elephant

Greater learnability is not sufficient to produce cultural universals

I always feel the need to mention these cultural learning in the lab papers when they pop up.

This one, by Rafferty, Griffiths & Ettlinger, to appear in Cognition, uses an iterated learning experiment to challenge the idea that tendencies across cultures  is the result of some structures and concepts being easier to learn than others, as things being easier to learn means they will be more accurately transmitted from one generation to the next. Mini artificial languages in iterated paradigms (most notably Kirby, Cornish & Smith, 2008), have shown that languages become more structured as the result of generational turnover (and with an added pressure for expressivity), and this is hypothesised to be because of pressures for learnability (as well as expressivity/communication).

If we can show empirically that cultural features which are more prevalent are more "learnable", than this adds extra weight to the hypothesis that the driving force because culturally universal concepts are the result of learnability. However, this paper finds the opposite, if a concept is more learnable, then that does not necessarily result in it being more prevalent in transmission chains.

Their first argument is that more learnable cultural features are not likely to be (re)produced in transmission failure. This was shown in an experiment which featured "distinctive items", such as the word "Elephant" on a shopping list. In this context, the word "Elephant" was much more likely to be remembered than other items on a list, but once it had been lost in a transmission chain, it was never regenerated. Participants were much more likely to regenerate mundane food items which are likely to feature on a shopping list, such as "apple".

They also showed this mathematically, showing that agents are more likely to arrive at H2 if they learn from an agent with H2, even if H1 is more learnable. This is based on the assumptions that learners rarely learn a particular hypothesis unless they receive data generated specifically from that hypothesis, less learnable hypotheses are more likely to be confused with one another and so will arise more often through transmission, and that learnable hypotheses are unlikely to arise as the result of transmission errors, just like the word "elephant".

main_list

The main point of this isn't to argue that learnability doesn't play a role in cultural transmission, but rather that we shouldn't look at only learnability bias, because we also need to take into account the rate at which hypotheses (or cultural features) change into other hypotheses.

This is quite a neat experiment, and works well in explaining certain cultural trends we see. The paper cites the example that across religions, concepts are generally “minimally counterintuitive” (Boyer, 1994). This is often cited to be because things that are minimally counterintuitive are more learnable, because counterintuitive elements are learnable because they are surprising, but if you have too many surprising things in a story, people can't remember all of them and so they can't be transmitted. Minimally counterintuitive stories have an optimal amount of counterintuitive elements to be memorable and so are transmitted with more fidelity (see Norenzayan et al. 2006 for a really neat study). However, Rafferty. Griffiths & Ettlinger show that more than just learnability needs to be taken into account because even one counterintuitive element is unlikely to be reinvented if forgotten. A good linguistic example of this is clicks, clicks have been shown to be the most acoustically salient sounds, but they appear in very few languages. This could be because they are unlikely to be spontaneously (re)produced, so are lost despite their learnability.

The scope of the implications for this study are limited however, as it makes the massive assumption that learnability is only born from an item being surprising. Sure, if all you're learning is a shopping list, the most learnable thing is the elephant, because it is the most surprising, but if you're learning a whole language, the surprising things are not going to facilitate learning an entire lexicon and grammar. In the context of language learning, structure is the thing that makes it more easily learned, rather than things being surprising.

The paper then goes on to test whether, if a set of languages that lack a property far outnumber the set with a property, then this can override a learnability bias. They carry out an individual artificial language learning experiment which shows that the learnability and generalisability of words with vowel harmony is greater than words without vowel harmony. However, they then go on to do an iterated learning experiment and find that no matter how much of an initial language consists of words with vowel harmony, all languages ended up with approximately 50% words with vowel harmony. This shows that, while one language is more accurately transmitted than others in an individual learning task (or in one generation), because of the large number of non-harmonic possibilities, a completely harmonic language can not win out through transmission.

There's quite an extensive paragraph explaining why vowel harmonic languages should exist at all then:

In contrast with the results of our experiment, harmony does exist in many languages of the world. Several factors might result in harmony being more common in these languages than in the final generations of our chains. Our experiments focus on cognitive learning biases, but it is likely that there are also sensorimotor biases that favor the articulation and perception of harmonic languages (Blevins, 2004). There may also be qualitative factors not included in our experiment that lead to the harmony bias being stronger in natural language than in the lab. For instance, children could have stronger harmony biases. Additionally, the quantitative bias towards harmony may be stronger than we found in Experiment 1. This could occur due to the existence of more words and a longer period of learning and use in naturalistic settings. Since all of our participantsLearnability and cultural universals 26 were adult speakers of English, their bias could also be weaker due to the fact that English is not a vowel-harmonic language. Another factor that could lead to a divergence between our results and natural language learning is the use of a linear transmission structure. In natural language learning, children may learn from people who are part of generations other than the prior generation. Transmission patterns are likely also influenced by other factors, such as language contact. This can result in speakers borrowing phenomena from other languages, resulting in the spread of properties that are unlikely to be generated spontaneously. Finally, there may be increased noise in transmission in lab experiments due to the fact that learning occurs over a relatively short period. This might mean that we would expect harmonic languages to eventually become less prevalent due to transmission errors, but that this process will be much slower in natural language than in the lab. In the experiment, we see that by the third generation the languages no longer contain more harmonic words than would be expected by random chance. This rapid shift may indicate that participants exhibit very little bias towards harmonic words when the input language is not 100% harmonic; in a naturalistic context, generalization is likely to be somewhat more robust due to the longer learning period and broader exposure to the language. Despite these differences, our experiment provides evidence for the fact that a bias need not lead to a universal tendency, something which is born out in the pattern of vowel harmony in existing languages: vowel harmony is relatively common, but it is not present in the majority of languages.

The conclusion is that there are a whole host of reasons why a language or cultural phenomenon might be successful in being transmitted and the learnability isn't necessarily enough. But I'm not sure anyone was arguing it was in the first place.

References

Boyer, P. (1994). The naturalness of religious ideas: A cognitive theory of religion. Univ of California Press.

Kirby, S., Cornish, H., & Smith, K. (2008). Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language. Proceedings of the National Academy of Sciences105(31), 10681-10686.

Norenzayan A, Atran S, Faulkner J, & Schaller M (2006). Memory and mystery: the cultural selection of minimally counterintuitive narratives. Cognitive science, 30 (3), 531-53 PMID: 21702824

Rafferty, A. N.Griffiths, T. L., & Ettlinger, M. (in press). Greater learnability is not sufficient to produce cultural universals. Cognition.

 

  • Catriona Silvey

    Nice post!

    It feels to me like treating 'languages and concepts' interchangeably allows the authors to do a bit of sleight of hand. As far as I know, the learnability-based accounts the authors cite do not generally suggest that the learnability of _individual items_ in a language determines whether they will survive the process of cultural transmission. It's the learnability of the entire system. And a learnable system has the property that forgotten/unseen items can be easily generalised on the basis of remembered items. So participants not being able to regenerate 'elephant', but being able to regenerate 'apple', is actually completely in line with this system-wide learnability account. This ties in to your point about learnability != surprising; what would it mean for a whole language to be surprising?

    Also, the shopping list example is odd, because the background assumptions of the task actually provide you with a prototypical set of items (which is why you can regenerate 'apple', but if your mind has gone blank, you're very unlikely to guess 'elephant'). I can't think of a comparable feature for language transmission, other than perhaps the general observation that the helpfulness of context and our inferential abilities sometimes relieves learning pressures (e.g. allowing underspecification & homonymy to exist). But this is a bit of a stretch!

    Their general points that learnability isn't everything and that we shouldn't draw massive conclusions from limited experiments are obviously sound, though. I like the point about religions - makes me think of Bartlett's War of the Ghosts study (1932). Cultures find their own bits of counterintuitive myth perfectly fine (angel visits a virgin to tell her she's going to have God's baby) but other cultures' so weird as to think 'how could anyone ever come up with that' (a man shot in a war with ghosts doesn't feel sick until something black comes out of his mouth & he dies).

  • Sean Roberts

    Nice post and an interesting study. A few random thoughts: It should be pointed out that the paper fits conveniently with the concept of random noise as the source of cultural innovation, which is the mechanism in all the Bayesian modelling (although the models also assume a closed-set of possibilities).

    The reasoning is a bit odd - an elephant appearing in a shopping list is unlikely, but in this study it did actually appear, and it's a countable, tangible thing. Therefore presumably it's more likely to appear than something extremely unlikely like the sound of liquid glass. Also, if you asked people to name an odd thing to put on a shopping list, I bet Elephant would come up. Basically, I'm not convinced that learnability = suprise. Cat's comments about the system hit the nail on the head.

    It reminds me of my 'evolve a band name' experiment (http://www.replicatedtypo.com/results-of-evolve-a-band-name/5313.html) where people had to remember 10 band names and reproduce them. One person wrote "God my memory sucks", and this remained alive in 4 chains that split off from this. The religion stuff I'm not so sure about - a while ago I wrote about Konrad Talmont-Kaminski's idea of religious beliefs beings selected to be 'super-empirical' rather than just memorable (http://www.replicatedtypo.com/the-evolution-of-religion/2343.html).

    Clicks are an odd example to use because, although they aren't used in many phone inventories, they are used paralinguistically in many langauges (WALS lists 123 languages with clicks from all over the world, versus 20 that don't use clicks, http://wals.info/feature/142A).

  • Hannah Little

    Great comments guys. I want to see what a constantly surprising language looks like.

  • Marc Ettlinger

    I'm a reader of your blog. Want to chat about the paper?

  • Hannah Little

    Hi Marc, thanks for commenting! I'd love to hear what you were originally trying to find, perhaps you could even write a guest post on replicated typo?

  • Marc Ettlinger

    Sure, happy to. Just let me know when and how and I'll put something informal together.

  • Hannah Little

    Hi Marc, whenever is best for you is fine, you can either email it to me and I'll post it as a guest blog, or you can email James for a login of your own. (email addresses available in the authors section above). Thanks, Hannah