But we’re getting ahead of ourselves. We need a little conceptual equipment before considering the example. It’s the conceptual equipment that’s in question. Make no mistake, the concept of memes is conceptual equipment, and it’s confused and confusing.

Roles in Cultural Selection

Genes and phenotypes play certain roles in a more or less standard account of biological evolution. The phenotype interacts with the environment, where it either succeeds or fails at reproduction, depending on the “fit” between its traits and that environment. Where the phenotype is successful at reproduction, it is the genes which are said to carry heredity from one generation to the next.

In one very widespread account genes are said to be replicators. That is to say, replication is the role they play in evolutionary change. Here’s what Peter Godfrey-Smith has to say about that (The Replicator in Retrospect, Biology and Philosophy 15 (2000): 403-423.):

In The Selfish Gene (1976), Richard Dawkins had argued that individual genes must be seen as the units of selection in evolutionary processes within sexual populations. This is primarily because the other possible candidates, notably whole organisms and groups, do not “replicate.” Organisms and groups are ephemeral, like clouds in the sky or dust storms in the desert. Only a replicator, which can figure in selective processes over many generations, can be a unit of selection.

At the same time Dawkins coined the term “meme” to name entities filling the replicator role in cultural evolution. Later on he used the term “vehicle” to designate the entity that interacts with the environment. In biological evolution it is phenotypes that are the vehicles. In cultural evolution, well, that’s a matter of some dispute. And that more general dispute–what are the roles in cultural evolution and what kinds of things occupy them?–is what interests me.

However, I don’t particularly like the term “vehicle.” As Godfrey-Smith has noted, following others, it is a gene-centric term, characterizing what entities do from the so-called “gene’s eye” perspective. I’d prefer a more neutral perspective and so will use a term coined by Richard Hull, “interactor.” Here are definitions as Godfrey-Smith gives them:

Replicator: an entity that passes on its structure largely intact in successive replications.
Interactor: an entity that interacts as a cohesive whole with its environment in such a way that this interaction causes replication to be differential.

We need one more role, that of beneficiary, as defined by Elisabeth Lloyd. “The beneficiary, for Lloyd, is the entity that ‘ultimately benefits’ from a process of evolution by selection” (Godfrey-Smith, see also Lloyd’s treatment of Units and Levels of Selection HERE). Godfrey-Smith goes on to suggest that “that most of the language of … ‘ultimate benefit’ in this context is merely metaphorical.” That may be so, however, I find it useful, at least heuristically.

Consider for example gene-culture coevolution school of thinking, which offers a technically sophisticated treatment of cultural change. In that tradition it is biological organisms, mostly humans, but also chimpanzees, songbirds and some other animal species, that occupy the beneficiary role. But that is not the case in any version of memetics, where it is the memes that are the beneficiaries. In fact, part of the appeal of memetics has been that it offers a way of thinking about cultural practices, such as life-long celibacy, that are at odds with biological interests. While I’ve got problems with memetics, I am very much interested in accounts of cultural evolution it which it is the cultural entity that benefits from cultural evolution.

So, we’ve got three roles on the table: interactor, replicator, and beneficiary. Let’s return to some jazz club in Philadelphia on some night in 1952 and talk about that performance of “Dexterity.”

Late Night Dexterity

The environment in which any cultural practice thrives or dies is, of course, the social group. The group relevant to the imaginary jazz performance I invoked at the beginning is, first of all, the people present in the club to witness it, the audience and, of course the musicians themselves, not to mention the other people working at the club. The jazz audience is, of course, larger and more diffuse than the people in a given club on a given occasion. But we need not worry about the demographic details, not for the purposes of this example. All I want to do is to explicitly state that the evolutionary environment for music is some population of people.

And it is the performance that plays the interactor role, not the tune, in this case “Dexterity,” but the performance of the tune. It is the performance itself that the musicians and audience will find to be satisfactory or not. The tune is simply a source musical materials on which a performance can be based.

What, then, is the beneficiary of the performance? Well, one would think it is those musical materials in the tune, “Dexterity” in this case. That’s not all, but let’s start with that.

Of what does the tune “Dexterity” consist? The conventions of jazz have it that such tunes consist of a melody and a harmonic structure. A melody is a sequence of musical tones having specific pitches and durations. Melodies are concrete and, in principle, one can hum, whistle, or sing them. In practice, that might be difficult. “Dexterity” is one such difficult case as it is fast and complex, a characteristic of many bebop tunes—bebop, of course, is a particular jazz style and “Dexterity” is a tune in that style.

A harmonic structure is more abstract than a melody and can be realized equally well in many different ways, ways that will sound different in matters of detail. Instruments that can play only one note at a time, such as trumpet or saxophone, can play melodies that are consistent with a given harmonic structure but they have trouble conveying the structure directly. Multi-note instruments, such as piano or guitar, can convey harmonic structure more directly. Within the jazz world harmonic structures are generally notated as short strings of alphanumeric characters, each of which indicates a single chord: B7, Am, EMaj9, etc.

The harmonic structure of “Dexterity”¬–and this is why I chose it as an example–happens to derive from another well-known song, George Gershwin’s “I Got Rhythm.” There are so many songs that derive their harmonic structure from “I Got Rhythm”, especially in bebop, that “Rhythm Changes” has become a term of art within the jazz world. Obviously enough it designates the harmonic structure (chord changes) of “I Got Rhythm” considered as an autonomous musical entity (for a little history, see Cultural Evolution 4: Rhythm Changes 2).

So, a successful performance of “Dexterity” will benefit the specific melody of the tune, increasing that likelihood that audience members will want to hear it again and that the musicians will perform it again. It will also benefit the specific harmonic structure, which happens to derive from another tune, “I Got Rhythm”. The “Dexterity” melody is specific in a way that satisfies the definition of replicator, and I suppose that’s true of the harmonic structure–Rhythm Changes–as well. That’s how I treated it a couple of years ago, in Cultural Evolution 4: Rhythm Changes 1, where I talked of it as being a memetic entity. Still, it’s abstract nature continues to give me pause.

But there’s more to a performance than the materials provided by the melody and chord changes. I’m imagining that this specific performance is by a quintet consisting of trumpet, tenor sax, piano, bass, and drums¬¬–a standard line-up. In this performance the group plays the melody once at the beginning and once at the end. In between we have improvised solos. “Dexterity” provides the opening and closing melody and it provides the harmonic framework for all the improvisations. More materials are necessary.

The drummer needs to play appropriate patterns throughout. The same is true for the piano player and the bass player, who have to supply accompanying materials throughout the performance. Things are a bit different for the sax player and the trumpet player, who are responsible for playing the melody at the beginning and the end. When they solo, however, they’re on their own.

A good many of these bits and pieces–variously called riffs or licks–may properly be considered to be replicators. Some of them will be standard throughout the bebop community, if not jazz in general, and so are played by many musicians and on many different tunes. But some of them may be specific to these musicians, and even to this specific performance of “Dexterity.” Perhaps the piano player, for example, comes up with a new lick that she likes a lot and that the audience likes as well. So she uses that lick in subsequent performances, not only of “Dexterity”, but of other tunes. One of those performances gets recorded and other musicians here it and take up that lick. Within two or three years it’s spread throughout the bebop community (and 40 years later gets sampled by a hip hop artist). It’s become a meme.

It is unlikely, however, that we’ll be able to analyze the whole performance as but a complex collection of replicators (that is, memes). For one thing, there’s likely to be at least some specific phrases that aren’t derived from prior materials. For another thing, the way all these derived and novel elements fit together is important. A performance, the interactor, is not a collection of replicators and other stuff joined in random order. The order is important.

Thus I do not think we can reduce the interactor (the musical performance) to a collection of replicators and other stuff, any more than we can reduce a biological organism to a collection of genes. In biology there is a developmental process in which phenotypes emerge through a process in which genes interact with an environment. In music there is a performance process, if you will, that accomplishes an analogous task.

Not only do the musicians have to perform the music, but members of the audience must do so as well. The process of listening to music is not passive in the way that recording devices passively record sounds. Members of the audience take the sounds that they hear and construct connections between them. Bits a piece of this process may be explicitly conscious, such as the aha! when you notice that that sax player has just quoted the theme song from the Woody Woodpecker cartoons (which also is based on Rhythm Changes), but most of it is unconscious. It just happens.

Regardless of one’s level of performance skills, one has to learn to hear music. The audience for our imaginary performance is likely to include some people with sophisticated performance skills, people quite capable of performing “Dexterity.” But most probably have few performance skills; they may sing in the shower or in church but otherwise they don’t perform any music. Still, they had to learn how to hear jazz.

Some who is not particularly knowledgeable abou bebop might well take great pleasure in a good performance of “Dexterity” even if they don’t quite know what’s going on. They might not know, for example that the harmonic structure is that of “I Got Rhythm.” And they might not know that that structure is 32 bars long and divided into four sections of eight bars each. But there’s something going on that they like, to they applaud loudly and seek out other performances, by the same musicians, of the same tune, and in the same idiom.

Memes: Internal or External?

Where are these replicators, these memes, to use Dawkins’s term? In a standard view within memetics (e.g. Daniel Dennett’s view) the memes are patterns in people’s brains. But that isn’t the view I took in the above discussion. The melody for “Dexterity” is a concrete physical entity out there in the physical world where, not only can it be heard by people, but it can be picked up by recording equipment. The same is true of all those licks and riffs and even of the rather more abstract harmonic structure (which I’ve argued in some detail).

If these things didn’t exist in the external world as physical objects or properties of physical objects, they would be of no use to musicians. In order to perform together musicians need to coordinate their actions, and do so in a very exacting way, one whose temporal characteristics must be measured in milliseconds. The only means they have of doing this is to listen to the sound. If the memes cannot be identified with properties of the sound stream they cannot play a role in coordinating the musician’s interaction, nor can they be conveyed to an audience.

The memes are those properties of the sound stream through which the musicians couple their actions into a single coordinated activity. The coupling extends to the audience and you can see and hear in the way they tap their feet, gasp a particularly audacious lick, stop chattering at the performance builds, and so forth.

That is, musical memes perform a coupling function during performance and it is these coupling functions that must be learned from one person to another if the individual memes are to earn a long-term place in the musical repertoire.

I’m not denying that, when a musician learns the melody of “Dexterity” that something happens in the musician’s brain. But it’s not at all obvious to me that we accomplish anything useful by saying that that something has been replicated from the brain of another musician or musicians. And the same is true in the case of audience members. In order to appreciate the performance they must already know the performance conventions to some degree or another.

I see nothing to be gained by saying musical replication is a matter of transferring replicators between brains. Very obviously there IS replication of sounds. Without that there is nothing. It is the sounds that are the replicators, not the neural patterns in brains.

What then, is the biological analog of whatever it is that’s taking place in human brains during musical performances? Ontogenetic development. In the case of single-celled organisms that’s a relatively simple process of cell division, but it’s not a null process. Not only must the genetic material be replicated, but cell structures and organs must be created. In the case of multi-celled organisms the developmental process can be quite complex.

The same is true in music. Using the melody of “Twinkle, Twinkle, Little Star” as the basis for a performance can be a relatively simple and straightforward process. But a skilled improviser could create a very sophisticated piece of music on that material. Wolfgang Amadeus Mozart was such an improviser, though he worked in a different idiom. We don’t really know whether or not he gave improvised performances of that tune, but the fact that, he composed a sophisticated set of variations on is suggests that he probably did.

That, of course, raises the question of more or less richly annotated compositions, such as those in Western classical music. But I’ve done enough for one blog post. We can consider that question some other time.

* * * * *

Addendum: When looking for memes in cultural evolution one can proceed under one of two assumptions: 1) Memes are a new kind of entity not yet identified in the human sciences. 2) Memes are entities that are already known but that have not heretofore been assigned to that role in the process of cultural evolution. I have taken the second position.

In contrast, Levinson suggested that we should be moving away from the picture of the modal individual with a fixed language architecture.  Instead, we should embrace population thinking and recognise the variation inherent at every level of language from typology to processing and brain structures.  While languages are constrained by the processing structures of the brain, these processing structures are plastic and adapt to the language and cultures in which they are embedded.  Adults lose the ability to distinguish sounds that are not part of their language.  Recent work on linguistic planning using eye-tracking shows that the elements of a scene that speakers attend to before starting to speak differs with the canonical word order of their language.  More fundamentally, brain structures can be affected by cultural experience, such as bilingualism or singing (indeed, the effect of bilingualism on processing shows that variation itself is a fundamental constraint).  So, brains do constrain learning and processing, but are themselves subject to constraints from interaction between individuals.  Brains also change over evolutionary time, adapting to a range of pressures.  Therefore, there is a complex ecology of systems that co-evolve to define the constraints on language, and understanding these systems requires focussing on diversity.

Hagoort conceded that there was impressive variation at each level, but wondered what was meant by “fundamental” differences.  For instance, how important is the precise neural architecture of an individual?  Even within the variation pointed out, complex linguistic processing isn’t being done in the thalamus, and this is a constraint that sets a boundary on variation.  Hagoort might have pointed out that, if there was so much variation between individuals, how do they communicate so effectively and how does basic interaction happen so easily between diverse individuals?  This points to brain processing universals that explain the constraints on language.

Both sides agreed that the basic aim of any science, including linguistics, is to discover general principles that explain the data.  However, are researchers focussing on the same data?  What is the object of study that linguistics are trying to find generalisations for?  It seems to me that the debate came down to what each proponent thought was the domain that was most likely to yield general explanations.  Hagoort suggests that we should be focussed on brain structures and processing in the individual.  Levinson, on the other hand, suggests that the interaction between individuals is a key domain (e.g. the interaction engine).  Proponents of cultural evolution such as Simon Kirby might argue that cultural transmission is a key domain.  It’s possible that the most relevant ‘universals’ in each of these domains may be very different.  A constructive step would be to describe how each of these domains constrain the other.  For instance, constraints on language processing in the brain certainly constrain interaction between individuals, but the requirements of interaction may affect how processing is employed.

There were some good points from the floor, including Peter Seuren pointing out that neither view was particularly close to proving their point, since proving universals, or their absence is very difficult.  A paper under review by Steven Piantadosi and Edward Gibson attempts to answer whether it is possible in principle or practice to amass sufficient evidence for a statistical test that would demonstrate a universal.  They conclude that it is possible in principle, but that there are not enough datapoints (languages) in order to achieve the required statistical power.  There was also an appeal for the study of diversity for the sake of diversity – that there are different motivations for explaining phenomena in the world, and that one of them is to understand human diversity.

The general message:  Proponents of universals need to take diversity into account, and proponents of diversity need to be more specific about how diversity maps onto processing and how different domains of language co-evolve.

1) The first are the videos of the plenaries from last year’s EvoLang conference in Kyoto. Unfortunately the page that did host these is here and now displays an error message: http://ocw.kyoto-u.ac.jp/international-conference-en/31 (I’ve posted the link incase they fix it).

FORTUNATELY, I still have the direct links to all the videos so here they are:

(Though I’ve completely lost the videos for the biolinguistics workshop videos hosted on the same site, so if anyone has these links please send them to me, or just comment. Thanks.)

2) On the CARTA website you can find videos of speakers such as Terence Deacon talking about Symbolic Communication: Why is Human Thought so Flexible? as well as V.S. Ramachandran, Colin Renfrew and Patricia Churchland.

3) The videos from 2011′s ProtoLang can be viewed here: http://www.protolang.umk.pl/videos_and_links

There’s a link to the videos from 2009′s protolang at the bottom of that too, but they all seem to be broken. But you can actually still find them by searching for the author’s name on http://tv.umk.pl/

For example, searching Bart de Boer, you can find: http://tv.umk.pl/#movie=521

4) YouTube.

Highlights include Simon Kirby’s inaugural lecture at Edinburgh University, Kenny Smith at the University of Southampton earlier this year, more Terence Deacon, Luc Steels on robots and loads of other stuff, I am sure you are capable or googling the names of some language evolution folk.

Also, you can watch bbc horizon’s why do we talk featuring Techumseh Fitch, Simon Kirby and others here: http://www.youtube.com/watch?v=75XxjJYuV7I&list=PL9DD35E568234CA7F

And by request in the comments: Peter Richerson – How Possibly Language Evolved http://www.youtube.com/watch?v=zxJMtZUaeZU

If anyone else has some good video resources please add them in the comments!

Some bits. From the beginning:

…evolution will occur whenever and wherever three conditions are met: replication, variation (mutation), and differential fitness (competition).

In Darwin’s own terms, if there is “descent [i.e., replication] with modification [variation]” and “a severe struggle for life” [competition], better-equipped descendants will prosper at the expense of their competitors. We know that a single material substrate, DNA (with its surrounding systems of gene expression and development), secures the first two conditions for life on earth; the third condition is secured by the finitude of the planet as well as more directly by uncounted environmental challenges.

The first question, then, is whether or not these conditions are met by human culture. Dennett thinks they are and so do I.

From the end, however:

Do any of these candidates for Darwinian replicator actually fulfill the three requirements in ways that permit evolutionary theory to explain phenomena not already explicable by the methods and theories of the traditional social sciences? Or does this Darwinian perspective provide only a relatively trivial unification?

We do not yet know. But are the prospects for non-triviality good enough to warrant considerable investment of conceptual time and energy? And so

We should also remind ourselves that, just as population genetics is no substitute for ecology—which investigates the complex interactions between phenotypes and environments that ultimate yield the fitness differences presupposed by genetics—no one should anticipate that a new science of memetics would overturn or replace all the existing models and explanations of cultural phenomena developed by the social sciences. It might, however, recast them in significant ways and provoke new inquiries in much the way genetics has inspired a flood of investigations in ecology.

* * * * *

Finally, a note on how my thinking about cultural evolution differs from Dennett’s. Those who’ve been memetics by the bug have two conceptions to choose from, both of which can be found in Dawkins. One can think of memes as cultural analogus to biological viruses or one can them of them as analogues to genes. Dennett chooses the former while I choose the latter. So I must also make a proposal about the cultural analog to the phenotypes in the biological model, and I have done so in various places (e.g. in the case of music, in Beethoven’s Anvil, and more generally HERE and HERE, which I’ve bundled into these notes HERE).

While I have my reasons for doing this, I’m wondering if I could make a really strong argument, one the might even capture Dennett’s attention? I don’t know. Here’s a quick thought or two, but not yet an argument.

One issue that’s come up in memetics discussions – here I’m thinking of some online discussions that took place on a memetics listserve back in the late 1990s – starts from the observation that cultural phenomena exist as relatively small things, like bricks, words, and notes, and relatively large things, like cathedrals, novels, and symphonies. How do we treat that? Some call the little things memes and the big things memeplexes. If you do this, then you don’t need special cultural analog for phenotypes as memeplex seems to fill the bill. Others, like me, want to think of the relatively large entities as cultural phenotypes (I don’t know of any word that’s been specifically coined to fill this conceptual slot).

However, while it is true the genes are smaller than phenotypes, even for the smallest cell, the distinction isn’t merely one of size. More importantly, it’s one of function. Yes, in phylogeny, genes replicated from one generation to the next. In ontogeny, however, and in ongoing maintenance, they code for the proteins out of which organisms are constructed and they perform roles in regulating that construction. It’s this distinction, in function, that’s important to me.

That’s the argument I’ve been pondering the last few days: How do I argue for a functional distinction between memes and cultural phenotypes? In particular, how do I formulate that distinction in the case of language? Can I argue, for example, that phonemes are memes, while words themselves are phenotypes? And when I talk of words in that previous sentences I don’t mean morphemes, but fullblown words, including syntatic, semantic, and pragmatic features all “anchored” by a string of spoken syllables or written letters. And if I can make that argument, what would it take to generalize across all of culture?

More later.

Perhaps.

However, most are boring old stock images.  So, I’m setting a challenge:  who can find the most awesome ‘broken telephone’ picture?

This is my submission:

Image by Craig Barritt / Getty Images, found at The 45 Most Legendary Pictures Ever Taken.

As the title indicates, Dennett focuses his attention on words and does so in a way that usefully brings their mystery, if you will, though mystery is rather low on Dennett’s intellectual agenda.

What then are words? Do they even exist? This might seem to be a fatuous philosophical question, composed as it is of the very items it asks about, but it is, in fact, exactly as serious and contentious as the claim that genes do or do not really exist. Yes, of course, there are sequences of nucleotides on DNA molecules, but does the concept of a gene actually succeed (in any of its rival formulations) in finding a perspicuous rendering of the important patterns amidst all that molecular complexity? If so, there are genes; if not, then genes will in due course get thrown on the trash heap of science along with phlogiston and the ether, no matter how robust and obviously existing they seem to us today.

For what it’s worth, I have it on good authority that there are languages which lack a word corresponding to our concept of word, though they generally have a word roughly corresponding to our concept of utterance (you can find this observation in, e.g., Alfred Lord, The Singer of Tales). That doesn’t bear directly on the point Dennett is making in those words as lacking a word for this is that really existing phenomenon is common enough, but it does indicate that words do have a rather diffuse or abstract character that makes it difficult to understand what they are and how they operate.

A bit later Dennett continues:

A promise or a libel or a poem is identified by the words that compose it, not by the trails of ink or bursts of sound that secure the occurrence of those words. Words themselves have physical “tokens” (composed of uttered or heard phonemes, seen in trails of ink or glass tubes of excited neon or grooves carved in marble), and so do genes, but these tokens are a relatively superficial part or aspect of these remarkable information structures, capable of being replicated, combined into elaborate semantic complexes known as sentences, and capable in turn of provoking cognitive, emotional, and behavioral responses of tremendous power and subtly.

I particularly like his phrase in that first sentence, “the trails of ink or bursts of sound that secure the occurrence of those words.” That secure the occurence, that’s nice. “Anchor” might also work, that anchor the occurence of those words in an utterance or a written text, as though the ink or sound were a tether holding the airy nothings of meaning and syntax to the ground.

Still later in the argument, now that he’s gotten us to think more deeply about words, he makes a point I’ve been making for years:

The “syntactocentric” (Jackendoff 2002) perspective on language that has dominated theoretical linguistics since the pioneering efforts of Chomsky (see, e.g., Chomsky 1957, 1980) tends to obscure the fact that words have an identity that is to a considerable extent language-independent. Like lateral or horizontal gene transfer, lateral word transfer is a ubiquitous feature, and it complicates the efforts of those who try to identify languages and place them unequivocally in glossogenetic trees. English and French, for instance, share no ancestor later than proto-Indo-European (see Fig. 2) but have many words in common that have migrated back and forth since their divergence (cul-de-sac and baton, le rosbif and le football, among thousands of others). Just as gene lineages prove to be more susceptible to analysis than organism lineages, especially when we try to extend the tree of life image back before the origin of eukaryotes (W.F. Doolittle, this volume), so word lineages are more tractable and nonarbitrary than language lineages.

Historical linguists arrange languages into trees, mostly though by no means entirely, on the basis of sounds. This obscures the fact words move rather freely horizontally from one branch of a tree to another and pretty much neglects the well-attested phenomenon of creolization in which, in a handful of generations, a new language can emerge from quite distinct and otherwise unrelated languages.

Nor is such horizontal transmission of hybridization confined to languages. Consider what happened in the Americas when musical traditions from West Africa met with musical traditions from Europe: ragtime, the blues, jazz, rock and roll, salsa, hip hop, tango, and so forth. The same historical movements crossed Catholicism with West African animism to yield voodoo, candomble, santaria and other syncretic belief systems.

More broadly, Dennett points out that thinking of culture as an evolutionary phenomenon makes as more alive to accident and happenstance than what he calls “the traditional wisdom”

according to which culture is composed of various valuable practices and artifacts, inherited treasures, in effect, that are recognized as such (for the most part) and transmitted deliberately (and for good reasons) from generation to generation. Cultural innovations that are intelligently designed are esteemed, protected, tinkered with, and passed on to the next generation, whereas accidental or inadvertent combinations of either action or material are discarded or ignored as junk. This is basically an economic model, where possessions, both individual and communal, are preserved, repaired, and handed down. This familiar perspective on culture is for the most part uncritically adopted by cultural historians, anthropologists, and other theorists, and it meshes nicely, it seems, with evolutionary biology. Cultural innovations, like genetic innovations, have to “pay for themselves” to survive, by providing a fitness boost to their possessors. A new way of catching fish, whether genetically transmitted as an innate instinct or cultural transmitted as a learned practice, will go to fixation only if it is better than the old ways of catching fish.

This is the common view, though I’m not so sure as Dennett that it is uncritically adopted by cultural historians and others, not these days.

Human history is made by people who specific intentions. Some of those intentions are realized, but many are not. The behaviors, ideas, and artifacts that one generation chooses to adopt from an earlier generation to not necessarily reflect the desires and intentions of that earlier generation.
reflect inadvertance and happenstance along with deliberate intention. Nor do we need to confine our view to a time scale measured in generations. In the domain of popular culture, for example, the songs, movies, books, games, etc. that survive from one year to the text cannot predicted or controlled, not even by corporate conglomerates spending 10s and 100s of millions to do just that, control the flow of cultural goods–Arthur Devany’s Hollywood Economics, for example, documents that story in considerable detail for America’s feature-film industry in the second half of 20th Century. But this intermixture of intention and accident is hardly confined to popular culture. It is, as Dennett asserts, ubiquitous, and can be seen in the plethora of “small” elements making up culture, the words, riffs, motifs, gestures, etc. and in the “large,” books and schools of thought, paintings and cathedrals, sporting events and factories.

* * * * *

I’ve written quite a bit about cultural evolution. In this particular context, however, I recommend my working paper on the Xanadu meme, which uses the web to trace one word, “Xanadu”, through 200 years of recent history.

The study uses a series of regressions to demonstrate robust correlations between grammatical gender and various economic variables from a range of databases.  The gender variables include whether a language has a sex-based gender system, how the genders are used in pronouns, the intensity of the gender system (languages with 2 genders vs languages with 1 or more than 2 genders) and whether gender is assigned semantically or formally.  The correlations control for geographical variables (distance from the equator), climate (tropics, frost days, access to the sea), history of colonisation, continent, religion and cultural beliefs and values.  The findings include statistics such as “Having a sex-based gender system decreases the female labor force participation rate by 13 pp % relative to the base-line value in countries with no gender system”.

The approach is very similar to Keith Chen’s study of future tense and economic savings behaviour, and uses some of the same data including the world atlas of language structures (WALS) and the World Values Survey.  Indeed, Gay et al. find that “women living in countries whose dominant language marks gender more intensively are less likely than men to save”.  The paper follows other studies on the cultural transmission of agricultural technology and the role of women in society (Alesina, Guiliano & Nunn, 2011, see here).

Gay et al. argue that this is evidence for a Whorfian effect of language on thought:  distinguishing gender in language leads people to think about genders differently, and therefore for the culture to distinguish opportunities for men and women to a greater extent.

“These findings are relevant to several long-debated questions about the origin of language, its evolution, and its effect on speakers…. We … show that language influences female labor force participation even after we control for current beliefs and values, for historical agricultural practices”

Why not argue the other way around?  Gay et al. argue that, because grammatical gender is a very stable linguistic feature, it’s unlikely that economic phenomena are influencing grammar.  The authors cite Wichmann and Holman (2009) who find gender to be very stable. However, in a very interesting paper, Dediu & Cysouw (2013) show that the number of grammatical genders is only within the top third of the most stable linguistic features (see below, number of grammatical genders marked with a red box).

The evolutionary explanation of Gay et al. includes Quiatt’s idea that the division of labor between men hunting and women gathering in early societies was a driving force in the emergence of language.   Together with the stability of gender, the authors recognise an alternative hypothesis that some aspects of culture change slowly, and (presumably) that gender inequality and grammatical gender may just have been co-inherited.  However, this hypothesis isn’t pursued.

The authors claim that they control for historical factors, but this just boils down to the use of the plough since the 17th century and whether the country was part of a colonial empire.  Like in Chen’s paper, there is no attempt to explicit deal with the non-independence of cultures and languages (Galton’s problem, James and I are working with Chen on doing some of these analyses for his hypothesis).  Phylogenetic comparative methods should be used to test this hypothesis (and are relatively easy to carry out with R packages caper and ape).  There seems to be little appreciation of the time-depth of languages in general, from comments like the 17th century being referred to as the ‘distant past’ and a confusing citation of evidence that grammatical gender existed as far back as the 5th century BC.

The authors do cluster the results by language family.  Although there are few details of the results in the paper, they indicate that language family has a strong effect on the results, but that this indicates that “the impact of gender marking intensity on socio-economic outcomes comes both from across and within linguistic family variations in languages grammatical structure”.

Population size might also be another confounding factor that the authors do not consider, since larger populations are more likely to be in contact with other languages, which can cause morphology to simplify (see Lupyan and Dale, 2010 and Hanna’s post on foreigner-directed speech).

I am quite skeptical of this paper.  James and I have written about spurious correlations and problems with correlational studies of cultural phenomena and this paper follows the recent trend of generating hypotheses from simple correlations.  The data and methodology are very similar to Chen’s work and it’s worrying to think that, since these studies are becoming easier to run, and since they’re being presented online before peer-review, erroneous or misleading hypotheses could gain the attention of the public or of policy makers.  Last month, several flaws were uncovered in a widely-cited large-scale cross-cultural economics study, but not before it had been used as testimony before the US senate budget committee in order to support budget cuts.

I can imagine this paper angering many people.  To put it cynically, it’s as if gender inequality is only due to humans being slaves to their language, rather than centuries of active patriarchal societies.  The hypothesis doesn’t seem to have a good reason why distinctions in gender should disfavour women over men.  Perhaps most disturbing is the authors’ clear appeal for these findings to be used in policy:

“The direct and possibly cognitive influence of a language on its speakers and on economic life may have important policy implications. For instance, understanding this connection would facilitate the debate regarding the need to implement quotas or to opt for market forces to drive women economic participation up and thereby increase overall economic prosperity.”

The authors suggest that this can be used to improve access to economic and political power for women, but how do they think the study will contribute to immigration quotas?  Could this argument be used to stop female economic migrants who speak languages with many genders, on the grounds that they are less likely to contribute less to the economy?

Whorfian hypotheses generated from large-scale cross cultural statistics are poorly understood on many levels, and it seems too easy to find ‘evidence’ that fits a particular agenda.  This kind of study is potentially dangerous, but it will no doubt garner media attention anyway.

Victor Gay, Estefania Santacreu-Vasut and Amir Shoham (2013). The Grammatical Origins of Gender Roles Berkeley Economic History Laboratory (BEHL) Working Papers  Online

“Music and the Origins of Language. International Summer School on Agent-based Computational Models of Creativity”.

15 – 20 September 2013, Cortona, Italy
http://ai.vub.ac.be/events/cortona-2013

For information and queries, please visit the website http://ai.vub.ac.be/events/cortona-2013/ or email cortona2013@ai.vub.ac.be.

Since its resurgence in the 90s Multi-agent models have been a close companion of evolutionary linguistics (which for me subsumes both the study of the evolution of Language with a capital L as well as language evolution, i.e. evolutionary approaches to language change). I’d probably go as far as saying that the early models, oozing with exciting emergent phenomena, actually helped in sparking this increased interest in the first place! But since multi-agent modelling is more of a ‘tool’ rather than a self-contained discipline, there don’t seem to be any guides on what makes a model ‘good’ or ‘bad’. Even more importantly, models are hardly ever reviewed or discussed on their own merits, but only in the context of specific papers and the specific claims that they are supposed to support.

This lack of discussion about models per se can make it difficult for non-specialist readers to evaluate whether a certain type of model is actually suitable to address the questions at hand, and whether the interpretation of the model’s results actually warrants the conclusions of the paper. At its worst this can render the modelling literature inaccessible to the non-modeller, which is clearly not the point. So I thought I’d share my 2 cents on the topic by scrutinising a few modelling papers and highlight some caveats, and hopefully also to serve as a guide to the aspiring modeller!

Anybody who’s ever gotten her hands dirty playing around with a multi-agent simulation will know how awesome it feels to send a program off for execution and wait a few seconds for it to produce a mountain of data that would have taken weeks or months to collect through classic psychological experimenting. This is the awesomeness of computational modelling, but it is also a curse. Because however tempting it is to run 1.000 instances of a stochastic simulation for 100.000 iterations each in 3.072 different conditions, all that data that is so easily produced will not only clog up your hard disk space, but it will also still need analysing.

The source of this luxury problem is that, rather than meticulously setting up two or maybe four conditions in which a human subject, acting as a blackbox, will hopefully perform in significantly different ways, it is very tempting (and admittedly very interesting!) for modellers to think about all the possible factors that might or might not lead to different behaviours, and implement them. More interestingly, because the unknown variable in multi-agent simulations are the interactions and cumulative effects we are free to not only vary environmental variables like population size and interaction patterns, but also control every little aspect of the agents’ individual behaviour. This means that in comparison to psychological experiments computational models will typically have many more parameters, many of which are continuous, leading to more and also more complex interactions between all of those parameters.

So no matter how ‘complete’ or ‘realistic’ you would like to make your model, any nonessential parameter n+1 that you put in will strike back in the form of a multiplicity of the amount of data and n (potentially nonlinear) interactions with all the other parameters which will be a pain to analyse, to fit, let alone to understand.

The first and most obvious point is of course that you ought to study (or in the case of modelling, make up) “the simplest system you think has the properties you are interested in” (Platt 1964), i.e. keep the number of parameters low! Having fewer parameters does not only allow you to sample the parameter space more exhaustively, it also makes eyeballing the data for interactions a lot easier, which can be followed up by some informed fitting of a descriptive model to the data.

A great (but rare) example of such an approach is the rigorous analysis of the Minimal strategy in the Naming Game by (Baronchelli et al. 2006). Their data is abundant, their model fits quantified, their figures mesmerising – and all that to study the impact of a single parameter (population size). I am not aware of any equally exhaustive and convincing studies of similarly interesting models investigating more than one parameter, most likely because precisely quantifying their interactions becomes complicated very quickly.

What I’m talking about here in terms of tractability are the natural limitations of what is called numerical modelling. Numerical modelling means that we determine the state that a model is in by using an incremental time-stepping procedure, which we have to iterate to learn about the development of the model over time. The results we get from such models are (as the name suggests) purely numerical, so if you want to study the model’s behaviour using different initial conditions you have to feed in different numbers at the start and run the same time-stepping procedure all over again. What’s worse, our updating function will usually contain some nondeterministic component, e.g. to randomise the order in which agents interact, which means that we will have to run the same condition multiple times to determine the interaction between the randomness and the model’s behaviour. The only exception to this general rule is when there is a transparent relationship between the probability distribution of the random variable and our updating function. In such cases we can sometimes run a numerical simulation of the development of the probability distribution over model states instead of running individual instances of our model, but this will generally not be the case. In most cases there will be nontrivial interactions where small quantitative differences in the random variables will lead to qualitatively different model behaviour, so we have to resort to running a large number of iterations using the same initial conditions just to get an idea of how our random influences affect the model.

That’s what makes numerical models a bit of a pain. They are (too) easy to run and you can implement almost anything you like with them, but they can be an absolute pain to analyse and, more problematically, it’s very difficult to draw any strict conclusions from them. This is probably also the reason why, following a frantic period of computational modelling from the mid-90s to the mid-00s, there have been significantly less simulations of evolutionary linguistics stuff carried out (or at least published) since then. Following the initial excitement over the very first results came a bit of a modelling hangover, most likely caused by some sort of resignation over just what we’re supposed to make of their results (de Boer 2012).

There is an alternative to numerical models though, and one that I would say is preferrable wherever possible, namely analytical modelling. Analytical just means that whatever model we happen to have cooked up, complete with parameter-controlled individual behaviour, parameter-controlled interactions and possibly some random components as well, has a mathematical closed form solution. “What on earth does that even mean?” I hear you asking. That just means that there is a solution to the equations which describe the changes to the state of our model (the time-stepping or update function that is called on every iteration in a numerical model) that can be expressed as an analytic function. If we were to squeeze our numerical model into an equation, it would most likely look something like xt = fα,λ,θ,...(xt-1), where xt captures the state of the model (i.e. the current prevalence of different traits/the agents and the languages they currently speak…) at time t and f applies the changes with which we update our population at every time step, the details of which are controlled by the parameters α, λ, θ etc. The observant reader will have noticed that this equation contains a recursion to xt-1, so if we want to know the state of our model after t iterations we have to first compute the state after t-1 iterations all the way down to 0, where we can just refer to the initial state x0 which is given. Evaluating this formula for a specific combination of parameters is basically what we do when we run a computer simulation.

Now the art of mathematical modelling is to transform this equation into something that looks like x(t) = f(x0, t, α, λ, θ,...) where f has to be a function of the given parameters containing only simple mathematical operations like addition, multiplication, the odd exponential or trigonometric function, but crucially no recursion to f(). If we find such an f then this is the analytical solution to our model, and if we want to know the state of our model after a certain number of iterations (note how t has become a simple parameter of f) given certain parameters and initial conditions we just have to fill in those values and perform a straightforward evaluation of this simple formula.

What’s more, because our parameters are explicit in f(), we might even be able to read off their individual effects on the overall dynamics without having to try out specific values. With a bit of luck we will be able to tell which parameters shift our system state x in which direction, whether their influence on the model is linear, quadratic, exponential, just constant, or whether they don’t affect the dynamics at all. (This might seem bizarre but I know of at least two papers covering computer simulations which discuss aspects of the model that are in fact completely irrelevant for the overall dynamics. One of them contains a parameter setting which completely cuts off the influences from another part of the model, but the results are still discussed as if the influence was there. The other contains a description of an aspect of the model that important influence is ascribed to, but the condition that would cause this part of the model to become active are never fulfilled, i.e. it describes bits of the simulation code that can in fact never be reached. In both cases it doesn’t seem like the authors have realised this. Spelling a model out in mathematical equations might not always be possible but it is certainly worthwhile to try and transcend purely verbal description to make it clearer what is actually going on.) Where was I? Oh yes, analytical models might even allow you to see whether the influence of specific parameters increases or decreases over time, whether the model converges towards some increasingly stable end state or whether it keeps on fluctuating, possibly exhibiting periodic behaviour etc etc.

It is important to note that analytical and numerical modelling are not two incompatible things. The ‘model’ itself, i.e. a description of how a system of certain shape changes over time, precedes both, analytical and numerical modelling are merely two different ways to figure out what the predictions of that model are. It is true however that some decisions in drawing up the model are taken with a particular kind of modelling approach in mind. Some of the core components of any model are going to be motivated or justified by common sense (an agent encountering a new convention should presumably add this form to his memory rather than, say, ignore the form but nevertheless scramble or invert the scores of all the conventions he already has in his memory), but other decisions are pretty arbitrary (de Boer 2012), in particular when it comes to how exactly continuous values/variables are updated. So while in numerical simulations you will tend to go for operations which are computationally cheap (examples being any multi-agent simulations of conventionalisation where scores of competing variants are simply incremented or decremented, still the cheapest operations on modern computers), analytic models are likely to make use of functions which might be computationally unwieldy, but mathematically very well understood (examples for this would be the use of an exponential distance function in (Nowak et al. 1999, p.2132), or the definition of the sample prediction function in Reali & Griffiths 2010, p.431-432).

It is natural to be a bit puzzled or sceptical when you come across those arbitrary choices made in papers – would the model behave completely different if they’d chosen a slightly different function? But one has to concede that in coming up with a model, some arbitrary choices have to be taken anyway. And if that’s the case then they might as well be taken in a way that simplifies whatever you’re doing, whether it’s numerical computation or mathematical analysis – because without that guided choice we might not have any results at all.

So while a lot of models can be implemented both numerically and analytically (which is a great way to cross-check their results!), some are only amenable to either approach. For many complex models analytical solutions are not obtainable, but conversely some analytical results based on infinite population sizes or continuous time models (where the updates between model states become infinitesimally small) can simply not be captured by running a discrete time-stepping function. But such models are the exception, and in most cases what will happen is that you implement and tinker with a numerical model first and, stumbling upon interesting results, attempt to figure out whether the behaviour can be captured and predicted analytically which would, if you’re lucky, relieve you from having to exhaustively run those 3.071 other conditions as well.

A brilliant example of how numerical and analytical models often coincide can be found in Jäger 2008 (link below) in which he shows that the behaviour of two seemingly disparate numerical models of language evolution (one an exemplar theory model of the evolution of phoneme categories a la Wedel/Pierrehumbert, the other (Nowak et al. 2001)’s evolutionary model of iterated grammar acquisition) can in fact both be captured by the Price equation, which is a general solution to predict the development of any system undergoing change by replication (i.e. evolution). The paper is a great read and easily understandable for anyone who’s not afraid of a bit of stats, so I highly recommend having a look at it!

To summarise, having an analytical model simply adds concision and predictive power to your results. But if analytical modelling is oh so awesome, why isn’t everyone doing it then? The first part of the problem is that a lot of models, and certainly a lot of the more interesting and realistic ones, simply don’t have such a closed-form solution. The ones that are solvable lead to powerful and irrevocable results though and it is very instructive to try and reconstruct them yourself, something that most biologists and the odd economist (but who cares about economists?) but sadly not enough social scientists will have some experience with.

A second problem is that even if an analytic solution is in principle obtainable, it is still far from trivial to do so, and pretty futile to attempt unless you have a background in maths or physics (disclaimer: I have neither and I have so far also not produced any analytical models myself).

The crux is that while not a lot of models are exactly solvable, many of them can be approximated by making certain assumptions, which is where we can finally look at the tradeoff between numerical and analytical models! In the aforementioned case where there is some stochastic component in the system (i.e. some aspect of the system is specified by a probability distribution rather than a value), any analytical solution would have to be a function transforming this distribution into a another distribution, which can get hairy very very quickly. So while it will typically not be possible to obtain a description of the entire distribution over model states, we can try to get at a description of the average state that the model will be in. This is often done based on so-called mean field assumptions, i.e. by disregarding the fluctuations around this average caused by stochastic interferences and sampling effects.

Most analytical models you encounter in the literature will rely on such approximations, and it is important to realise that they influence how we may interpret their results. This is something that can easily get lost in the quick succession of mathematical transformations and unwieldy equations full of greek characters. Most readers without a background in maths have no choice but to either disregard such models because they cannot evaluate their validity themselves, or otherwise accept their conclusions at face value. Both of these options are far from ideal of course and, given the merits of analytic models that I outlined above, it is absolutely worthwhile to consider those models and try to understand what their approximations mean for the results.

Funnily enough all this was just intended to be the introduction to a post in which I wanted to have a closer look at two analytical modelling papers and their conclusions, but since it might have some value by itself I think I’m gonna leave it at this for now! So check back in a week or so for some applied model dissection!

References

Baronchelli, Andrea, Felici, Maddalena, Loreto, Vittorio, Caglioti, Emanuele, & Steels, Luc (2006). Sharp transition towards shared vocabularies in multi-agent systems Journal of Statistical Mechanics: Theory and Experiment, 2006 (06) DOI: 10.1088/1742-5468/2006/06/P06014
de Boer, Bart (2012). Modelling and Language Evolution: Beyond Fact-Free Science Five Approaches to Language Evolution: Proceedings of the Workshops of the 9th International Conference on the Evolution of Language, 83-92
Jäger, Gerhard (2008). Language evolution and George Price’s “General Theory of Selection” Language in flux: dialogue coordination, language variation, change and evolution, Communication, mind & language 1, 53-80 Other: http://www.sfs.uni-tuebingen.de/~gjaeger/publications/leverhulme.pdf
Platt, John R (1964). Strong inference Science, 146 (3642), 347-353 : 10.1126/science.146.3642.347
Nowak Martin A, Krakauer David C, & Dress Andreas (1999). An error limit for the evolution of language Proceedings of the Royal Society B: Biological Sciences, 266 (1433), 2131-2136 PMID: 10902547
Nowak, Martin A, Komarova, Natalia L, & Niyogi, Partha (2001). Evolution of universal grammar Science, 291 (5501), 114-118 PMID: 11141560
Reali, Florencia, & Griffiths, Thomas L (2009). Words as alleles: connecting language evolution with Bayesian learners to models of genetic drift Proceedings of the Royal Society B: Biological Sciences, 277 (1680), 429-436 DOI: 10.1098/rspb.2009.1513

Cast your reminisce pods back a few days and recall Sean’s iterated learning experiment using the automated transcription of YouTube videos. The process went as follows:

1. Record yourself saying something.
2. Upload the video to YouTube
3. Let it be automatically transcribed (usually takes about 10 minutes for a short video)
4. Record yourself saying the text from the automatic transcription
5. Go to 2

Sean took a short extract from Kafka’s Metamorphosis and found that, as in human iterated learning experiments, both the error rate and compression ratio decreases with successive iterations. He also found that the process resulted in a text with longer and more unique words.

I was curious to see whether we could remove human participants entirely and run computer generated speech through this automated transcription. Here’s the process:

1. Generate an audio file from some text using a speech synthesis program;
2. Generate a transcription of the audio file;
3. Repeat from 1. with the new transcription.

For speech synthesis I used Festival, and for transcription I used an undocumented Google API that is presumably used in the YouTube transcription (here’s the code: ilm.sh). One thing to note is that this API only accepts short files, so longer passages had to be bitten off into red ball like sections.

Here’s the first few sentences of Metamorphosis:

One morning, as Gregor Samsa was waking up from anxious dreams, he discovered that in his bed he had been changed into a monstrous verminous bug. He lay on his armour-hard back and saw, as he lifted his head up a little, his brown, arched abdomen divided up into rigid bow-like sections. From this height the blanket, just about ready to slide off completely, could hardly stay in place. His numerous legs, pitifully thin in comparison to the rest of his circumference, flickered helplessly before his eyes.

And here’s what we get after one iteration:

wine bar in the ass cracker Samsung was waking up from my chest remains the discover about in his bed he had been changed into a monstrous reminisce pod. delay on his armor hard back and saw as he lifted his head off a little is brown autobell me to bite it off into red ball like sections. from the title blanket just about ready to slide off completely could hardly stay in place. his numerous legs pitifully Finn in comparison to the rest of his circumference Flipkart helplessly before his eyes

Given that there is far less variability in computer generated than human produced speech, we would expect that the process would rapidly converge to little or no error. After only five iterations, we get a stable state with no transcription error and the text:

wine bar in class cracker Samsung was waking up from my chest remains the discover about in his bed he has been changed into a monster some other spot. delay on his armor hardback and so has he left his head off a little is brown Autoblow me to bite it off at the Red Bull like sections. from the title blanket just about ready to slide off completely could hardly stay in place. his numerous legs pitifully Finn and comparison to the rest of his circumference flipped out Leslie before his eyes.

What’s immediately interesting is the number of brand names that are produced, in this and other samples I tried. Google being Google this is perhaps not surprising, but it does give a further insight into the biases of their algorithms and the limitations for their use as trustworthy models in these kinds of experiments.  I would suggest the rule that anything repeatedly passed through an advertising company rapidly converges to spam.

As with Sean’s production, the error rate decreases with successive iterations. In contrast to Sean’s production, however, the compression ratio actually increases.  In addition, this process resulted in an increase in the number of words and decrease in average word length. This suggests that the bias for word length might be more affected by the production than the transcription.

However, since Metamorphosis was originally written by Kafka in German, the passage we used had already passed through at least one level of production and transcription before we started. Here’s the actual original passage:

Als Gregor Samsa eines Morgens aus unruhigen Träumen erwachte, fand er sich in seinem Bett zu einem ungeheueren Ungeziefer verwandelt. Er lag auf seinem panzerartig harten Rücken und sah, wenn er den Kopf ein wenig hob, seinen gewölbten, braunen, von bogenförmigen Versteifungen geteilten Bauch, auf dessen Höhe sich die Bettdecke, zum gänzlichen Niedergleiten bereit, kaum noch erhalten konnte. Seine vielen, im Vergleich zu seinem sonstigen Umfang kläglich dünnen Beine flimmerten ihm hilflos vor den Augen.

And here’s what we get after nine iterations of German production and transcription (note that because of the limitations of the API discussed above I had to break the text not just at sentence boundaries but also after the phrase “geteilten Bauch”):

Amazon sind morgen Samsung Wave 2 Vapiano Markt Pforzheim Bett beiden Mannschaften. Wetter Santander an den Weihnachtsmann corporate website and merry christmas for my life and style, auf der Schanz jetzt ein Bett bei dem Fensterbrett steht gar nicht arbeiten bei den wecker how to upgrade. browser Android erscheint Samstag im Fernsehen Transformation morgen.

Again with the spam, and we still see a decrease in error and increase in compression ratio.

It differs from the English chain in other ways. German takes longer to converge (nine generations vs. five) and results in fewer words than the original text, though it’s unclear whether these words are significantly longer or shorter

Typically in iterated learning experiments we don’t begin with such structured initial data, but with random unstructured strings like the following:

nehomami wumaleli mahomine maholi wupa wuneho lemi manehowu nepa wunene maho howu nemine lemilipo hopa lemipo nehowu nemi lipapo lilema pohomali pamamapo wulepami lepali poliho powuma lemaho

(Kirby, Cornish & Smith, 2008)

Taking 30 random permutations of the above initial data I ran 30 chains for ten generations in both English and German. The above string itself yields:

I hope mommy will email the holy holy holy place the whole house full of people no hold me like a polar pop the Molly pop over the popular poly-poly hope out of my hoe

We see similar trends here as for the literary samples above.  Both languages show a decrease in error and increase in compression ratio over generations.  For the English translation chains, the number of words increases while the average word length decreases. For the German translation, the number of words decreases, while the average word length stays more or less constant.

For the English transcription

For the German transcription

Using computational speech synthesis reduces the variability in production, and allows us to more accurately investigate the inductive biases of the transcription algorithm. The differences we see in English and German transcription suggest that many biases may be introduced not through some hard coded algorithm but by the data used to train it. Much like human language.

Merry christmas for my life and style,
Justin