Constructions and creativity

1 Blending

Research on blending per se and by that name began in 1992. The first presentation occurred in 1993 and the first technical report appeared in 1994 (Fauconnier and Turner 1994). A quarter-century after that first technical report, several million words at least have been published on blending, in nearly every field that studies human affairs: art, music, religion, law, decision-making, mathematical insight, scientific discovery, AI, dance, fashion, advanced tool use, advanced social cognition, creativity in all fields, and so on. Blending has been the core subject of several major longitudinal scientific grants in various fields in North America, South America, Europe, and Asia. Most publications on blending do not treat language at all. An attempt has been made to list major publications at http://blending.stanford.edu, but the field is too large to survey. For an overview, see especially Fauconnier and Turner (2002) and, secondarily, in chronological order, Turner (1996), Fauconnier (1997), Turner (2001), Fauconnier and Turner (2008), and Turner (2008, 2011, 2014).

Principal findings of blending theory include the following:

1. While rudimentary blending seems to be available to several species, cognitively modern human beings are a step up the cline. Human command of advanced blending is indispensable to human creativity. The small extra step up the cline makes a vast difference in ability and products.

2. For cognitively modern human beings, blending is not at all costly. On the contrary, it is a constant and almost entirely invisible operation. The cognitively modern human mind is built to attempt to blend when possible. Blending is opportunistic. Nearly all blending attempts fail almost immediately by failing to meet the constitutive requirements of building a blending network. The relatively small number of attempts that survive rarely attach to a purpose or survive optimality constraints on blending. Few of those actually penetrate to action (such as expression). Few of those ever enter into consciousness. Very few of those achieve uptake by the community. Very few of those become entrenched for a community or a culture.

3. Blending is not a mental pattern with a fixed template or sequence. It is instead a principled mental operation that can run across any cognitive network that a human being is activating. Very many different kinds of blending products have been studied. There are some standard generic patterns of blending that arise repeatedly, but they are only standouts in the general landscape.

4. Blending work can be done anywhere in the blending network at any instant in the construction of the network, including in developing various spaces and developing vital conceptual relations between them.

5. Blending networks recruit from stable structures and can themselves become stable, widely-shared, and entrenched.

6. Blending networks create emergent structure not only in the blended space but throughout the network.

7. Blending plays a crucial role in everything we tend to view as distinctly human, and no analysis of higher capacities will be adequate if it leaves blending out of account. A theory of art, mathematics, science, religion, spirituality, learning, phylogenetic, human neurobiology, cultural evolution, etc. that leaves blending out of account is fundamentally flawed. But in the great scheme of things, the role of blending is still very small. Socially and consciously, we are often aware of blending phenomena, which we use for social distinctions, technological advances, aesthetic experience, spirituality, and so on. Accordingly, we notice blending in some cases. But billions of years preceded the relatively recent evolution of advanced blending; elaborate social and demographic work is needed to exploit the capacity fully. It is easy to make the mistake of discounting the vast run-up to advanced blending by falling for the “Last is Everything” fallacy. Advanced blending seems to be evolutionarily recent, but it is very small relative to what it built upon.

8. Although advanced blending comes last in this evolution of mental operations, it is not a linear addition but on the contrary changes the way we deploy all those other mental operations.

9. In the great scheme of things, blending is just one of very many important mental operations. Blending research is not a reductionist program. There is no attempt to argue that everything is a variety of blending, or that a particular mental or social achievement is explained by blending tout court.

Examples used to give an elementary taste of blending need to be recognizable as blends, if the examples are to be pedagogically effective. Consequently, these examples are typically pyrotechnic. Such pyrotechnic examples are useful, but insidious, because their visible flashiness misleads students into assuming that blending is a special, pyrotechnic, unusual, and visible operation instead of what it is: the basic harmony of the way we think. Blending is almost entirely invisible at first glance.

A chestnut example where the blending is visible is “the riddle of the Buddhist Monk.” A Buddhist monk in the pre-dawn light, standing for a while at the foot of a mountain path that leads to the summit, decides to climb the path. He begins at dawn walking up the mountain, reaches the top at sunset, meditates at the top overnight until, at dawn, he begins to walk back to the foot of the mountain, which he reaches at sunset. Make no assumptions about his starting or stopping or about his pace during the trips. Riddle: is there a place on the path which the monk occupies at the same hour of the day on the two separate journeys?”

Many people, including trained mathematicians, fail to see the answer quickly or at all. But separate the action into two mental spaces, one with the monk ascending and one with the monk descending, both from dawn to dusk, and then blend them. Think of two videos being superimposed. In the blend, there are two monks, one at the foot of the path and one at the top, and they begin traversing the path at dawn and finish traversing the path at sunset. We complete the blend with the knowledge that two people who traverse a path in the same time interval in opposite directions must, so long as they stay on the path, meet. Given the way the blend is assembled, the place the monk occupies at the same hour of the day on the two separate journeys is the spot where “the monk” “meets” “himself” in the blend. There is selective projection to the blend (calendrical dates are not projected to the blend, nor is the intentional state of someone who actually encounters himself, etc.) and there is emergent structure in the blend (two monks, a meeting). The two paths are fused, as are the two dawns and the two dusks. There are many complexities to how the blend is assembled. Some of them are reviewed in Turner (2012), which includes three mathematical proofs, all of them launched by blending of standard types deployed in mathematics, as reviewed in Alexander (2011) and Lakoff and Núñez (2000).

A blend is a “mental space” in a conceptual network of potentially many such mental spaces, in a potentially highly dynamic network, and helps to manage that conceptual network, including its construction and dynamism. A slightly less pyrotechnic chestnut example of blending is the “miserable Bay Area stockbroker.” Consider Bill and Peter, brothers-in-law, each happy. Bill is a mathematically-talented professor in the Eastern time zone who likes both investing and San Francisco. Peter is a stockbroker in San Francisco. Bill, in the East, wonders, analogically, should he move to San Francisco and be a stockbroker and get a huge raise? No: Bill is a night owl but Peter must arise at 5:30 AM Pacific Time to pull himself together for an hour to then deal with the stock market's opening at 9:30 AM Eastern Time. Peter likes the morning but Bill hates it. Mentally, Bill has created a new, blended person—Bill-as-Peter, who is miserable. Everyone is happy in the inputs but the central structure of the conceptual network is the miserable stockbroker in the blend. That misery is emergent structure. The selectivity of the projection is crucial. The stockbroker in the blend is not married to his own sister, even though the person in the blend is a stockbroker married to the woman who is in fact the sister of the stockbroker in the inputs.

A shockingly creative example which at first almost nobody takes as involving blending is the chestnut case of the cyclic day (Turner 2014). In our experience, there is actually just one day and then another day and then another day and then another day, in a sequence that stretches out indefinitely, forward and backward. The days in that sequence are all quite different. They do not repeat. If we woke up today and it was exactly the same as yesterday because it was in fact the same day in every detail, we would be sure we had lost our minds. And then it would not even be the same day, because yesterday we did not think we had lost our minds. Day after day after day, indefinitely, with all those differences between days, is too much to comprehend, too much to fit inside working memory, too much to carry around and manage. It is not mentally portable. So, we blend these different days into a conception of a cyclic day. We do this by using a very common general blending template. There are analogies and disanalogies across the different days in our experience. The analogies are packed into one thing in the blend: the day. The disanalogies are packed into change for that thing: The day is cyclic; it starts over every dawn and repeats. No one of the individual input days to the cyclic day blend is cyclic; no one of those individual input days repeats or starts over. But because of blending, all the days that have ever happened or will happen can be packed into a single idea, a tight, tractable, manageable, human-scale idea — the cyclic day, which repeats. Thinking of the cyclic day, we can say, “dawn is coming around again,” “it's time for my morning coffee,” or “this park closes at dusk.” We know how these words and concepts apply to the compact blend, and we can unpack from that blend to any parts of the mental web that interest us, to any day at any time, and even to a tractable stretch of days. The cyclic day is a compact touchstone, a congenial blend for thinking about the vast sequence of days. This vast sequence of days is of course itself too big for working memory. The compact blend — the single cyclic day, a new idea — makes it possible for us to work with concepts of time that stretch far beyond what we would otherwise be able to manage.

The riddle of the Buddhist Monk is not solved by abstracting what is common to the ascent and the descent. The miserable stockbroker is not imagined by abstracting what is common to the two brothers-in-law. The cyclic day is not conceived by abstracting what is common to every day (none of which repeats). Other chestnut examples from Fauconnier and Turner (2002) and the other introductory texts cited above demonstrate a vast range of blending in which the constructed meaning is not an abstraction over inputs but goes far beyond it (see, e. g., the chestnut analysis of “dinosaurs turned into birds”). Blending is very complicated.

2 Blending at work in specific communicative patterns

To know a language is to know a relational network of form-meaning pairs and how they blend to produce expressions. Form-meaning pairs are often referred to as “constructions.” Speakers mentally combine various form-meaning pairs to construct completely specified communicative performances and enact them. Such a performance is commonly referred to as a “construct.”

Language science needs a theory of creativity to account for:

1. Relatively quick (fewer than thousands of years, which of course includes anything down to milliseconds), non-species-wide creation of a form-meaning pair, at any point in human descent;

2. The blending of form-meaning pairs to create expressions, or more generally, communicative performances (“constructs”);

3. The creation of novel form-meaning pairs, using established ones as inputs, whether or not these new form-meaning pairs become entrenched to some degree;

4. The phylogenetic origin of language;

5. The ontogenetic acquisition of language.

The remainder of this section gives some examples of constructions that have received blending analyses.

3 Related work on blending in communication

Graph theory provides mathematical tools for crafting formal representations of patterns of various types, determining relationships (such as subsumption) between those formalizations, and unifying those formalizations. Graph theory is the foundation of formal approaches in unification grammars (Shieber 1995) such as Functional Unification Grammar, Definite-Clause Grammars, Lexical-Function Grammar, Generalized Phrase Structure Grammar, and Head-Driven Phrase Structure Grammar. Unification grammars typically represent a construction as an attribute-value matrix (AVM), and in fact an AVM is simply an object in graph theory.

Many readers will be familiar with the extraordinarily influential work on AVMs in the never published but legendary Construction Grammar Coursebook by Charles Fillmore and Paul Kay, such as the AVM for the lexical construction shoe,a very simplified AVM for the subject-predicate construction,

and the “unified” AVM representing the construct that is the sentence “trout relish worms.”

In graph theory, an AVM is an instance of a Directed Acyclic Graph (DAG) (Harary 1969). DAGs can be unified in graph theory through certain established algorithms. A unification of two DAGs is the most general graph that includes all the information in all of the inputs, if it exists.

The operation postulated as Merge in the Minimalist Program (Chomsky 1999) is to a great extent a notational variant of DAG unification in graph theory. (Of course, AVM unification in unification grammars operates over various assumed types of structure — syntactic, semantic, pragmatic, phonological, etc. — while Merge in the Minimalist Program is restricted to syntactic objects.) Both DAG unification and Merge embrace recursion.

DAG unification and Merge are interesting theoretically, but the major question from blending theory is: what is the cognitive basis for stipulating that DAG unification or Merge are human capacities? Certainly, human beings can do such things, but stipulating that they can do so takes out a loan of theory that needs, from a cognitive perspective, to be cashed out: what is the non-linguistic mental operation that makes these linguistic powers possible? The assertion of the present article is that the indispensable mental operation is blending.

To be sure, even if we are concerned only with language behavior rather than cognition generally, blending theory has an objection to such DAG unification approaches. They are indeed handy if the goal is to craft algorithmic processes for natural language processing: the mathematical theory and computational practices are strong and productive. But DAG unification models are bad as linguistic theory, because they do not have room for selective projection and emergent structure, argued to be indispensable in the blending analyses of linguistic phenomena, some of which are mentioned in the preceding section of this article. Most important, DAG unification and Merge do not account for the creativity we see in the origin of form-meaning pairs, the ontogenetic acquisition and development of form-meaning pairs, the phylogenetic development of form-meaning pairs, and the creation of new form-meaning pairs. A theory that does not account for such creativity is not a theory of human language, although it might, and manifestly often does, provide structure for the development of efficient and useful natural language processing tools.

The justification for stipulating the existence of a process in a scientific model is usually that such a scientific generalization allows us to “explain” a great deal in a compendious fashion, one that is not deeply disconfirmed by vast stretches of available evidence. The successful generalization is one that applies aptly to ranges of data that were not in the original data the contemplation of which led to the original hypothesis of the scientific generalization. This is the justification upon which the assertion of operations like DAG unification rests. Blending theory also asserts this justification, and additionally argues that it does a better job of accounting for the evidence, such as the evidence for various kinds of invention, innovation, and creativity. But there is a second justification, the principal emphasis of this article, that blending asserts: it does not stipulate the existence of language-specific mental operations, leaving it to cognitive scientists and psychologists and neuroscientists to go and locate them. On the contrary, the existence of blending and of its mechanisms is asserted on the basis of broad analysis of evidence from non-linguistic thought and action. Blending theory does not need to stipulate to the existence of a linguistic operation with certain powers in order to account for language; on the contrary, it adduces one that has already been analyzed outside of language science.

Within cognitive and functional approaches to language, there are repeated presentations of processes of language that look much like blending. The most prominent and thorough of these theories is Cognitive Grammar (Langacker 2006). Langacker writes, for example, that “one constructional schema can be incorporated as a component of another” (ibid.: 46). These many analyses include considerations of coercion when two units combine (Taylor 2002: 287), partial sanctioning of a unit, and extension and innovation during the combination of units (Langacker 2008: 215–255). The thorough attention to varieties of unification and combining is present in Cognitive Grammar from its earliest days. For example, in his landmark 1986 introduction to the field in Cognitive Science (1986), Langacker writes,

When a head combines with a modifier, for example, it is the profile of the head that prevails at the composite-structure level. (Ibid.: 13)

Each sense of ring […] combines with the phonological unit [ring] to constitute a symbolic unit. (Ibid.: 18)

An auxiliary verb, either have or be, combines with the atemporal predication and contributes the requisite sequential scanning. (Ibid.: 27)

A modifier is a conceptually dependent predication that combines with a head, whereas a complement is a conceptually autonomous predication that combines with a head. (Ibid.: 34)

[Of the sentences “liver likes Alice” and “Alice likes liver”] It should be apparent, however, that the same composite structure will result if the constituents combine in the opposite order, with Alice elaborating the schematic trajector of likes, and then liver the schematic landmark of Alice likes. This alternative constituency is available for exploitation, with no effect on grammatical relations, whenever special factors motivate departure from the default-case arrangement. (Ibid.: 35)

It may be that Langacker and other cognitive and functional linguists would agree (no doubt with illumination or objection, and only to a certain extent) that such analyses presuppose general cognitive mechanisms not specific to language and that indeed the chief one they presuppose is blending.

In response to this section, Langacker writes (p.c., and published here with his permission):

I fully agree that CG “analyses presuppose general cognitive mechanisms not specific to language and that indeed the chief one they presuppose is blending.” I have long taken it for granted that conceptual and grammatical structures can be characterized in terms of mappings between mental spaces, and in particular, that ‘composition’ amounts to (bipolar) blending. This conforms to my traditional description that component structures are ‘integrated’ to form the composite whole. Perhaps because it is so evident, I have not made the connection to conceptual integration theory as explicit as I perhaps should have done. Here is one succinct statement: “In composition, component structures undergo conceptual integration to form a composite structure that is more than just the sum of its parts” (Langacker 2017, 118). Slightly more elaborate indications of the affinity are found in Langacker (2009: 47) and (Langacker 2015: 135–136).

4 Multimodal communicative blending

To know a language is to know a relational network of form-meaning pairs and how they blend to produce expressions. More generally, to know a communicative system is to know a relational network of form-meaning pairs and how they blend to produce (actually performable) constructs. Part of the reason that we can understand text is that we can simulate mentally, in the backstage of cognition, a whole human performance whose language we would write down as that text. This is easy to demonstrate: if we hand anyone a text to read aloud, for example, they will do so with a great range of auditory and bodily performance that is nowhere “encoded” in the text but that seems to them obviously to be part of its actual performance. It is an oversight to claim that these aspects of performance are irrelevant to the text on the ground that we do not have to perceive them to understand the text. What we do not perceive we may well imagine. All understanding happens in the hearers, not in the texts.

There are many forms — auditory, visual, and textual — that we take to have a special relationship to a language, traditionally considered. But blending of forms that have communicative function is quite general; this blending for communication includes co-speech gesture broadly considered; the use of material affordances in the environment as props broadly considered; and a great range of auditory, visual, and motor performances, including pauses, laughter, clicks, taps, shoves, pokes, winks, nods, singing, whistling, cartoons, movies, music, television news, and on and on. Theoretically, we draw distinctions between performances in this ocean of performance and are unlikely to agree to homogenize all such form-meaning pairs and their blends to the same communicative status. But the general theory of form-meaning pairs and their blends applies to all such performances (Turner 2018).

Consider a chestnut example (used inter alia in Turner 2018) of creative blending in language: someone sees a termite doing something to wood. The meaning attached to the form “food” certainly does not include wood, but people can make a blend of a person eating food and what the termite is doing to the wood. This blend is not an algorithmic unification of these two quite different input spaces, but instead a new mental space constructed through selective projection from these inputs and conceptual elaboration to produce emergent structure in the blend. Crucially, forms that apply to any of the input spaces (and there can be many) can be projected to the blend to pick out corresponding meaning there, despite its not conforming to the meaning in the input from which it was projected. The wood can now be referred to as “food.”

In the same way, the young child who sees a lion at the zoo for the first time can point and say “kitty” to the delight of charmed adults. As analyzed in Fauconnier and Turner (2002), projection and blending of forms to refer to meaning developed in new situations solves the problem of equipotentiality.

Nesset et al. (2013: 229) explores linguistic creativity under blending by looking at a corpus of TV news clips in Russian and in English and investigating the use of the five Russian deictic words that correspond to the English meanings ‘here’ and ‘now:’ zdes', tut, sejčas, teper' and vot. These are forms with routine if quite complicated use in actual scenes of face-to-face classic joint attention (Turner 2017). They acquire slightly different sets of distributions for TV news. The English and Russian words acquire in the usage domain of TV network news distinct radial category profiles, in the sense that they display different centers of gravity in the semantic network. The authors propose the “Minimal Adaptation Hypothesis,” according to which “language makes adaptations that are as small as possible when applied to a new setting, such as the one created by TV” (Nesset et al. 2013: 229). This is unrecognized but impressive creativity.

This creativity is open-ended. In the case of web tutorials, the tutor can say “If you have any questions about getting the data from your old hard drive to your new hard drive, we made a video for you that covers the process, and we will link to it, right there” and points to and looks at the word “HERE” which appears suddenly in the blended ground when she points and says “there.” The printed word “HERE” is a form, and we know that its deployment in the blend carries the new, creative meaning that clicking on the visible spot directs a web browser to the linked URL.

Accordingly, the form “HERE” acquires a new meaning for the blend that it does not otherwise have. Turner (2017) discusses such creativity through blending for a range of linguistic forms, from lexical to clausal.

The need for cognitive (as opposed to merely linguistic) models of communication is suggested by a range of recent publications, such as the articles in the 2017 special issue of Linguistic Vanguard: Towards a Multimodal Construction Grammar (Zima and Bergs 2017). The author views proposals in that volume as potentially inspiring for the development of blending theory and takes as a starting point for further research that those approaches are compatible with the blending approach but less specific as models and less general in range of application. Additionally, the author takes the view, based on many extended conversations consistently over a period of more than 40 years with Charles Filmore, that a blending approach is compatible with most of his view of Construction Grammar, although explicitly and sometimes (notably in conferences and workshops at Berkeley) publicly contrary to the unification approach maintained by his collaborator Paul Kay, specifically on blending's fundamental assertion of the pervasive role of both selective projection from inputs and emergent structure in the blend. The blending approach has been discussed publicly and privately with Adele Goldberg when she was on the faculty at UC San Diego in 1992–1993 and many times since. Her exemplary career in Construction Grammar has deservedly inspired a strong international following, and it would be good for linguistics to arrange for a consideration of whether there are any incompatibilities between her tradition and the blending analyses.

5 Conclusion

In 1996, in a chapter titled “Language” in The Literary Mind: The Origins of Thought and Language, I argued that “Cognitive mechanisms whose existence we must grant independent of any analysis of grammar can account for the origin of grammar” (Turner 1996: 141). That argument showcased blending in the creation of constructions and the combining of constructions and in human linguistic innovation. The present article maintains that claim, more than 20 years later, but this time cites some of the work on blending in language and communication done recently. This article does not claim that any specific analysis of a specific linguistic phenomenon in the literature should be discarded or displaced by a blending approach, or that processes that look much like blending have not been discussed widely by language scientists during the last three thousand years. On the contrary, those many stipulations of linguistic processes need, in order to count as cognitive linguistics, a grounding in mental operations whose existence and power we must grant independent of any analysis of grammar. My claim is that the obvious candidate for this cognitive anchoring is blending. To be sure, in the conversation between theory of blending and theory of linguistic phenomena, there will be many detailed moments of analytic contest. Contest is indispensable to the progress of science. Properly conducted, it will strengthen and improve both blending theory and cognitive linguistics.