baBouma - Monkeys learn to tell words from nonsense letters

[mike_duggan]

This topic was imported from the Typophile platform

https://web.archive.org/web/20120415232740/http://www.dispatch.com/content/stories/national_world/2012/04/12/monkeys-learn-to-tell-words-from-nonsense-letters.html

[froo]

This explains many things. Thank you.

[jacobsievers]

I see an opportunity here: fonts for baboons?

Frankly, I don't get it. How is distinguishing learned letter groups from unlearned ones anything like reading? Their conclusions overreach, I think.

[John Hudson]

How is distinguishing learned letter groups from unlearned ones anything like reading?

That's not what happened. In the learning phase, the baboons were exposed to sets of words and non-words, but in the testing phase they were exposed to different sets that they had not seen before. So what they were demonstrating was the ability to recognise patterns in how letters are ordered or not ordered in English words, what the researchers call orthographical processing. It doesn't surprise me that this level of orthographical processing is independent of language knowledge, because I can easily tell the difference between, say, real Russian text and random arrangements of Cyrillic letters, even though I can't read the language. But I am surprised -- and delighted -- that baboons can do the same sort of thing, especially since they're primates that are not even in the hominid line, which means that a shared perceptual functionality that makes this possible must have evolved an incredibly long time ago.

The comments at the end of the article seem misleading, though, or at least poorly expressed. 'Recognising the entire word'? If anything, in this regard, the study suggests that what is being recognised is patterns derived from letter ordering principles. Entire words can't be being recognised because the baboons don't know the words. What they are recognising is patterns of letters that correspond to learned rules.

[hrant]

"More research is needed to determine whether this technique can be used to improve the reading skills of teenage humans."

Actually Jacob, I think it's a nice bit of proof that recognizing abstract, arbitrary patterns that mean something (in this case "I get food") requires no formal education, that it's strongly related to the brain's "firmware".

In fact what I find extremely interesting is the possibility of seeing what kinds of letterforms (think sans versus serif) primates do better at! And I would also test crows.

hhp

[Christopher Dean]

The databases gave me the same article.

[enne_son]

There seems to be an online pdf of the study that is described in the The Columbus Dispatch story:
Orthographic Processing in Baboons (Papio papio)

[enne_son]

Yes, the paper is available for a fee in other databases, but the link I gave is to on open access copy on a Cornell database of readings for a course in perception, cognition, and development. I provided the link so typophiles can examine how Grainger et. al., come to the conclusion that the baboons were “computing an orthographic code.”

Using results from primate performance, Grainger et. al., try to make the case “that orthographic processing may, at least partly, be constrained by general principles of visual object processing shared by monkeys and humans.” I've been arguing for some time, as Grainger et. al., do here that the perceptual proccessing dimension, in other words, what happens in the visual cortex, is often under-explored by cognitive scientists.

To make their case that baboons are perhaps computing an orthographic code, Grainger et. al., point to analogies in performance statistics between the baboons and humans, and cite the work of S. L. Brincat, C. E. Connor, on how the macaque brain synthesizes low-level neural signals for simple shape parts (simple contour fragments) into coherent representations of complete objects. The idea of Grainger et. al., is that full letters are analogous to these parts.

An alternate explanation that seems to be a more “ecologically valid” extension of the S. L. Brincat, C. E. Connor analysis might be that the baboons learned to recognize words from what we might call word-specific trans-glyphic featural patterns, rather than as orthographic sequences, that is, rather than “coding the word and non-word stimuli as a set of letter identities arranged in a particular order,” the results might have been based on across-the-word feature-conjunction similarities.

[John Hudson]

Peter, yes, that alternative explanation is possible, as it is for human reading. Of course, the baboons are performing a different task than the human reader: they're not recognising words as semantic objects, they're only distinguishing between patterns of ordering, which is something that the human reader has to do too, but only as a sub-task process. In the past, we've posited super-letter pattern identification as an efficiency, something that makes it possible for us to get to where we really need to be, which is word recognition, faster than letterwise recognition. The baboons, it seems to me, don't need that efficiency, recognising patterns in letter order is where they need to be to complete the task. And I'm not convinced that 'word-specific trans-glyphic featural patterns' are more 'ecologically valid' as extensions of Brincat & Connor's findings than, say, starting from even lower level, intra-glyph feature recognition, e.g. recognising the relationship between stem and bowl in /b/ as constituting a distinct object, then distinguishing the order of that object in a sequence such as 'blob' from a sequence such as 'lbbo'.

Do you think there is any way to isolate 'across-the-word feature-conjunction' patterns from letter ordering patterns in a way that would make it possible to test the functional distinction?

[enne_son]

[John] “The baboons, it seems to me, don't need that efficiency, recognising patterns in letter order is where they need to be to complete the task.”

I was suggesting recognizing patterns in across-the-word (trans-glyphemic) feature distribution is where baboons are to complete the task. I wasn't thinking of ‘supra-letter’ features, but of ‘sub-letter’ features. I doubt if baboons used letter identification to execute the task. They had no training in letter identification.

[enne_son]

[John] “Do you think there is any way to isolate 'across-the-word feature-conjunction' patterns from letter ordering patterns in a way that would make it possible to test the functional distinction?”

I think to do this effectively we first need to compare quantitatively the potential supercapicity benefits (because of interfacilitation) that would attach to redundancy when calculated at a letter level versus at a role-unit or sub-letter feature level. If these turn out to be distinct values, fits to performance variance in Word Superiority Effect tasks could be decisive, provided quantification of Word Superiority Effects could be made sensitive enough. Because of the limitations of screen refresh rates, actual tachistoscopes would need to be used.

Interfaciliation usually isn't in the picture in most models of perceptual processing in words, because channel independence at a feature and letter level is widely assumed. Role units, as distinct from both features in a strict and primitive sense, and letters, aren't even recognized.

[John Hudson]

I wasn't thinking of ‘supra-letter’ features, but of ‘sub-letter’ features.

Ah. I had misinterpreted 'trans-glyphic'. So you mean gestalt recognition of sub-letter features in super-letter patterns, bypassing the letter-as-object level, yes? This seems reasonable. I wouldn't discount the possibility, though, that the baboons make the step of recognising the combinations of features that make up letter objects, since it is the latter that are actually being reordered in the word objects, i.e. that the baboons could teach themselves letter recognition as part of the task.

[Ryan Maelhorn]

The thing I'm trying to wrap my head around is that if one were to start designing with idea that what we recognize word shapes before we recognize the actual word, or its actual letters; which I believe is true; then what we are really talking about here is taking into account almost as many different designs/overall-shapes as there are words. Now we are talking about designing something like 50,000 things, instead of the 150 or so that make up the standard latin unicode set. Do we really want to open this box?

Speaking of word shapes, would broom and bloom have the same shape? Or does the ascender of the l render it a different shape?

this all hurts my brains

[hrant]

I don't think you need to design that many different things; you try to design things that strike a balance of: what the current "average" bouma is for that word*; and what the bouma should be, to maximally diverge it form others. The balance you strike depends on: how progressive you want the font to be; and how quickly bouma familiarity is actually gained...

* Which is why a descending "f" in an Italic is bad.

would broom and bloom have the same shape?

No, but the deeper into the -increasingly blurry- parafovea, the more the boumas will converge. In fact this example is one small pointer to the need to make the "el" more than just a dumb stick (and the "r" wider than might seem to make sense on a formal, deliberative level). But that's just one pair of words; we're supposed to be worrying about entire lexicons...

hhp

[John Hudson]

Ryan, the notion that we recognise whole word shapes isn't given credence by anyone familiar with reading research. The open question is whether we recognise words by building up from letters or from combinations of letter features across the word. In this forum, I tend to argue the letterwise side, mainly because the most interesting and knowledgeable contributors to discussion make the case for various kinds of feature arrangement recognition.

[hrant]

It's true that the whole-word model is skewed and/or over-simple. But the belief that we read anything beyond individual letters (even just sometimes) is what opens up this whole hidden world that requires the sort of optimization that a real text face designer must worry about.

hhp

[enne_son]

I think ‘building up’ is a misleading metaphor. The visual cortex works on the principles of learned responsiveness and task optimizing channeling. It doesn't ‘rebuild’ the letters inside the brain out of tiny bytes supplied by rods and cones. With it's tuning mechanisms and form-resolution capabilities, the visual cortex (a many-layered thing) connects us dimensionally to the marks on the page in the way that best supports the tasks we need to do with these marks.

I like to say that in humans in the visual cortex, for material in the uncrowded foveally centered span, there is responsiveness to role-units, their local combinations (to form letters) and their across-the word distribution. These configurations of role units in local combinations and word-specific distributions are what the brain has 'representations’ for, for familiar words. Responsiveness to local combination accounts for the fact that we see words as lettered things, but without explicit labeling. As Edmund Burke Huey argued in 1908, there is an inhibition of incipient recognitions for letters.

Explicit labeling would be fatal, because it introduces the “[phonological] disambiguation overhead” issue, which some learners can get hung up on to the point of developing enduring reading difficulties.

‘By-parts’ or letter-wise processing imagines a chunking process that tends toward explicit labeling at a letter level, as a process where connected role-units are processed in hermetically-sealed slots or independant channels with no opportunities for across-the-word interfacilitation between role-units while the processing is going on, just (a computationally less efficient) interactive activation in hierarchical schemes.

While the human visual cortex learns a kind of role-unit-analytic processing in writing and forms distributional mappings at this level from repeated exposure during reading acquisition (which is largely and properly ‘phonological awareness’ based), I believe baboons stick with a kind of feature-analytic processing such as Frank Smith described, that responds to patterns of features properly so-called. Features are things like disturbed or undisturbed expressedness, or closure / absence of closure, in forms with round or oblong envelopes; aspect or extendedness, when it comes to items with lack of curvature. In printed words, configurations of these types of features make up distinctive patterns which can become familiar, to the point where similarity can be assessed, even by species without humanoid language capabilities.

[John Hudson]

Thanks, Peter. This is one of the clearest expositions of your ideas that I have read.

Explicit labeling would be fatal, because it introduces the “[phonological] disambiguation overhead” issue, which some learners can get hung up on to the point of developing enduring reading difficulties.

Fatal seems too strong a word. The fact that some readers get hung up on phonological disambiguation means that this issue is, in fact, introduced -- presumably strongly in the case of some learning methods --, but on the other hand only some learners have difficulty with it.

Of course, phonological disambiguation isn't an issue at all for the baboons, since they neither associate the visual objects with sounds nor with meaning. 'Bear' and 'Fear' are simply like-arrangements of sub-word objects to the baboons.

[oldnick]

'Bear' and 'Fear' are simply like-arrangements of sub-word objects to the baboons.

This observation should shade the takeaway from this exercise: the baboons were “reading” for rewards, not for meaning. Precisely which perceptual and cognitive systems are operative may be dependent on the purposes to which they are directed.

[enne_son]

John, perhaps fatal is too strong a word.

My view is that phonological awareness is a necessary and important part of reading literacy. And phonological mapping is a strategy we use almost automatically whenever we encounter (are stopped by) an unfamiliar word.

In reading acquisition we are taught a letter-wise, or bi-gram or tri-gram, etc., parsing of the word which sets up, so it seems to me, a parallel letter recognition / interactive activation-based processing architecture and neural representation.

Repeated exposure in a typographical spacing environment where there is 1) a distinctive narrow phase alignment that avoids auto-correlation in the vertical parts of strokes as well as distracting harmonics, and 2) rhythmic cohesion in the whites, creates a situation where, with non-word letter clusters, short-exposure foveal crowding would normally occur, but also — and because of the potential crowding — provides, in orthographically legal wards, a catalyst for a perceptual learning process that remaps words at a more primitive (than orthographic) visual level, the level I'm calling the role-unit level. Synaptic interfacilitation replaces interactive activation here.

We don't in my opinion arrive at effortless immersive reading until this remapping occurs.

So perfect spacing is beneficial to the development of skilled readers. For the rapid, automatic visual word-form resolution of skilled readers, though, explicit labelling at a letter level by the visual system, such as is encouraged in reading acquisition, can disturb the immersive reading process.

[John Hudson]

Again, very clearly stated. Thank you.

Will you be at the 'Reading Digital' conference at RIT this week? You should be.

[enne_son]

I had the RIT conference in my sights.I was sent a heads up a while ago, but unfortunately prior commitments have boxed me in.