This book is about network linguistics, based on the idea that language is a network.Footnote ^[1] In a sense, this is an old idea, going back at least to the early structuralist dictum that a language is a system where everything holds together (‘La langue est un système où tout se tient’, attributed to Meillet – Peeters Reference Peeters1990). But the search for structure rapidly turned into a search for compartments and modules, resulting paradoxically in the very opposite of what Meillet seemed to have in mind – an analysis focusing on boundaries rather than on interconnections. For most modern linguists, the idea that language is a network seems like a denial of structure. After all, an attractive metaphor for networks is a plate of spaghetti, and who wants to think that language is like that? In contrast, modern theories offer information packaged neatly into units called constructions, phrases or lexical items, which in turn are stacked up neatly in compartments such as morphology, syntax and the lexicon.

Rolf Kreyer has written a worthwhile contribution to the growing literature on network linguistics, which in modern anglophone linguistics started with Stratificational Grammar, later named ‘Neurocognitive Linguistics’ (Lamb Reference Lamb1966, Reference Lamb1998), and includes my own Word Grammar. So in this review I have to declare a strong interest: I’m a believer. I believe that language is, indeed, a network, and that this network is, indeed, a structure. Many readers may protest that they too see language as a network; after all, cognitive linguists envisage ‘an elaborate network comprising any number of conventional units linked by categorizing relationships’ (Langacker Reference Langacker, Barlow and Kemmer2000: 12) or a ‘network of constructions [which] captures our grammatical knowledge of language in toto, i.e. it’s constructions all the way down’ (Goldberg Reference Goldberg2006: 18). But notice that in these cases the complex units which the network connects have their own internal structure which is not part of the network. Similarly, there are network-based theories of individual areas, such as Network Morphology (Corbett and Fraser Reference Corbett and Fraser1993, Brown et al. Reference Brown, Corbett, Fraser, Hippisley and Timberlake1996, Fraser and Corbett Reference Fraser and Corbett1996), and there are network-based theories of how usage affects storage (Bybee & Beckner Reference Bybee, Beckner, Heine and Narrog2010). But network theory goes further by claiming that ‘it’s networks all the way down’.

As Kreyer explains, one of the attractions of network linguistics is its orientation towards neurophysiology, where everyone agrees that the neurons in our brains form a network. This leads Kreyer to his ‘neurophysiological challenge’ (2):

Sooner or later, linguistics must meet up with its neighbouring disciplines including neurophysiology, so the more compatible with the neighbours our current models of language are, the better. Better still, as Kreyer also shows, the network model is just what we need to accommodate a lot of the facts about language that are becoming increasingly hard to ignore, and that are hard to reconcile with the more familiar boundary-based models. For example, a neural network carries activation levels which help to explain why frequency affects the accessibility and ‘entrenchment’ of the various bits of language that we know; and on a much shorter time-scale, these same activation levels explain the priming effects so familiar in cognitive psychology. Of course it could be objected that linguists can safely leave such things to their colleagues in psychology and psycholinguistics; but the point is that activation follows structure, and structure is our province, not theirs. And finally, network linguistics offers a real chance of showing that language is just like other bits of cognition, without any of the special apparatus that some linguists like to imagine.

The real challenge, of course, lies in the detail. What is this network like? What kinds of nodes and links are there, and how do they interact? In addressing these questions, we are not just passive readers of research by neuroscientists because our research complements theirs. They can tell us about the physiology of neurons in general terms, but not about the linguistic detail where we are at our strongest. The goal of network linguistics, then, is a theory which meshes on one side with the generalities of neuroscience and on the other with the details of theoretical and descriptive linguistics. This requires not only good theory, but also some knowledge of neuroscience and descriptive linguistics, and good modelling skills – a serious challenge for the 21st century.

How, then, does Kreyer fare in relation to his challenge? Since my Word Grammar (WG) is one of the theories against which he reacts, this is the standard against which I will measure his work. Surprisingly, perhaps, Kreyer does not name his theory, so I shall refer to it simply as the Kreyer theory (KT).

Starting with their similarities, WG and KT both go well beyond linguistic theory as such, by offering an elementary theory of mental structures which can accommodate what we know about language, not only in terms of how we store it but also in terms of how we learn it and how we use it. This being so, both theories have two parts: a cognitive part, consisting of a general theory of cognition, and a linguistic part, where this theory is applied to the particular case of language.

The cognitive part of KT is presented as a symbolic network, both in the sense that each node corresponds to (‘symbolizes’) one concept and also in the sense that each node in the diagrams is labelled. As in WG (following Lamb), the labels are simply a presentational convenience; but unlike WG, Kreyer hints that ‘some concepts are represented by a single node ... while others are represented by the co-activation of a number of nodes’ (6). For example, while ‘pronoun’ is a single category, there is no category for ‘plural demonstrative pronoun’ (i.e. for these and those) (114). Instead, these two words are picked out by the co-activation of ‘pronoun’ with the features (not categories) ‘plural’ and ‘demonstrative’. The claim may be right, but the only evidence for it (115) is that it produces a short hierarchy, with these and those immediately under ‘pronoun’ rather than separated by several nodes such as ‘plural pronoun’ or ‘demonstrative pronoun’. There are two arguments for this analysis: that it avoids difficult choices about the relative ranking of competing contrasts, and that it is more efficient in terms of access time. The arguments are presented too briefly to evaluate properly: for instance, it’s not clear why ‘plural’ and ‘demonstrative’ can’t be categories, or why they can’t apply simultaneously to define an intersecting sub-category.

Another major cognitive issue is how activation spreads through the network. The specific issue raised by KT concerns inhibition: can the activation in one node inhibit the activation in another? KT recognises inhibitory links as well as excitatory ones, whereas WG only recognises excitatory links. Maybe KT is right, but again it deserves more discussion. It’s true that inhibition is well documented at the level of neurons, but the strongest neurocognitive evidence for inhibition seems to lie in perception rather than in cognition, so we need a great deal more debate before we assume that inhibition is available in a cognitive network. This matters, because inhibition plays a major role in KT, and the theory even allows one node to inhibit the inhibitory link from another; for example, inhibition of inhibition is used to handle non-default word order (149).

A further point on which KT and WG agree is the importance of activation: linguistic elements must be able to carry activation. This is obvious in a theory of language use, but even a model of language structure will eventually have to accommodate the effects of frequency and recency. But KT demands much more work than this from activation; for example, the ordering of elements (whether phonemes or words and phrases) follows activation levels, on the assumption that the most active element is realised first (153). The trouble with this kind of claim is that the model is so complex that claims are hard to prove without some kind of demonstration machine such as the computer model that Kreyer himself calls for (268). In the present case, the complicating factor is that activation spreads from all sorts of sources, including the context, so even if we want element A to precede B, we can’t be sure that the word-order pattern will achieve this by assigning more activity to A because B might receive even more activity from some other source.

Another point of agreement is the very small vocabulary of elements and relations, but the vocabulary for KT is even smaller than for WG. In particular, KT does not recognise the basic relations which are so fundamental to WG: ‘isa’, ‘argument’ and ‘value’. According to KT, ‘all links are of a similar kind’ (47). One could object that KT does distinguish excitatory and inhibitory relations (as noted above), but this would miss the real challenge of the KT claim: since the brain uses only one kind of associative link, the same must be true of a cognitive model. When applied to language, the result is an analysis in which the various categories and units are simply associated with one another; for example, in the sentence John reads books, the word John is linked to ‘Agent’, which is linked to ‘Subject’, then to ‘sing.sub’ and ‘3rd.sub’, which are both linked, via the word reads, to ‘read’ and ‘verb$_{\text{sem}}$’ (151). More traditional analyses would recognise fundamental differences between argument roles, syntactic relations, word-classes, meanings, realisations and so on; but KT claims that these distinctions can be dispensed with. Kreyer doesn’t develop this claim seriously, but the issues are clearly important, and if he is right we need to pay attention.

The purely linguistic part of KT combines an unusual mixture of approaches, including Halliday’s Scale and Category grammar (74), the Quirk/Aarts theory of gradience (99) and Goldberg’s Construction Grammar (185). The bits sometimes conflict – for example, having introduced Halliday’s rank scale Kreyer also insists on the importance of ‘rank permeability’, where the rank scale is irrelevant (194). Moreover, as was evident in the discussion of John reads books, the actual analyses offered don’t follow any of these theories. Unfortunately, Kreyer also ignores the WG arguments for dependency structure, so the linguistic analyses are generally conservative, and he faces the usual range of syntactic problems. For example, he claims that the meaning of a ditransitive construction resides in the entire construction, not the verb (191), only to admit four pages later that the participant roles (i.e. the meaning) are provided by the verb (195). It is true that verbs such as BAKE aren’t inherently ditransitive but can be used ditransitively (as in He baked her a cake), but a grammar has to explicitly allow this extension somewhere, so simply providing a ditransitive construction isn’t enough. The extension is an extension of the verb’s valency, so why not invoke a valency change?

In short, although I approve whole-heartedly of Kreyer’s enthusiasm for network linguistics, I’m not yet persuaded that KT offers a better direction than WG for future research. However, if we see the version of Kreyer’s theory described here as ‘work in progress’, it is very promising. I have criticised the very limited assumptions about relation types, but starting from scratch, with the minimum of cognitively plausible tools, is a good research strategy, provided there is also a testing strategy for recognising counter-evidence which requires some kind of enrichment of the apparatus. If Kreyer can find a convincing testing strategy, the results will certainly be interesting.