Saturday, 25 March 2017

Language, thought and the ecological approach; A Purple Peril

As part of a class on cognitive psychology, I give a seminar in which we talk about the research on the relationship between language and thought. In particular, I show this great talk by Lera Boroditsky as a starting point. She talks about the kind of research in this area, and talks about results such as how we linguistically interact with space and time affecting how we physically interact with these things. For example, some languages like English use an egocentric frame of reference when talking about space (e.g. describing things as being to the left or right, where the origin of this space is the speaker). Other languages use a geocentric frame of reference (e.g. describing things as being to the south of you). In order to be able to speak and understand the language, you therefore have to be able to remain oriented in space, and speakers of these kinds of languages have been shown to be capable of impressive feats of dead reckoning previously thought impossible in humans. 

The reason this is all interesting is in the context of how the field is changing how it thinks about language; is it magical, or merely interesting? If the former, language becomes a unique human cognitive capacity that requires specific neural mechanisms that serve language and nothing else. If the latter, language becomes an integrated part of our cognitive systems and we should expect it to show these connections to other capacities. 

The weight of evidence right now I think favours the latter view. In fact, one whole strand of embodied cognition (Shapiro’s ‘conceptualisation’ hypothesis strand) explicitly pursues these connections between language and other capacities, for example Lakoff’s work on metaphors being grounded in action. Language, while still phenomenal in what it can do, is not different in kind to the rest of cognition. 

The field is still very much at the ‘functional model’ stage of developing explanations, however. The research mostly just catalogues linguistic differences and cognitive differences and works to map those onto each other in a fairly metaphorical, word-association kind of way (e.g. politics is talked about in terms of left and right wing so this should connect to physical movements to the left and the right). Our ecological questions has become, what kind of mechanism might allow this kind of cross-talk, and as I’ve been chatting to students I’ve been connecting a few dots for myself. This post sketches the outline of a mechanistic, ecological research programme for attacking the fascinating problem of the relationship between language and thought.

Our ecological approach grounds all explanations of the structure in behaviour at the level of ecological information and our interactions with that information. Other things that can structure behaviour happen (e.g. the formation of ecological neural representations) but all of those work the way they do because of the ecological way we perceive and act in the world. 

Sabrina has been chipping away at the annoying problem of language for a long time (see the 'Language' posts listed here). Other ecological attempts go looking for linguistic affordances, which frankly never works at any level other than the functional, metaphorical level seen in the work above. Sabrina’s key move (Golonka, 2015 and this post) is to go to information alone and see if it can do things that might support something like language. Turns out, yes, it can. 

Ecological information is complex structure in an energy array that is specific, though not identical, to the dynamical process that created it. Sabrina’s insight is that there are actually two ways for an organism to use these structures. The first is law-based. If you use that structure as if it were about the dynamical process that created it, then you can select, coordinate and control your actions in real time to complement that dynamic. This is the kind of real time action control we ecological types normally study. The second is convention-based. If you use that structure as if it were about anything other than the dynamical process that created it, then you can select but not control actions with respect to that thing. She describes this as convention based because typically, the only way this works is that there are enough organisms who agree to use the information that way to make it pan out. You can’t have a language community of one person; no functional behaviour happens when no one else respects the same conventions. 

Language, therefore, like everything else, is a behaviour shaped by our use of information (primarily convention-based use). This view immediately makes language the same in kind as things like perception and action; it’s all behaviour shaped by our interactions with the very real ecological information in which we are densely embedded. This provides the only existing unitary framework for analysing language as embodied and enacted, as well as figuring out (beyond mere metaphor and description) how language and thought are related and interwoven. 

Cognitive capacities are connected to the extent that the underlying mechanisms use and interact with the same information. I’ve argued this in a recent empirical paper (Snapp-Childs, Wilson & Bingham, 2015) showing that the extent of transfer of learning is explained by overlap in information and the details of the differing stability of the two task dynamics. Information governed whether transfer happened by being the necessary ‘thing in common’, and then there were consequences dictated by some of the details. 

From this mechanistic ecological stance,

  1. Of course language and the rest of cognition can be interconnected because they are all the same type of thing; behaviour shaped by our use of ecological information
  2. The nature of those interconnections will be shaped by the informational overlap between tasks, and
  3. The details of the interconnected behaviours will depend on the details of the dynamical mechanisms implementing the cognitive activities, and in particular on the details and differences in how the overlapping information is being used by the two dynamics

This, I propose, is a decent first draft of an ecological research programme into the relationship between language and thought, and given that this field has documented many interesting connections without every successfully explaining any of them, I also propose that this research programme has a lot to offer the field. 


  1. This comment has been removed by the author.

  2. I agree, there's a lot of promise in an ecological approach to language. But I'm not sure that I see much of a research programme in your outline (or the places you linked to).

    1. I'm not even sure that the language-thought relationship is a good way to phrase what the field should be about. It's really only a question that comes up when the artificial distinction is being debated - but simply dissolving it would do just fine. I've outlined a way of doing it here:

    2. You need a much richer image of what 'language' comprises. See here and here We need a model that can process the laughter at a comedy gig as much as giving instructions, small talk along side rhetorical flourishes, code switching, turn taking, conversation repair, bilingualism, idioms, syntax, dictionaries, ... I'd also include phenomena that are commonly included under culture such as linguistic identity, language change, etc. I'd suggest you start thinking about those rather than imagining that you can deal with simple interactions and scale those up. I'd look for inspiration in construction grammar, local grammar, corpus linguistics. There's just so much out there. The current psycholinguistic paradigm is completely moribund and just toying at the edges of language. Although, I think discursive psychology has a lot to recommend it.

  3. 3. I think the ecological approach can tell an interesting story but I want to see something a practicing linguist or even psycholinguist can get out of it. Solve one of our real problems. For example, incredibly rich morphological paradigms (Russian, Finnish, Swahili) with hundreds of selections take no time to process (or acquire) when compared to languages with no morphology. Yet, relatively simple syntagmatic phenomena - like embedding - seem to take a processing toll - as if there was a processing module. I've argued that we need a model that's more like face recognition than chess - that is massively parallel as opposed to serial algorithm resolution.

    4. There's also the venerable problem of dual or multiple articulation in language. We seem to be solving a number of extremely complex tasks at once with some building on the others. We we have to articulate phonetically/phonologically, pick the right lexicon (words, socially appropriate expressions, genre conventions), pick the right grammar (morphology and word order), orient our body just to say something as simple as 'You dropped your glove'. But there's also convention. In Czech, you would say 'The glove dropped for you' (Spadla [Dropped{3sg, past, fem}] vám [you {polite, dative}] rukavice [glove{nom, sg}]. All of that just to alert your interlocutor to something a simple grunt and a finger point could do. Literally noone has a clue how this all actually happens all the time even with people with severe intellectual impairments. Often in multiple languages.

    5. Given the above. I'm not sure you can really get rid of mental representations as an analytical category. I'm perfectly prepared to admit that they don't actually exist in the brain, but we behave as if they did. For instance, presupposition. 'When did you stop cheating on your taxes?' How do I know that the addressee had been accused of cheating on their taxes? Or even more complex 'In France, Clinton would not have had a problem.' Or, how come 'Jewish holocaust' and 'German holocaust' mean exactly the same thing. But you probably have to Google it if I say 'Spanish holocaust'. Or I cannot say 'The Browns (Mr and Mrs) conquered England.' But I can say 'The Normans conquered England.' Or 'The Beatles conquered America'. What are the tasks involved, what is the information? I can imagine you could rephrase it in those terms, but I'd like to see it done in a way we could build on, rather than just translating it. You have to be able to somehow account for the 'knowledge of language'. At the moment, we have very poor models for doing that. We have the computational metaphor with processing and storage of information - dead in the water, we have the connectionist model which suffers from the isomorphism fallacy and still struggles with the legacy of functional localism. I see a glimmer of something in the ecological approach - but mostly a liberation from the old models - not necessarily an affirmative framework for thinking about all these things in a tractable manner. Here's something to try. I've written about is the phenomenon of 'call back' in comedy shows. It is a completely puzzling phenomenon - none of the information processing and memory storage models seem to have much to say about it. You can explain it in a connectionist manner but not very satisfactorily. Can an ecological approach make a dent here (while also taking into account the dual articulation issues described above)?

    PS: A tiny niggle: Lakoff does not actually claim that metaphors are rooted in action. Simply that they often rely on our bodily orientation to the world. But you can even have metaphors like 'Marriage is a contract' that does not draw on that. A much more fruitful concept of Lakoff's Idealized Cognitive Models (later 'frames').

    1. Thanks for all this; these details are all very relevant (although a little over my 'language researcher' paygrade so I have no good answers just now :). At this point I was just articulating an idea; I'll think about how to apply it to a specific problem.

      Two quick things I can address;
      I'm with you on the details of Lakoff; I'm not nailing myself here to anything specific, just the general idea.

      And while we do need to address the issues of 'language knowledge' effects (for lack of a better term) I will continue to argue that mental representations are the wrong analytic tool for many, many reasons. This argumwnt comes up a lot - "language is clearly more than perception and action, so there must be something intervening, and those must be mental representations", to which I answer, in order, "yes, it is; yes, there is; no, they don't have to be and there's reasons to think they aren't.". I know that's not a satisfying answer but it's the stance I'm taking right now :)

    2. I actually totally agree with you that there are good reasons nothing like mental representations exists at any biological level (stored, activated, or whatever by the embodied brain). But as a linguist or even a psychologist, they are the only way I know how to think about the actual phenomena. I see the benefit of your programme in allowing me to stop myself looking for projections of these are 'lower' levels or even to think of this in terms of levels. But nothing I've seen of the ecological approach actually suggests direct usefulness at the, for lack of a better term, higher levels of description.

      Here's an analogy. I'm looking at a manga image on my screen. I want to say things about colours and shapes, meanings and intentions. But I could also talk about whether this is a vector or raster image. This is crucially (one could say 'generatively') important for the image's existence but I almost never care. I may care if I need to know how much space the image file is going to take or if I want to enlarge it. But I'm still going talk about areas of colour in a vector way, even though the image is actually just a collection of pixels on the screen. I may also not talk about brushstrokes, if I know the colouring was done in Illustrator rather than Photoshop, etc.

      That's how I see this discussion. It matters a lot of the time (e.g. when it comes to scalability, consideration of processing loads, etc.) but rarely to anything I consider to be substantive to the analytic 'task' at hand.