3.2.1 Processing and linguistic theory

The question of how language is processed is logically inseparable from questions regarding the specific nature of language. It should not be controversial to say that research and theory in language processing stand to benefit greatly from an understanding of language (or that the study of language can benefit from an understanding of processing). This is, again, one of the strengths of a Jackendovian approach, that it takes both seriously. There are of course widely differing views on the nature of language, so adopting a particular linguistic theory is by no means a guarantee of success. But disregarding the rich bodies of research and theory on language may well be a guarantee of failure.

So linguistic theory and processing theory must ultimately come together. This coming together can take various forms. Knowledge of language could be treated as a largely independent system that is consulted by processing mechanisms as they carry out their work. A more parsimonious approach would be one in which the two are shown to be essentially one, sharing the same principles. This amounts to treating linguistic principles as processing principles, an approach pursued by, for example, Pritchett (Reference Pritchett1988,Reference Pritchett1992), Weinberg (Reference Weinberg1993,Reference Weinberg, Epstein and Hornstein1999), Crocker (Reference Crocker1996), and Dekydtspotter (Reference Dekydtspotter2001). Connectionist approaches to language also fit in this category, though accounts of the nature of language are generally not so well developed in them as they are in generative-oriented approaches to processing. As we shall see later on, it is eminently feasible to combine a generative-oriented approach with some version of connectionism, albeit not the one most commonly associated with this term (see the discussion in 1.2).

3.2.2 Modularity and interaction

One highly controversial issue is the degree and nature of interaction between syntactic and conceptual processing, views ranging from Frazier's (Reference Frazier1979) strictly modular approach to the unrestricted interaction of McClelland, St. John, and Taraban (Reference McClelland, St. John and Taraban1989). This range naturally parallels the diversity of views on modularity described in Chapter 2. In processing terms, the logic of a hypothesised separation between syntactic and semantic processing is that this specialisation allows extremely efficient processing at each level. The existence of an independent syntactic processor is suggested by evidence that syntactic structures can be primed independently of semantic factors (e.g. Bock Reference Bock1986; Ferreira and CliftonReference Ferreira and Clifton1986; Branigan et al. Reference Branigan, Pickering, Liversedge, Stewart and Urbach1995), by studies of self-paced reading and eye movements during reading (Frazier and RaynerReference Frazier and Rayner1982; Ferreira and HendersonReference Ferreira and Henderson1990; Meunier and LongtinReference Meunier and Longtin2007; Omaki Reference Omaki2010), and through the use of the visual world paradigm, which involves monitoring participants gaze when they are presented with pictures of potential referents while they process a sentence (Omaki Reference Omaki2010). Neurolinguistic research using various ways of measuring brain activity indicates that syntactic and semantic processing are neurally distinct, providing further support for a specifically syntactic representation (see Brown and HagoortReference Brown, Hagoort, Crocker, Pickering and Clifton2000; Friederici Reference Friederici2001). The same can be said for studies of the effects of brain damage (Breedin and Saffran,Reference Breedin and Saffran1999). van Gompel and Pickering (Reference van Gompel, Pickering and Gaskill2007) provide a useful review of the overall evidence, concluding that it suggests largely separate syntactic and semantic processing. But the case for some interaction is compelling (e.g. Steedman and AltmannReference Steedman and Altmann1989; MacDonald, Pearlmutter, and SeidenbergReference MacDonald, Pearlmutter and Seidenberg1994; Pickering and TraxelReference Pickering and Branigan1998).

So while the issues are complex and the evidence mixed (see also, for example, LevyReference Levy1996), the apparent conclusion is that syntactic and semantic processing are distinct but do interact, the issue being how limited the interaction is and what its exact nature is. A successful model should respect the distinction between syntax and semantics in terms of both representation and processing but at the same time should allow meaningful but constrained influences between the levels.

One approach that meets these requirements is the incremental-interactive theory (IIT) of Crain and Steedman (Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985), Altmann and Steedman (Reference Altmann and Steedman1988), and Steedman and Altmann (Reference Steedman and Altmann1989) (see also Gorrell Reference Gorrell1995). In IIT, the syntax overgenerates, producing multiple possible representations in parallel, without reference to conceptual information. While they are being constructed, the bits so far assembled are passed on to conceptual processors (incrementally), which reject those they find least acceptable, thereby cutting off those derivations in the syntax. Thus, syntactic representations are constructed purely in terms of syntactic principles but the ultimate selection of a suitable representation is based on conceptual factors. IIT has proven problematic in some respects, despite the appeal of its leading idea. There is reason to believe, first, that the selection process is not entirely conceptual but also involves syntactic factors (Mitchell, Corley, and GarnhamReference Mitchell, Corley and Garnham1992). The details of IIT's conceptual selection process have also been challenged (e.g. Hickok Reference Hickok1993; Nicol and PickeringReference Nicol and Pickering1993).

3.2.3 Competition

A major theme in the processing literature has been competition among items available for use in processing. When input is being processed or output is being produced, many more items are available than will actually be used in the ultimate representations. A selection process picks out those that are most appropriate and eliminates the others. The selection process can be neatly captured in the notion of competition among the candidate items.

This idea of competition has appeared in a great many approaches to various aspects of cognition, including the acquisition and use of language. An early example can be found in Morton's (Reference Morton1969) Logogen Model of word recognition (see also Morton Reference Morton, Morton and Marshall1979). Morton did not present his theory as a ‘competition’ model, but it does have the key attributes that now qualify it as one. Logogens representing words receive signals from sensory analysis; when enough stimulation has accumulated in a given logogen it puts out a response, in effect announcing that it is the word being received. This signal goes to the Output Buffer by way of a single channel that allows only one signal at a time, so the first logogen to reach a threshold is selected. In effect, the various logogens are constantly competing for control of the single channel. Related ideas can be found in a great many sources on language processing (e.g. Marslen-Wilson and WelshReference Marslen-Wilson and Welsh1978; McClelland and RumelhartReference McClelland and Rumelhart1981; Dell Reference Dell1986; Clifton and StaubReference Clifton and Staub2008; Abdel-Rahman and MelingerReference Abdel Rahman and Melinger2009; Roelofs, Piai, and Schriefers 2013).

Outside the area of language processing, competition also plays a central role in Baars’ (Reference Baars1988) theory of consciousness, which has influenced our thinking in a number of ways, as will become clear in Chapter 8. He hypothesised a global workspace (GW), serving as a platform on which the work of a processor can be shared with the rest of the cognitive system whenever it has general value, i.e. provides information that is useful to other processors. The contents of the GW at any given time are determined by competition among processors. The visual system might at this moment be producing a representation of the words on this page while the auditory system processes the speech of someone nearby and thoughts of plans for the evening percolate in the back of the mind. At any given moment, only one of these can dominate the GW and thereby become conscious. The stream of consciousness is in effect the product of a continuous competition process among candidate representations.

MacWhinney developed a related idea in his Competition Model (e.g. Bates and MacWhinney Reference Bates, MacWhinney and MacWhinney1987; MacWhinney Reference MacWhinney and MacWhinney1987), focusing on the ways that the conceptual role of agent is assigned to one, and only one, of the noun phrases in a sentence that is being processed. This idea has been adopted in a great deal of work based on his approach (e.g. MacWhinney and Bates Reference MacWhinney and Bates1989). It should be stressed, though, that our use of competition was not based on and should not be identified with that found in the Competition Model. Another possible confusion stems from the fact that the term competition is sometimes identified with highly non-modular processing theories hypothesising unrestricted information flow. This free flow of information produces a variety of possible interpretations for the input, which compete with one another (see Vosse and Kempen Reference Vosse and Kempen2009). Our use of the term should not be taken as an endorsement of such models.

A final example of competition can be found in the account of second language learning proposed by Felix (Reference Felix1987, Reference Felix, Flynn and O’Neil1988), which he referred to as a competition model. In his approach, L2 learners have a specifically linguistic system and also non-specific General Problem Solving mechanisms, the two systems competing for input and therefore getting in each other's way in the process of language development.

3.2.4 Incremental processing

One relatively clear finding is that semantic characteristics of a sentence can affect processing long before the end of the sentence is reached (e.g. Tyler and Marslen-WilsonReference Tyler and Marslen-Wilson1977; Traxler and PickeringReference Traxler and Pickering1996; Williams Reference Williams2006; van Gompel and Pickering Reference van Gompel, Pickering and Gaskill2007; Omaki Reference Omaki2010). Thus, conceptual processing does not wait for syntactic analysis to be completed; results of syntactic processing are presented to conceptual processors incrementally. When a person hears the sentence Mary kissed Bill, for example, the word Mary is subjected to phonological and syntactic processing before kissed, simply because it is received first. Semantic processing of Mary then begins as soon as the phonology and syntax identify it as a word, and before the remaining words have yet been identified. When the phonology and syntax make kissed available, semantic processing then works with Mary kissed, before Bill is yet available to it. We will take this to be a very general characteristic of processing, inside and outside the language module.

3.2.5 Activation

A common idea in processing work is that items vary in their level of activation and that the more active an item is the more readily available it is for use in processing. Thus, activation commonly serves as the mechanism underlying processing competition, sometimes in these terms and sometimes not. An item's activation level is commonly seen as a function of its past use in processing and therefore as reflecting learning by the system. It also serves as an account of priming effects, in which the use of one item in processing raises the likelihood that a related item will be selected for use shortly afterward, or reduces the time required for the latter (primed) item to be selected. In the classic example, interpretation of the ambiguous word bank will be influenced by whether the person just heard river or money.

The underlying mechanism is spreading activation. When money is activated during processing, this activation spreads, to varying degrees, to items that are in some way related to it, such as the financial institution meaning of bank. Activation of this semantic representation makes it more readily available for subsequent use, biasing the interpretation of bank toward financial institution and away from river bank. Spreading activation can also explain the common observation that the stream of consciousness tends to follow paths of association. All these aspects of the activation idea are incorporated in MOGUL, so a brief survey of past work is appropriate.

A classic example of these ideas is provided by Dell's (Reference Dell1986) model of production, which in some respects resembles the approach that we will adopt below. The model was made to account for speech errors. Each item in the system is assigned an activation level, with 0 as the basic resting level. Whenever the number rises above 0, activation spreads to neighbouring items. During production, a number of conflicting items or rules are available for use, and the ones actually selected are those that currently have the highest activation level; in other words, activation level serves as a means of resolving competition among various candidates for inclusion in current processing. Errors occur because inappropriate items are activated as a result of their relation to the target items, creating the possibility that they will be wrongly selected.

In terms of the modular vs. interactive distinction discussed above, Dell's version of spreading activation falls very much on the interactive side. An alternative that is more modular and is for that reason closer to our own approach, is that of Levelt (Reference Levelt, Brown and Hagoort1999; Levelt, Roelofs, and Meyer Reference Levelt, Brown and Hagoort1999). Levelt split the production process into a series of levels, each responsible for a particular aspect of the sentence being generated, ranging from the development of a message to its actual articulation. Activation spreads in two ways: within the items at a level and then downward, active items at the higher level activating their counterparts at the next level down.

Spreading activation accounts like Dell's fit into the area of connectionist work and the idea of spreading activation has become closely associated, even identified, with such theories. But this association is by no means necessary, as can be seen in the fact that Levelt's approach is not normally classified as connectionist. The concept of spreading activation in fact preceded connectionist work as such, originally becoming established with the work of Collins and Loftus (Reference Collins and Loftus1975) on semantic networks, which was based on the earlier work of Ross Quillian. The ideas of activation and spreading activation have since been used very extensively in a variety of work (e.g. McClelland and RumelhartReference McClelland and Rumelhart1981; Anderson Reference Anderson1983a; BockReference Bock1986; Dijkstra and van HeuvenReference Dijkstra, van Heuven, Grainger and Jacobs1998; Saunders and MacLeodReference Saunders and MacLeod2006; Schwartz et al. Reference Schwartz, Dell, Martin, Gahl and Sobel2006; Eder and KlauerReference Eder and Klauer2007; Dijksterhuis and AartsReference Dijksterhuis and Aarts2010). Our own use resembles these applications but differs from most of them, crucially, in that it occurs within a modular system, a point that we will develop below.

The priming effects described here are most often lexical, one word priming another word, but no less interesting is the phenomenon of syntactic or structural priming (Bock Reference Bock1986; Ferreira and CliftonReference Ferreira and Clifton1986; Hartsuiker and KolkReference Hartsuiker and Kolk1998; Hartsuiker et al. Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and Vanderelst2008; Pickering and Ferreira Reference Pickering and Ferreira2008; Thothathiri and SnedekerReference Thothathiri and Snedeker2008; Weber and IndefreyReference Weber and Indefrey2009; Reitter, Keller, and MooreReference Reitter, Keller and Moore2011). When a particular structure has been processed recently, it makes the use of that structure more likely in subsequent processing. A person is more likely to use a passive sentence, for example, after hearing a passive sentence than after hearing an active sentence. This effect has also been analysed in terms of activation (e.g. Pickering and BraniganReference Pickering and Branigan1998; Cleland and Pickering Reference Cleland and Pickering2006). The structure that has recently been used has an elevated activation level for that reason. This level persists for a time, making its subsequent use more likely. Research has shown that the effect can be quite lengthy, leading some to see syntactic priming as a form of implicit learning, producing lasting changes in the system (Bock and GriffinReference Bock and Griffin2000; Chang et al.Reference Chang, Dell, Bock and Griffin2000; Cleland and Pickering Reference Cleland and Pickering2006; Hartsuiker et al. Reference Hartsuiker, Bernolet, Schoonbaert, Speybroeck and Vanderelst2008; Pickering and Ferreira Reference Pickering and Ferreira2008). We will return to this point in the following chapter.

3.2.6 Dual storage and processing as a race

Another recurring theme in the processing literature is the distinction between processing input in terms of stored chunks and analysing it into simpler component parts. When a person hears the phrase more often than not, for example, we can imagine two possibilities for the way it is processed. First, it might be analysed in terms of the individual words of which it is composed, their meanings being combined to form an overall meaning for the phrase. Alternatively, the entire phrase might already be stored, as a result of prior experience with it, and the processing could then be simply a matter of activating the stored meaning. The analogous distinction in production is that between expressing a message in terms of stored chunks, possibly including more often than not, and constructing the intended utterance by combining simple parts on-line.

There has been a temptation in the past to think of the issue in absolute terms: accessing complex forms is entirely a matter of decomposing them or is entirely a matter of direct whole-form recognition. An improvement on this all-or-nothing idea is the view that derivational forms are retrieved as wholes and inflected forms by decomposition. But research described in Section 2.6 provides considerable reason to think that all these views are overly simple. This research has in fact left little doubt that language consists of both stored chunks and simpler parts and that processing is both retrieval and on-line construction/analysis. It also indicates that while the inflection–derivation distinction is important, it does not provide a reliable line between types of storage or between types of processing. Inflected forms are more likely to be processed in terms of their component parts, but they are sometimes stored and accessed as wholes. What this research seems to suggest is that any complex form can in principle be retrieved as a whole during processing or can be constructed online and that the actual outcome can be explained as a function of the past use of the wholes and their parts in processing. Whether a person processes more often than not as a single chunk or as a combination of simpler parts depends on how many times the phrase has been processed previously and how often the component words have been processed individually.

Thus, the processing system has two options for accessing complex forms during comprehension. The whole form could be recognised as such in the input signal or the component parts could be recognised individually and the composite form then assembled from them. These two routes might well differ in the amount of time they require in any given case. And of course speed is crucial in processing, so this difference could be very meaningful. Not surprisingly then, a number of researchers have hypothesised that aspects of processing are essentially a race between the two routes (Caramazza, Laudanna, and RomaniReference Caramazza, Laudanna and Romani1988; Frauenfelder and SchreuderReference Frauenfelder, Schreuder, Booij and van Marle1992; Schreuder and BaayenReference Schreuder, Baayen and Feldman1995; Baayen and Schreuder Reference Baayen, Schreuder, Dijkstra and de Smedt1996; Baayen, Dijkstra, and Schreuder Reference Baayen, Dijkstra and Schreuder1997; Pinker and UllmanReference Pinker and Ullman2002; Niswander-Klement and PollatsekReference Niswander-Klement and Pollatsek2006). In fact, much of the evidence cited above for dual storage consists of findings that the amounts of time taken for lexical access vary in ways that can be explained in terms of a race between a whole-form route and a decompositional route.

The race notion can be generalised beyond the question of chunks vs. on-line construction to potentially serve as a general way of understanding competition in processing. If one assumes parallel processing, the processing line that produces an acceptable result first is the winner of the competition. At a lower level, the individual items that are the fastest getting into the representations under construction are more likely to be the ones that ultimately find a place there, and the sub-processors that are able to impose their representations on a store are more likely to control the overall construction process.

Kuperman, Bertram, and Baayen (Reference Kuperman, Bertram and Baayen2010) argued for significant revisions in the dual-route view, based on interaction effects found in their experiments on visual word recognition (see also Kuperman, Bertram, and Baayen Reference Kuperman, Bertram and Baayen2008; Kuperman et al. Reference Kuperman, Schreuder, Bertram and Baayen2009). Specifically, they concluded that we need to adopt a more complex view, a ‘multiple-route’ model which allows for the use of a greater variety of information (coming from ‘morphemes, combinations of morphemes, morphological paradigms and structurally complex words’, p. 95) and greater interaction between information types than is commonly assumed. One implication is that there cannot be any fixed order in which whole-form and decompositional processing occur; processing cannot even be analysed in terms of two distinct routes. Note that this sort of freedom in the parser's use of information does not imply that a single processor can freely use phonological, syntactic, and conceptual information; i.e., it does not challenge modularity. (It does not support modularity either; the two are distinct issues.) Another implication the authors drew is that the race idea, at least as it is commonly presented, does not work. We will examine these conclusions and the arguments for them in more detail below, in considering their relevance for MOGUL processing.

3.2.7 Serial vs. parallel processing

Another debate in the processing literature is over whether syntactic processors produce analyses one at a time, i.e. serially, creating a second analysis only after the first has been abandoned, or produce multiple possible analyses at the same time, i.e. in parallel. A major factor driving and sustaining this debate is the extreme difficulty of distinguishing the two possibilities in terms of the output they produce (Townsend Reference Townsend1971, Reference Townsend1976, Reference Townsend1990). The characteristics of one can be mimicked by the other. So the debate is necessarily more subtle and indirect, and more difficult to resolve, than might be expected in advance.

Parallel processing is suggested by findings that properties of a rejected analysis produce priming effects (Hickok Reference Hickok1993; Nicol and PickeringReference Nicol and Pickering1993), indicating that this secondary analysis is present, but it is difficult to rule out the possibility that the rejected analysis was produced before the one that was ultimately accepted, i.e. that the processing was actually serial. Meng and Bader (Reference Meng and Bader2000), noting that experimental evidence for serial processing was disputed, produced novel evidence for it, but noted that parallel models can handle their results if they include early cut-off of unpromising analyses, as occurs in the incremental-interactive model described above (see also Hsieh et al. Reference Hsieh, Boland, Zhang and Yan2009; Schlesewsky and BornkesselReference Schlesewsky and Bornkessel2003). There is also reason to think that the reanalysis mechanisms usually associated with serial models are necessary (LewisReference Lewis2000), but early cut-off again allows parallel approaches to remain viable: reanalysis is the resurrection of a rejected analysis after all else fails. Gibson and Pearlmutter (Reference Gibson and Pearlmutter2000) also explored possible ways of distinguishing serial from parallel processing, and further demonstrated the extreme difficulty of doing so. In more recent years, interest in this issue seems to have declined somewhat, perhaps because of the difficulty of resolving it, though it has not been entirely forgotten (see Farmer et al.Reference Farmer, Cargill, Hindy, Dale and Spivey2007; van Gompel and Pickering Reference van Gompel, Pickering and Gaskill2007; Novais-Santos et al. Reference Novais-Santos, Gee, Shah, Troiani, Work and Grossman2007; Clifton and StaubReference Clifton and Staub2008; Boston et al. Reference Boston, Hale, Vasishth and Klieg2011). The essential conclusion continues to be that no clear conclusion can be drawn.

3.3.1 Processing and linguistic theory

As described above in reference to the language module, we have adopted the position that the processors are the embodiment of UG principles. We thus follow Pritchett (Reference Pritchett1988, Reference Pritchett1992), Weinberg (Reference Weinberg1993, Reference Weinberg, Epstein and Hornstein1999), Crocker (Reference Crocker1996), and Dekydtspotter (Reference Dekydtspotter2001) in treating processing mechanisms and linguistic principles as the same entities. This point about ‘embodiment’ needs some elaboration. It does not mean that descriptions of linguistic entities which take no account of processing factors but which do serve as a source for formulating a processing-based approach such as MOGUL will provide the totality of the descriptive apparatus of the linguistic processors. For example, freed of any responsibility to processing concerns, a linguistic-theoretic account of the relationship between different levels of linguistic structure might include an elegant sequence of steps that provides a neat way of describing how one structure relates to another. Such time-free accounts are full of useful time-based metaphors like checking and movement. It is tempting to take the leap and claim that they also represent processing steps. The derivational theory of complexity in the early days of generative linguistics made just such a claim in attempting to associate transformation rules with processing routines (Miller and ChomskyReference Miller, Chomsky, Luce, Bush and Galanter1963). Unfortunately, this endeavour led to a dead end: a more complex transformation turned out to be more easily processed than a less complex one (Fodor, Bever, and Garrett Reference Fodor, Bever and Garrett1974: 369, cited in Marantz Reference Marantz2005). All that is claimed here is that linguistic principles govern the operations of the parser, not that a purely linguistic account is strictly equivalent to the psycholinguistic account of the parser.

Returning now to the related positions of Pritchett, Weinberg, Crocker, and Dekydtspotter cited above, we still differ from them in important respects. Most importantly, our proposal is a framework rather than a specific theory. As a consequence, we will not attempt to develop any particular aspects in anything like the depth that these more specific proposals offer on their chosen aspects of processing. We also follow linguistic theories in the nature of the representations in syntactic structures and the basic entities from which they are constructed. The implication is that linguistic theory can be used to fill in the details of the framework. It is important to stress here that MOGUL does not require a commitment to any specific proposal regarding these details. In our analyses we will often use the assumptions of minimalist and/or principles and parameters accounts, but alternatives are available, such as construction grammar. The essential point is that our approach seeks a unification of processing theory and linguistic theory.

3.3.2 Modularity and interaction

MOGUL is, as has been already made clear, a modular framework. The key to modularity in our approach is the nature of the interfaces, as described above. More specifically, we hypothesise that the functions of these interfaces are very constrained, consisting entirely of efforts to match activation levels in adjacent processing units. A high degree of modularity follows, though substantial influence from adjacent modules is allowed, as is necessary given the fact that phonological processing for example clearly does exert an important influence on syntactic processing. The hypothesis that interfaces are this constrained in their function may turn out to be overly strong, requiring adjustments to be made in the future, but the goal of parsimony makes it worthy of serious exploration.

3.3.3 Competition

The notion of competition plays a crucial role in MOGUL, as it does in other approaches to processing, determining which items from a store are selected for inclusion in a representation under construction. It takes on special importance in MOGUL because this selection process is the key to acquisition in our framework. The competition is based on the activation level of the candidate items, so we will delay further discussion until the character of activation in MOGUL has been explicated.

3.3.4 Incremental processing

Because the incremental nature of processing is generally considered an established fact, it must be incorporated in any approach to processing. In MOGUL, it is a natural consequence of the nature of the interfaces and the nature of the processors.

Interfaces constantly seek to match activation levels in the two stores they connect; they do not wait until the construction has been completed on one side. Indeed, the hypothesis of a mechanism for determining when a representation is ready for the next level would be an inherently undesirable complication. Incremental activation thus represents the natural default hypothesis.

The same considerations apply to the MOGUL processor. Its one purpose in life is to construct a coherent representation from whatever elements are currently active on its store. This activity is constantly shifting, partly as the result of the incremental activation coming from adjacent modules. The most natural conclusion is that the processor also works incrementally; i.e., the competition that occurs within each processing unit is not between completed candidate representations but rather between current possibilities based on the state of the store at any given moment. The alternative, that the processor waits for some indication that the adjacent module is finished with its work, would bring in considerable additional complications. Assuming this general feature of processors, it is not surprising to find that many researchers have argued that the parsing system tries to make an analysis quickly, incorporating each word in the developing structure as it becomes available rather than waiting for additional input to guide the incorporation (e.g. CrockerReference Crocker1996; Dekydtspotter et al. Reference Dekydtspotter, Edmonds, Fultz, Renaud, Iverson, Ivanov, Tiffany, Rothman, Slabakova and Tryzna2010; Fodor and InoueReference Fodor2000). Adoption of this idea in MOGUL does not appear to require an explicit principle, simply a recognition of the general nature of processors.

3.3.5 Activation

Activation plays a crucial role in the MOGUL framework, as it is activation that underlies the competition described above. An item's current activation level is its resting level plus any additional activation it has received during the current processing. Its resting level is determined by the extent of its past use (and possibly by innate specification in some cases; see below). Every item in a store has an activation level, resting and current, so the notions apply to features, values, and indexes, as well as the more prototypical contents of the store.

3.3.6 Dual storage and processing as a race

As described above, research has shown that processing involves both the retrieval of stored chunks and on-line construction and analysis, though large questions remain open regarding the details. MOGUL architecture and processing properly allows for both aspects of processing. It does not in itself answer the big questions about the details but rather offers a framework within which these questions can be productively addressed.

In MOGUL, the distinction is made possible by the compositional nature of representations, as described above. Each store contains simple elements and arbitrarily complex combinations of these elements. No utterance consists entirely of primitives, so in this sense every instance of processing involves the use of stored chunks. Considerable variation occurs, however, in the amount of on-line construction that can or must be done. In a sentence like Aging zebras seldom attack anthropologists, there are no apparent chunks above the word level, so a great deal of on-line construction will be necessary. But when the input can be analysed largely in terms of stored chunks, both types of processing can occur. A possible example is Criminal gangs hunt wild animals. Because of the common co-occurrence of the items wild and animals, and criminal and gangs (and perhaps hunt plus wild animals), these combinations are good candidates for storage and use as chunks.

In either type of case, PS activity results directly in activation of the component parts on SS as their PS counterparts become active (based on activity in the auditory system and therefore on AS). These active SS representations then activate any chunks in which they are included, without intervention from the processor. At the same time, the processors will try to integrate the various more basic elements as they become available. We will consider below the details of how each process occurs and how competition between them gets resolved, and in Chapter 4 we will discuss the development of complex representations, i.e. how they come to exist and become stable items in the linguistic stores.

In Chapter 2, we briefly discussed the ‘words and rules’ view of language promoted by Pinker and others (e.g. PinkerReference Pinker1999; Pinker and UllmanReference Pinker and Ullman2002). We can now do a bit more to clarify the status of this distinction in MOGUL. Rules, again, are the workings of the processors, while ‘words’ (stored items, actually) are representations in the stores. The two clearly interact, but the emphasis in the words and rules view is on their separability, particularly the case of past tense forms. In English the regular past tense form is the product of computation – a rule – while irregular forms are individually stored. The situation in German, the other extensively studied language in this respect, is somewhat more complex but shows the same clear split between computation and storage (ClahsenReference Clahsen1999; Marcus et al. Reference Marcus, Brinkmann, Clahsen, Wiese and Pinker1995).

In MOGUL terms, there is no past tense rule as such; the essential point is that computation is involved, rather than just retrieval of a stored representation (see Pinker and Ullman Reference Pinker and Ullman2002, for a similar clarification). The category T is innately present,² and development of the ‘rule’ consists primarily of the coindexing of this item with the appropriate PS and CS representations. The processors will then insert it into representations – during processing – in accordance with their own principles. In contrast, the use of a stored irregular form is based on direct activation of a complex representation; the processor is not directly involved. It is no surprise, then, that ‘words’ and ‘rules’ can be dissociated in the many ways Pinker and his colleagues have identified. Note that this is not to say that regular English past tense forms are never stored and accessed as wholes; they occasionally are, as described above. The point is that the computational route is available and is routinely used.

The distinction between retrieval of stored chunks and on-line analysis/construction is commonly seen in terms of a race, as described above. The distinction between the processor doing the work of constructing the overall representation on the one hand and a stored representation being automatically activated without its intervention is a competition between two processes, the outcome of which is determined by the speed with which each occurs. Is the complex representation activated before the processor has time to build a representation from the component parts as they become active? (The assumption here is that both routes will lead to a result that is compatible with current CS activity. If one has a substantial advantage in this respect, it might well be the ultimate winner of the competition even if its speed is below that of its rival.)

Competition based on activation levels is at the heart of the MOGUL approach to processing. The race idea suggests that one way to instantiate the competition is to treat activation level as the speed with which an item can become available for processing or can be incorporated in the representation under construction. We have described activation as a representation rising into working memory and thereby becoming available to the processor. On the reasonable assumption that this rise can occur at different speeds, depending especially on resting activation level, this conceptualisation can be readily translated into a competition based on speed. The framework we are proposing does not require an approach of this sort to the competition; other instantiations are quite possible. But its adoption is natural and has some good consequences, as we will show in Chapter 7.

We noted above Kuperman, Bertram, and Baayen's (Reference Kuperman, Bertram and Baayen2010) arguments that in order to account for characteristics of visual word recognition we have to adopt a multiple-route model and that the idea of a race between two processing routes is untenable. We will suggest that their findings, while interesting and challenging for many specific models, do not pose any problem of principle for the dual-route race approach or for its incorporation in the MOGUL framework. Within our framework, the sort of freedom of information suggested by Kuperman and colleagues is in fact the natural default. Processors work with whatever is currently active on their store. A principle that whole-form and decompositional processing must be separated and carried out in a specific order would be an additional complication, to be avoided if at all possible, as would a hypothesis of limits on the types of morphological information that can be used by the syntax processor. Empirical evidence against such stipulations is to be welcomed, as it facilitates the quest for parsimonious modelling.

More specifically, the conclusions drawn by Kuperman, Bertram, and Baayen (Reference Kuperman, Bertram and Baayen2010) were based primarily on two major insights that they drew from their data. The first was the presence of an interaction between characteristics of the whole word and characteristics of a suffix it contains. They confirmed the familiar finding that recognition time for complex words is influenced by their frequency (more frequent words are more quickly recognised) but also obtained the more novel finding that words with longer suffixes show a weaker effect of frequency than those with shorter suffixes. For very long suffixes the effect essentially disappeared. They attributed this finding to the relation between length and salience. Short suffixes are more easily passed over by the parser, meaning that the analysis is done entirely in terms of the whole form; its frequency then becomes crucial. The presence of a long, and therefore salient suffix pushes the parser to treat the word as a composite, with the result that the frequency of the whole form becomes unimportant. This relation between salience of the component part(s) of a complex word and the parser's tendency to rely on decompositional analysis was already well established in the literature (e.g., Sereno and JongmanReference Sereno and Jongman1997; Bertram, Schreuder, and Baayen Reference Bertram, Schreuder and Baayen2000; Järvikivi, Bertram, and Niemi Reference Järvikivi, Bertram and Niemi2006).

This finding of Kuperman and colleagues and the analysis they offered for it are straightforwardly incorporated in the MOGUL framework. An affix that is salient (meaning long, in this case) is one that is relatively active. If an element is active, the processor will seek to incorporate it in the representation it is constructing. In other words, longer affixes encourage decompositional analysis, making the frequency of the whole form largely irrelevant. Non-salient (short) affixes are those with low activation levels. A whole form competing with them is therefore more likely to triumph, its prospects of success directly varying with its own activation level, which is a function of the word's frequency. It is no surprise then that longer suffixes attenuate or entirely cancel the effect of the complex word's frequency.

The second insight that Kuperman, Bertram, and Baayen (Reference Kuperman, Bertram and Baayen2010) derived from their findings was that an interaction occurs between characteristics of the morphological family of the base and the morphological family of the suffix. The morphological family of a morpheme is the set of words that include it. The family of the English base act, for example, includes active, action, activity…while the family of the suffix –ness includes happiness, sadness, quickness…(in other words, family size represents the productivity of the affix). Kuperman and colleagues found that recognition time is minimised when the family of the base and the family of the suffix are of similar sizes and rises substantially when an imbalance occurs between the two. They were very tentative in their efforts to explain this relation. One possibility that they considered involved activation spreading from the components to their family members and then feeding back to the actual components, thereby enhancing their activation. The larger the family, the larger the enhancement would be.

This type of account is quite natural within the MOGUL framework and, within that framework, appears to handle the findings. If both the base and the affix are highly active, as a result of support from their families, the processor will make use of them, readily combining them to make the ultimate representation of the input. In this case the computation route will be maximally successful and so recognition time will be relatively low. If both base and affix have low activation levels, the whole form will have only weak competition. The whole-form route is then maximally successful, again resulting in relatively low recognition times. But if one component of the complex word is highly active and the other is not, conditions will not be good for either route and recognition times will be higher as a result. So there does not appear to be any problem for the adoption of a dual-route race approach in the MOGUL framework, with the understanding that interaction occurs between the routes. We should note, though, that the findings of these studies and their analysis are quite complex, dictating some caution in our conclusions.

3.3.7 Serial vs. parallel processing

In MOGUL there is only one syntactic store, and a representation written on it (currently active in a dominant manner) must ultimately be coherent, so in an important sense processing as we have described it is serial. But it involves simultaneous activity by a number of subprocessors, each seeking to construct that representation in accordance with its own nature and the current state of SS. There is no overall control mechanism but rather an ongoing competition among processors to shape the representation their own way. So in another important sense there is an element of parallel processing. Early cut-off of some possible analyses is inherent in the approach, as a budding representation on one store that is inconsistent with activity in an adjacent processing unit will for that reason be challenged and frequently terminated. The parallel activities of the various subprocessors will often result in the temporary activation and use of elements that do not appear in the ultimate representation, so some priming of such elements should be expected, as has been observed in the experimental literature.

3.4.1 The process

During comprehension, potentially relevant items at PS have their activation levels raised by the interface connecting phonological structure to auditory processing – the level of auditory structure (AS), which we will develop in more detail in Chapter 5. The PS–SS interface then raises the levels of the corresponding items in SS, leading the SS–CS interface to do the same with their CS counterparts. Each type of information about the selected word is thus made available for processing at the appropriate level. Throughout, the three types are kept in registration through their common indexes. The process is incremental; i.e., when an item in one module is activated the interface will promptly activate any items coindexed with it in the adjacent module, without waiting to see what else occurs.

Spreading activation within a store is also relevant. Those items that share features with items already activated will also experience a rise in their current activation levels, the degree of the change depending on the number of shared features, the current activation levels of those features, and the resting level of the items being activated.

To construct its representations, a processor uses the items in its store that are most active at the moment. There are, again, various ways to realise this idea, and the one we will tentatively favour involves the speed with which an item becomes available for use in the construction of representations during processing (equivalently, the speed with which it enters working memory). The higher an item's current activation level the more visible it is to the processor. The implication is that items that attain high levels most quickly have the first opportunity to enter the representation under construction and those that lag behind may well be excluded because an adequate representation has already been completed and in effect passed on to the next module before they become available. Resting level is of great importance in this process as it represents the starting point for each of the competitors.

An item that has become sufficiently active to play a role in the current processing will not necessarily be included in the ultimate representation. In other words, the psycholinguist's distinction between activation and selection is relevant here. Whether an active item is selected depends on how well it fits into the representation being constructed and the closely related question of how long it maintains its elevated current activation level and therefore continues to be available for processing activity. The processor seeks to make a coherent representation from the currently active items. An item that receives no continuing stimulation will fall in current activation level and thereby remove itself from the current processing activity.

The construction of the SS representation involves multiple syntactic subprocessors, each trying to make the representation fit its own requirements. The ultimate representation is thus a kind of best fit among those requirements. But adjacent modules also play a role in this process. PS is obviously relevant because of its role in initial activation of SS items. The PS–SS interface probably also seeks to maintain those activations, as long as the PS representation lasts.

The role of CS is no less important. Throughout the process of the syntactic processor constructing a representation in SS, the SS–CS interface activates the CS counterparts of whatever SS representations are active, incrementally, and conceptual processors seek to construct from them a suitable representation. This process necessarily involves activating, or further activating, certain existing CS representations and allowing others to fall back toward their resting levels (in effect deactivating them). It includes the effect of on-going CS activity resulting from context, continuing non-linguistic input, and demands of conceptual processors based on their own nature. Throughout, the SS–CS interface seeks to balance activation levels of coindexed SS and CS items. Where the embryonic representation on CS conflicts with that on SS the result will be a challenge to the latter (and the former). The SS–CS interface, constantly seeking to match current activation levels, will therefore raise the levels of other SS items while withholding its support from those making up the incompatible representation. This conceptual interference in the syntax opens the door to other SS items that were not initially included, items that might lead to an SS representation that is more compatible with CS activity.

Thus, the process is simultaneously modular and interactive in roughly the sense of IIT. The syntax processor constructs a representation on SS based on its own principles, but when a representation being constructed creates problems for conceptual processing the SS–CS interface will interfere with its construction.

To this point we have focused on comprehension, but the same principles apply to production. In fact, production in MOGUL is simply comprehension in reverse. A message to be expressed forms as a representation in CS, not differing in any principled way from the message representations that are the end product of comprehension. The SS–CS interface activates SS representations coindexed with active elements in this representation and the syntax processor seeks to construct from them a coherent syntactic representation. The PS–SS interface activates PS representations coindexed with the active SS representations and the phonology processor seeks to construct from them a coherent phonological representation. Throughout the process, activation spreads within each store just as it does in comprehension and the interfaces constantly seek to reconcile the representations on each side of them in terms of activation levels.

3.4.2 An example

Consider now an example of processing in MOGUL. The input sentence we will consider first is the following:

Again, we will have little to say about the phonological aspects and not much more to say about semantics, our main attention being on the syntax.

Comprehension begins with activity in the auditory system, which stimulates (via an interface) various PS representations, from which the phonology processor seeks to construct a coherent overall representation. This construction includes drawing word boundaries in the input received from auditory structures, i.e. activating the PS representations of the four words in the sentence plus the past tense affix. Activation of these PS representations incrementally activates coindexed SS representations, from which the syntax processor attempts to construct a coherent overall representation. The details of these representations – that for the sentence as a whole and those for its individual components – depend very much on the particular linguistic theory one adopts, but they will certainly include the syntactic categories of the words, plus Tense.

Activation is not likely to be a simple matter of PS items activating corresponding SS items, though. Subcategorisation frames, in particular, exert an important top-down influence on the process. Activation of the SS representation of kick leads to activation of any frame of which it is a part, which activates the other syntactic elements of that frame. Thus, the [__NP] frame will be activated and therefore the NP representation that it contains. Similarly, any items that collocate with kick and therefore appear with it in complex representations will undergo an increase in current activation level. This will presumably include ball. Those SS representations activated in this way will then activate their PS counterparts, which might or might not be those already activated via stimulation from auditory input. The result is a matching process between SS and PS elements.

The same sort of matching process occurs between SS and CS elements. The latter are activated directly by the SS/CS interface in response to activation of their SS counterparts. Some SS items other than those associated with Ron, kicked, the, and ball will inevitably be activated, by spreading activation from the appropriate items and possibly by noise at PS. The CS representations coindexed with them are stimulated as a result, though probably not as strongly as RON, KICK, PAST, DEFINITE, and BALL. The conceptual grid of kick is activated along with KICK, in much the same way that its subcategorisation frame is activated at SS, and its role in processing is also comparable. The conceptual grid includes an agent and a patient, which must be matched (coindexed) with active SS items. The active status of the SS representations of Ron and the ball allows a successful matching and a mutual reinforcement of the SS and CS items. The Case items at SS must also match with coindexed conceptual role items at CS. The Cases, obligatory in the presence of the heads that assign them, are combined with the NPs that are active. A parallel process must occur at CS, the conceptual role items coindexed with the Case items combining with the CS representations that are coindexed with the SS NPs.

Contextual influences also play a role in construction of the CS message representation. Sensory input is constantly available to conceptual processors and therefore constantly influences CS processing. To most American listeners, the linguistic processing of the word football would probably activate the representation of an American football much more strongly than that of a European football. In this case, all other things being equal, the CS representation of an American football would win the competition and enter the message representation, as shown in Case 1 of Fig. 3.1.

Figure 3.1 Processing Ron kicked the football: three different cases.

But if the person is seeing a European football while hearing the sentence, this visual experience will strongly activate its representation, giving it a large and no doubt overwhelming advantage in its competition with the otherwise dominant AMERICAN FOOTBALL. The ultimate message representation will therefore include a European football rather than an American football, as in Case 2.

Another possible situation is that in which the word football was not heard clearly, due to such factors as noise, an unfamiliar accent, slurred speech, or a distraction. This situation is depicted in Case 3 of Fig. 3.1. Here a number of different CS representations will initially receive relatively weak activation based on the alternative PS representations activated by the unclear sound appearing at AS. These would include both AMERICAN FOOTBALL and EUROPEAN FOOTBALL but also other CS representations coindexed with PS representations that resemble the sound. If the hearer is simultaneously seeing a European football, as in Case 2, this visual experience would again make EUROPEAN FOOTBALL dominant and lead to its inclusion in the message representation. Once it has received a high current activation level and has been inserted in the overall representation, the SS/CS interface will stimulate the SS of football, leading to its insertion in the overall SS representation, after which the PS/SS interface will stimulate the coindexed PS, with a parallel result. In other words the ultimate outcome of the processing will be a PS–SS–CS chain of coindexed representations. This last point has important implications for the characterisation of development in the MOGUL framework, to which we will return in Chapter 4.

In these examples perceptual experience led to the activation in CS that established EUROPEAN FOOTBALL as the winner of the competition. An alternative source for the additional activation would be nonsensory sources. If the person already knows that Ron is a professional soccer player, for instance, or the linguistic context already contains clear information as to the identity of the game, then any representations associated with soccer will have temporarily elevated activation levels.

In more familiar terms, the additional information leads to the activation of a soccer schema, which in MOGUL is simply a complex CS representation, functioning in the same way as any other. It can be activated by the linguistic label associated with it (i.e. the PS–SS coindexed with it) or with one or more of its component representations, or by sensory activation of one or more of its components, or by purely system-internal processing – thinking, imagining, free associating. When it is activated, the current activation level of each of its component representations is also elevated as a result, including that of the particular type of ball, in this example.

Such contextual influences will play an especially large role if the activation coming upward through the language module is weak, perhaps as the result of a degraded signal. In other words, top-down processing becomes especially important when bottom-up processing is inadequate or unreliable. If linguistic processing does not lead to activation of RON, for instance, visual input (a view of the kick) might take its place, activating RON and thereby setting up its incorporation in the message being constructed.

Production of this sentence is essentially the same process in reverse. A CS representation of the concept forms as a result of activity at CS, based on current sensory input interacting with the current state of the store, including the stored representations and their current activation levels, the latter reflecting to a large extent their resting levels. The most active elements will be RON, KICK, PAST, BALL, and DEFINITE, though other items will certainly have elevated levels as well, some because of their relation to these (items in KICK's conceptual grid, especially) and others as a result of other perceived or remembered activity. Possible examples are conceptual items capturing specific aspects of the kick, such as its quality, its outcome, its location, etc. The question of how the contents of an utterance are selected is of course extremely complex and our intent, again, is to provide a framework within which such questions can be studied rather than a particular proposal.

Activation of these CS items will lead the SS–CS interface to stimulate the coindexed SS items, which will result in the PS–SS interface stimulating any PS items that are coindexed with them. On SS, the active items will include some that are inappropriate, as a result of both the related CS activity and spreading activation from the SS counterparts of the five appropriate CS items. Under normal circumstances, the items that correspond to the CS components of the intended message will be more strongly activated than any others, but the collateral activation opens the door to errors. The activity on SS and PS prompts each processor to construct a coherent representation that includes the most active items. The ultimate outcome, within the language module, should be a coherent PS representation comparable to that which was produced during comprehension of the sentence Ron kicked the ball. The presence of this representation on PS then triggers activity by motor processors (via an interface) to produce from it a spoken (or written or signed) utterance.

There are of course a great many details we have not dealt with in this example, regarding both linguistics and processing. The logic, again, is that MOGUL is intended to be a relatively high-level approach, allowing various specific instantiations. It is, in other words, a framework rather than a specific theory of language or language processing.

3.4.3 Another example: input including a fixed expression

A complication is added when a fixed expression, such as an idiom, is part of the utterance being comprehended, so we will present a second example to illustrate this somewhat altered situation.

In most respects the same processes are involved as those described above, so we will disregard much of the processing and focus on the differences. The idiom, kick the bucket, is comparable to a word in that it consists of a PS–SS–CS chain, where the CS is DIE. The primary difference, for our purposes, is that the SS is more complex, composed of the SS representations of the individual words combined to form a verb phrase.

During comprehension, the PS string is activated, starting with /kick/. Activation of /kick/ spreads to the PS of the idiom, which contains it. Activation of this composite PS representation then spreads to /the/ and /bucket/, reinforcing the stimulation they are receiving directly from the auditory signal. This more direct stimulation will further activate the /kickthebucket/ representation of which they are components. Activation of /kick/ also activates the SS of kick and this activation spreads to the SS of the idiom. At the same time, activation of /kickthebucket/ also activates its SS.

As the SS representations corresponding to kicked, the, and bucket become active, the syntax processor seeks to make a combined representation from whatever is available to it at the moment, following normal in-built principles. But if the idiom representation that contains them is active before this work can be done, the effort will be abandoned in favour of a synthesis of this representation with the rest of the input, namely the SS component of Ron and any additional SS elements that have been activated as part of the process (functional categories, phrasal categories, intermediate categories). Whether this happens will depend largely on the resting level of the idiom representation (see below). CS activity also plays a central role. Active CS representations that are consistent with the idea of dying raise the activation level of the idiom's CS, DIE, and therefore its SS and PS representations. Those that are consistent with kick or bucket raise the activation level of the component parts. So if the context of the utterance supports the literal interpretation and not the idiomatic reading, the latter will be abandoned in SS in favour of a construction process using kick, the, and bucket as independent items.

3.4.4 Processing as dynamic equilibrium

It is in the nature of a processor to try to make a coherent representation (coherent in its own terms) from the current state of its associated blackboard. Whenever representations (coherent or otherwise) are active on a blackboard, the processor tries to use them to construct its own representation there. Interfaces respond to input on boards they can read by trying to establish a match between the activation levels of the representations on that board and the other boards they can access. The tendency toward a dynamic equilibrium, expressed by a unified set of representations, follows.

As part of this dynamic equilibrium, strong internal demands at one level can force other levels to adapt. The order of items in a PS representation is a good example. In comprehension, the phonology processor sticks strongly to the order that comes from below, so syntax has little choice but to construct a representation that respects this order. (This can probably be analysed as an artefact of speed and incremental processing, but we will not explore this point here.) Crain and Steedman's (Reference Crain, Steedman, Dowty, Karttunen and Zwicky1985) idea of semantic rejection of syntactic representations during the construction process translates into the idea that the conceptual processor is doing things on CS with the current input that are not compatible with the current state of SS – in the sense that the set of active items at one level does not match that at the other – so the interface tries to reconcile them. This reconciliation often takes the form of raising the activation levels of SS items that are not yet part of the in-progress SS construction. The syntax processor then has different input and is therefore likely to construct something a bit different. The influence could go in the other direction as well: if the demands of SS are stronger than those of CS, the set of active CS items will be altered by the interface.

One implication of the dynamic equilibrium view of processing is that the system does not know or care if it is involved in comprehension or production. Whenever input enters, from above or below, the whole system seeks to create a stable, coherent, unified set of representations for it. ‘The system’ here can be taken as the language faculty or the whole cognitive system or other portions of it. Production in the model involves the same items and mechanisms as comprehension, consistent with findings that activation of a syntactic structure in comprehension can prime it in production (e.g. Branigan, Pickering, and Cleland Reference Branigan, Pickering and Cleland2000; Pickering et al. Reference Pickering, Branigan, Cleland and Stewart2000; Cleland and Pickering Reference Cleland and Pickering2003; Pickering and FerreiraReference Pickering and Ferreira2008). Similarly, Kempen, Olsthoorn, and Sprenger (Reference Kempen, Olsthoorn and Sprenger2012) recently provided evidence that production and comprehension make use of a ‘shared grammatical workspace’; in our terms, SS is used in both directions. When a CS representation is active, the SS/CS interface similarly activates the SS items that are coindexed with the component CS items, leading the PS/SS interface to do the same with the coindexed items in PS. Each processor works to build a legitimate representation, in its own code, by connecting some of these items along with any others it needs in order to produce an acceptable representation. In the process, it allows the current levels of items that turn out to be unsuitable to fall back to resting levels. The end result is that exactly those items that are ultimately selected will stand out in terms of activation. This relatively simple picture of processing will become more complex in Chapter 5, where we discuss the role of representations outside the language module.

3.4.5 A note on neurological plausibility

Given our frequent use of ‘activation’ and associated terms, it is worth emphasising at this juncture that we are not talking about neural architecture. The neural instantiation of structures and processes described within MOGUL represents a different level of description. However, in the same way that a Jackendovian account facilitates the integration of purely linguistic explanations and real-time processing, so we hope that processing accounts such as the one we propose in MOGUL will be open to a useful degree of harmonisation with explanations of neural functioning. It is still important, however, that the levels be kept conceptually distinct. To take one example, what is explained as varying degrees of ‘activation’ in psychological processing terms may require an account at the neural level that involves both activation and inhibition (GreenReference Green1998; but see Finkbeiner et al. Reference Finkbeiner, Almeida, Janssen and Caramazza2006). In other words, the two-way distinction may be crucial to explain neural activity that subserves psychological processes but perhaps unnecessary when describing those processes themselves. Put another way, in describing the functional architecture of processing, we do not have to deal with the specifics of neural activity: the desire to harmonise the two levels of description as much as possible certainly does not imply any simple equation between psychological and neurological processes. Even the networks posited by connectionists are not neurologically plausible in a way that the term ‘neural network’ implies. In other words, neurological plausibility is a relative concept.