncratic However, in our view, given the domain-generality morphosyntactic constraints because it channels of DS, the same mechanisms set out here can ac- processing towards accomplishing interpretation count for contextual modulation of word meaning and production in an incremental manner by re- in terms of updates selecting, through affordance stricting the available predictions. Considerable competition, the relevant aspects of meaning of but not unconstrained flexibility is achieved due words in context. Words can be seen as trigger- to various types of licensed underspecification and ing packages of sets of potential actions (macros) updates and action selection requirements are im- formalised as TTR types (Gregoromichelaki et al., plemented despite the multiplicity of action oppor- 2019). tunities in the environment. During the process Accounting for contextual shift in meaning, cur- of formulating or comprehending an utterance, rent DS work on the incorporation of a vector- the initial availability of various options for up- based semantics for the affordances words provide date is gradually restricted via incremental culling is ongoing, with the goal of modelling the per- of options. This is carried out as constraints vasive variability of word meaning and its incre- accumulate from various sources and at various mental resolution, without invariably invoking lex- stages, with a process of affordance competition ical ambiguity. This approach builds on previous serving as the necessary action selection mech- work on compositional distributional semantics anism. This mechanism replaces “high order” for pregroup grammars and the Lambek Calculus cognitive individualistic and internalist notions (Coecke et al., 2010, 2013), and for Combinatory like intention recognition, mind-reading, and plan- Categorial Grammar (CCG) (Maillard et al., 2014; ning (see also Mills and Gregoromichelaki, 2010; Moot, 2018). There is current work in this area, Silberstein and Chemero, 2011; Froese, 2018) and where distributional meanings of words are com- implements relevance considerations at the low- bined with their grammatical types, albeit more in- level of subpersonal mechanisms of action con- formally (e.g. Baroni and Zamparelli (2010)). But trol. Within this perspective, it is notable how this these do not consider an incremental setting with action control mechanism applies across sources underspecified nodes in the grammatical construc- of constraints bridging the levels of both “lower-” tion, a challenge which our program addresses. In and “high-level” cognition. this vein, Purver et al. (forthcoming) suggest two We believe that, by abandoning representational distributional counterparts to underspecificational individualism and grasping the potential that re- notions, and present experimental results regard- thinking language as an action system offers, we ing affordance competition that results in optimal can explain in an integrated way the dynamics of choice in the disambiguation task dialogue interaction and the seamless interweav- ing of NL utterances with nonverbal actions. Ad- 4 Conclusion ditionally, given that NL constraints are here sub- The shift of emphasis from static structural gen- sumed under a theory of general action, the prac- eralisations to actions is central here. The goal tices guiding NL users are a source of normativity of NL processing and NL interaction is not to independent of themselves, hence avoiding “pri- generate a sentence structure or proposition as vate language” Wittgensteinian objections. On such: it is to promote progressive context-updates the other hand, the established practices guiding via pursuing selected action opportunities (affor- action need to be enacted afresh every time in dances) yielding further affordances guided by new contexts, and either reinforced if successful the various practices constraining and enabling or modified if ineffective (see, e.g., Eshghi et al., the actions of the interlocutors. Given the DS 2017b). We can then get a handle on both the architecture’s domain-generality, constraints aris- apparent stability of a language for its users and ing from non-verbal practices blend seamlessly yet its endless potential for ongoing variability and with morpho-syntactic and semantic constraints. change. Furthermore, combining these perspec- Context-updates ensuing from NL sequential sig- tives, the ontogenetic trajectory of acquiring a first language can then be seen not as maturational de- Patrick Blackburn and Wilfried Meyer-Viol. 1994. Linguis- velopment executing a genetic program but as a tics, logic and finite trees. Logic Journal of the Interest Group of Pure and Applied Logics, 2(1):3–29. grounded process crucially involving interaction with the social environment and scaffolded by J. Briglia, P. Servajean, A. H. Michalland, L. Brunel, and D. Brouillet. 2018. Modeling an enactivist multiple-trace non-verbal expressive and communicative joint ac- memory. ATHENA: A fractal model of human memory. tions like gesture and turn-taking (Clark and Kelly, Journal of Mathematical Psychology, 82:97–110. to appear). Jelle Bruineberg, Anthony Chemero, and Erik Rietveld. 2019. General ecological information supports engage- Acknowledgements ment with affordances for ‘higher’ cognition. Synthese, 196(12):5231–5251. We wish to thank three anonymous reviewers for their careful reading and thought-provoking com- Joan Bybee. 2010. Language, Usage and Cognition. Cam- bridgge University Press. ments which led, in our minds, to significant im- provements to this paper. We would also like to Ronnie Cann, Ruth Kempson, and Lutz Marten. 2005. The thank the Technical Programme Chairs for impor- Dynamics of Language. Elsevier, Oxford. tant practical assistance. Charles S Carver and Michael F Scheier. 2012. Cybernetic Chatzikyriakidis and Howes were supported by control processes and the self-regulation of behavior. In The Oxford handbook of human motivation, pages 28–42. grant 2014-39 from the Swedish Research Coun- Oxford University Press New York, NY. cil, (VR), which funds the Centre for Linguistic Theory and Studies in Probability (CLASP) in Stergios Chatzikyriakidis. 2010. Clitics in four dialects of Modern Greek: A dynamic account. Ph.D. thesis, King’s the Department of Philosophy, Linguistics, and College, London. Theory of Science at the University of Gothen- Stergios Chatzikyriakidis and Ruth Kempson. 2011. Stan- burg. Howes was also supported by the VR grant dard Modern and Pontic Greek person restrictions: A 2016-0116 – Incremental Reasoning in Dialogue feature-free dynamic account. Journal of Greek Lin- (IncReD). Purver is partially supported by the gusitics, pages 127–166. EPSRC under grant EP/S033564/1, and by the Anthony Chemero. 2009. Radical Embodied Cognitive Sci- European Union’s Horizon 2020 programme un- ence. MIT Press. der grant agreements 769661 (SAAM, Support- Morten Christiansen and Nick Chater. 2008. Language as ing Active Ageing through Multimodal coaching) shaped by the brain. Brain and Behavioral Sciences, and 825153 (EMBEDDIA, Cross-Lingual Embed- 31:489–509. dings for Less-Represented Languages in Euro- Paul Cisek and David Thura. 2019. Reach-to-Grasp Be- pean News Media). The results of this publica- havior, chapter Neural circuits for action selection. Rout- tion reflect only the authors’ views and the Com- ledge. mission is not responsible for any use that may be Andy Clark. 2016. Surfing Uncertainty: Prediction, Action, made of the information it contains. and the Embodied Mind. Oxford University Press, Ox- ford.

Eve Clark and Barbara Kelly. to appear. Constructing a References system of communication with gestures and words. In M Eulalia Bonet i Alsina. 1991. Morphology after syntax– Goldin-Meadow and Morgenstern Aliyah, editors, Ges- pronominal clitics in romance. Ph.D. thesis, Mas- ture in language. Mouton de Gruyter. sachusetts Institute of Technology. Herbert H. Clark. 1996. Using Language. Cambridge Uni- Michael L. Anderson. 2014. After Phrenology. Cambridge versity Press. University Press. Bob Coecke, Edward Grefenstette, and Mehrnoosh Marco Baroni and Roberto Zamparelli. 2010. Nouns are vec- Sadrzadeh. 2013. Lambek vs. lambek: Functorial tors, adjectives are matrices: Representing adjective noun vector space semantics and string diagrams for lambek constructions in semantic space. In Proceedings of the calculus. Annals of Pure and Applied Logic, 164(11):1079 2010 Conference on Empirical Methods in Natural Lan- – 1100. Special issue on Seventh Workshop on Games for guage Processing, pages 1183–1193, Cambridge MA. As- Logic and Programming Languages (GaLoP VII). sociation for Computational Linguistics. Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen Clark. Johan van Benthem. 2011. Logical dynamics of information 2010. Mathematical foundations for a compositional dis- and interaction. Cambridge University Press. tributional model of meaning. Linguistic Analysis, 36(1– 4):345–384. Special Issue: A Festchrift for Joachim Lam- Mark H Bickhard. 2009. The interactivist model. Synthese, bek. 166(3):547–591. Robin Cooper. 2012. Type theory and semantics in ﬂux. In Eleni Gregoromichelaki, Christine Howes, and Ruth Kemp- Ruth Kempson, Nicholas Asher, and Tim Fernando, edi- son. 2019. Actionism in syntax and semantics. In Gothen- tors, Handbook of the Philosophy of Science, volume 14: burg Proceedings of Conference on Dialogue and Percep- Philosophy of Linguistics, pages 271–323. North Holland. tion. CLASP.

Robin Cooper and Jonathan Ginzburg. 2015. Type theory Eleni Gregoromichelaki, Ruth Kempson, Christine Howes, with records for natural language semantics. The Hand- and Arash Eshghi. 2013b. On making syntax dynamic: book of Contemporary Semantic Theory, pages 375–407. The challenge of compound utterances and the architecture of the grammar. In Ipke Wachsmuth, Jan de Ruiter, Matthew S. Dryer. 2013. Order of subject, object and verb. Petra Jaecks, and Stefan Kopp, editors, Alignment in Com- In Matthew S Dryer and Martin Haspelmath, editors, The munication: Towards a New Theory of Communication, World Atlas of Language Structures Online. Max Planck Advances in Interaction Studies (series editors: Kerstin Institute for Evolutionary Anthropology, Leipzig. Dautenhahn, Angelo Cangelosi), pages 57–85. John Ben- jamin. Arash Eshghi, Julian Hough, and Matthew Purver. 2013. Incremental grammar induction from child-directed dia- Eleni Gregoromichelaki, Ruth Kempson, Matthew Purver, logue utterances. In Proceedings of the 4th Annual Work- Greg J. Mills, Ronnie Cann, Wilfried Meyer-Viol, and shop on Cognitive Modeling and Computational Linguis- Patrick G. T. Healey. 2011. Incrementality and intention- tics (CMCL), pages 94–103, Soﬁa, Bulgaria. Association recognition in utterance processing. Dialogue and Dis- for Computational Linguistics. course, 2(1):199–233.

Arash Eshghi and Oliver Lemon. 2014. How domain-general Martin Haspelmath. 2020. Human linguisticality and the can we be? Learning incremental dialogue systems with- building blocks of languages. Frontiers in Psychology, out dialogue acts. In Proceedings of Semdial 2014 (Dial- 10:1–10. Watt). Willy van Hoeke. 1996. The Latin dative. In William van Arash Eshghi, Igor Shalyminov, and Oliver Lemon. 2017a. Belle and Willy Langendonck, editors, The Dative Volume Bootstrapping incremental dialogue systems from mini- 1: Descriptive Studies, pages 3–38. John Benjamins. mal data: the generalisation power of dialogue grammars? In Proceedings of the 2017 Conference on Empirical Meth- Julian Hough. 2015. Modelling Incremental Self-Repair Pro- ods in Natural Language Processing (EMNLP). cessing in Dialogue. Ph.D. thesis, Queen Mary University of London. Arash Eshghi, Igor Shalyminov, and Oliver Lemon. 2017b. Interactional Dynamics and the Emergence of Language Julian Hough and Matthew Purver. 2012. Processing self- Games. In Proceedings of ESSLLI 2017. repairs in an incremental type-theoretic dialogue system. In Proceedings of the 16th SemDial Workshop on the Se- Nicholas Evans and Stephen Levinson. 2009. The myth of mantics and Pragmatics of Dialogue (SeineDial), pages language universals: language diversity and its importance 136–144, Paris, France. for cognitive science. Brain and Behavioral Sciences, 32:429–492. Daniel D. Hutto and Erik Myin. 2012. Radicalizing Enac- tivism: Basic Minds without Content. MIT Press. Jerry A. Fodor. 1983. The Modularity of Mind. MIT Press. Daniel D. Hutto and Anco Peeters. 2018. The Roots of Re- Tom Froese. 2018. Searching for the conditions of genuine membering: Radically Enactive Recollecting. In Kourken intersubjectivity: from agent-based models to perceptual Michaelian, Dorothea Debus, and Denis Perrin, editors, crossing experiments. In Albert Newen, Leon De Bruin, New Directions in the Philosophy of Memory, pages 97– and Shaun Gallagher, editors, The Oxford handbook of 4e 118. New York: Routledge. cognition, pages 163–186. Oxford University Press, Ox- ford, UK. Gerard Kempen. 2014. Prolegomena to a neurocomputa- tional architecture for human grammatical encoding and Tom Froese and E. J. Izquierdo. 2018. A Dynamical Ap- decoding. Neuroinformatics, 12(1):111–142. proach to the Phenomenology of Body Memory: Past In- teractions Can Shape Present Capacities Without Neuro- R. Kempson and J. Kiaer. 2010. Multiple long-distance plasticity. Journal of Consciousness Studies, 25(7-8):20– scrambling: Syntax as reﬂections of processing. Journal 46(27). of Linguistics, 46(01):127–192.

C. Randy Gallistel and Adam Philip King. 2009. Ruth Kempson, Ronnie Cann, Eleni Gregoromichelaki, and Memory and the Computational Brain. Wiley-Blackwell, Stergios Chatzikiriakidis. 2016. Language as mechanisms Oxford, UK. for interaction. Theoretical Linguistics, 42(3-4):203–275.

E. Goffman. 1981. Forms of Talk. University of Pennsylva- Ruth Kempson, Eleni Gregoromichelaki, Arash Eshghi, and nia Press. Julian Hough. 2018. Ellipsis in Dynamic Syntax. In Jeroen van Craenbroeck and Tanya Temmerman, editors, Eleni Gregoromichelaki. 2018. Quotation in dialogue. In Oxford Handbook of Ellipsis. Oxford University Press. The Semantics and Pragmatics of Quotation, pages 195– 255. Springer. Ruth Kempson, Wilfried Meyer-Viol, and Dov Gabbay. 2001. Dynamic Syntax. Blackwell. Eleni Gregoromichelaki, Ronnie Cann, and Ruth Kempson. 2013a. On coordination in dialogue: subsentential talk Jieun Kiaer. 2007. Processes and Interfaces in Syntactic The- and its implications. In Laurence Goldstein, editor, On ory: the case of Korean. Ph.D. thesis, King’s College Lon- Brevity. OUP. don. Jieun Kiaer. to appear. Pragmatic Syntax. Bloomsbury. Mehrnoosh Sadrzadeh, Matthew Purver, Julian Hough, and Ruth Kempson. 2018. Exploring semantic incremental- Gregory M Kobele. 2012. Eliding the derivation: a minimal- ity with dynamic syntax and vector space semantics. In ist formalization of ellipsis. In Proceedings of the 19th Proceedings of the 22nd Workshop on the Semantics and International Conference on HPSG. Pragmatics of Dialogue (SemDial 2018 - AixDial), Aix- en-Provence. Masatoshi Koizumi. 2000. String vacuous overt verb raising. Journal of East Asian Linguistics, 9:227–85. Yo Sato. 2011. Local ambiguity, search strategies and parsing in Dynamic Syntax. In E. Gregoromichelaki, R. Kemp- P. Linell. 2005. The Written Language Bias in Linguistics: Its son, and C. Howes, editors, The Dynamics of Lexical In- Nature, Origins and Transformations. Psychology Press. terfaces. CSLI Publications.

Jean Maillard, Stephen Clark, and Edward Grefenstette. 2014. Michael Silberstein and Anthony Chemero. 2011. Dynamics, A type-driven tensor-based semantics for CCG. In Pro- Agency and Intentional Action. Humana Mente, 4(15):1– ceedings of the EACL 2014 Workshop on Type Theory 19. and Natural Language Semantics (TTNLS), pages 46–54, Gothenburg, Sweden. Association for Computational Lin- Michael Silberstein and Anthony Chemero. 2012. guistics. Complexity and Extended Phenomenological-Cognitive Systems. Topics in Cognitive Science, 4(1):35–50. David Marr. 1982. Vision: A Computational Approach. Free- man & Co, San Francisco. T. Slama-Cazacu. 1976. Nonverbal components in message sequence: ‘mixed syntax’. In W.C. McCormack and S.A. G. Mills and E. Gregoromichelaki. 2010. Establishing coher- Wurm, editors, Language and Man: Anthropological Is- ence in dialogue: sequentiality, intentions and negotiation. sues. Mouton, The Hague, pages 217–227. In Proceedings of SemDial (PozDial). Paul Smolensky. 1986. Parallel distributed processing, vol- Richard Moot. 2018. Combining logical and distributional ume 2, chapter Neural and conceptual interpretation on methods in type-logical grammars. Journal of Language PDP models. MIT Press, Cambridge. Modelling, 6:288–317. Dan Sperber and Deirdre Wilson. 1995. Relevance: Commu- Ezequiel A. Di Paolo, Elena Clare Cuffari, and Hanne De nication and Cognition, second edition. Blackwell. Jaegher. 2018. Linguistic Bodies: The Continuity between Life and Language. MIT Press. Mark Steedman. 2000. The Syntactic Process. MIT Press, Cambridge, MA. David M Perlmutter. 1970. Surface structure constraints in syntax. Linguistic inquiry, 1(2):187–255. Markus Werning. 2020. Predicting the past from minimal traces: Episodic memory and its distinction from imagina- Alan Prince and Paul Smolensky. 1997. Optimality: tion and preservation. Review of Philosophy and Psychol- From neural networks to universal grammar. Science, ogy, 11(2):301–333. 275(5306):1604–1610. L. Wittgenstein. 1953. Philosophical Investigations, Trans. Bradley L. Pritchett. 1992. Parsing with Grammar: Is- G.E.M. Anscombe. Oxford: Blackwell. lands, Heads, and Garden Paths, pages 321–349. Springer Netherlands, Dordrecht. Jeffrey Witzel and Naoko Witzel. 2016. Incremental sentence processing in japanese: a maze investigation into scram- Geoffrey Pullum and Barbara Scholz. 2001. On the distinc- bled and control sentences. Journal of Psycholinguistic tion between model-theoretic and generative-enumerative Research, 45:475–505. syntactic frameworks. In Proceedings of the 4th Interna- tional Conference on Logical Aspects of Computational Linguistics (LACL), pages 17–43. Springer.

Matthew Purver, Arash Eshghi, and Julian Hough. 2011. In- cremental semantic construction in a dialogue system. In Proceedings of the 9th International Conference on Com- putational Semantics, pages 365–369, Oxford, UK.

Matthew Purver, Mehrnoosh Sadrzadeh, Ruth Kempson, Gijs Wijnholds, and Julian Hough. forthcoming. Incremen- tal composition in Distributional Semantics. Journal of Logic, Language and Information.

Erik Rietveld, Damiaan Denys, and Maarten Van Westen. 2018. Ecological-Enactive Cognition as Engaging with a Field of Relevant Affordances. In The Oxford Handbook of 4E Cognition, page 41. Oxford University Press.

Erik Rietveld and Julian Kiverstein. 2014. A rich landscape of affordances. Ecological Psychology, 46.