[Humanist] 31.378 not as text but marks on pages; maps and networks

Humanist Discussion Group willard.mccarty at mccarty.org.uk
Tue Oct 24 07:07:17 CEST 2017

                 Humanist Discussion Group, Vol. 31, No. 378.
            Department of Digital Humanities, King's College London
                Submit to: humanist at lists.digitalhumanities.org

  [1]   From:    "Bradley, John" <john.bradley at kcl.ac.uk>                  (22)
        Subject: RE:  31.377 maps and networks?

  [2]   From:    "William L. Benzon" <bbenzon at mindspring.com>             (123)
        Subject: Re:  31.376 sustained reading from screen; not as text but
                marks on pages

        Date: Mon, 23 Oct 2017 09:59:36 +0000
        From: "Bradley, John" <john.bradley at kcl.ac.uk>
        Subject: RE:  31.377 maps and networks?
        In-Reply-To: <20171023062317.A9B468148 at s16382816.onlinehome-server.info>

About Paul's question below: I believe that my Pliny project (http://pliny.cch.kcl.ac.uk), now largely inactive, involved some thinking about how modelling approaches similar to Mindmapping might work in humanities scholarship.  It attempted to connect annotation into the issue, and someone might find the most recent publication that came out of the Pliny project in DHQ (http://www.digitalhumanities.org/dhq/vol/11/1/000279/000279.html), has something to say about this.

.. John Bradley

Visiting Senior Research Fellow,
Department of Digital Humanities,
King’s College London

-----Original Message-----

                 Humanist Discussion Group, Vol. 31, No. 377.
            Department of Digital Humanities, King's College London
                Submit to: humanist at lists.digitalhumanities.org

>        Date: Sun, 22 Oct 2017 16:07:24 -0500
>        From: Paul Fishwick <metaphorz at gmail.com>
>        Subject: mind maps/concept maps/semantic networks in the humanities?
>        In-Reply-To: <20171022065129.24CDE8110 at s16382816.onlinehome-server.info>

I noted Mike Cosgrave’s response on the “sustained reading from screen” thread, and was curious about:

Concept Maps
Semantic Networks

that humanist scholars and their students use. If you do use these can you specify how, and using what package (if any)? Exploring this practice seems a good way to bridge concepts in computer science to the humanities. Last year, I was intrigued to find out that Art History AP material in some US high schools use a “concept map” although this concept map is different than Novak’s map. The Art History map places an art work at the center and surrounds this work with arrows pointing to inquiries surrounding the work.


        Date: Mon, 23 Oct 2017 10:56:02 -0400
        From: "William L. Benzon" <bbenzon at mindspring.com>
        Subject: Re:  31.376 sustained reading from screen; not as text but marks on pages
        In-Reply-To: <20171023062221.CE952813E at s16382816.onlinehome-server.info>

Hi Tim,

Thanks for your thoughtful reply to my Babel post.

Yes, I know that DLTs work with pairs of texts where one is a human-translated version of the other. And, in that post, I’m not interested in DL systems that learn games, such as Go. Why not? Because those games are entirely self-contained. They aren’t about some world external to the game itself. Human texts, the ones translated by DLTs, are about the world. I should have been explicit about this.

“Classical” work in machine translation had hand-coded rules about language and meaning. Now, some of this work was pure engineering while other work sought to model human performance (insofar as we could figure it out). But, regardless of just what was done, the enterprise crumbled. But these more recent statistical systems, without hand-coded rules about syntax or semantics, they’ve succeeded.

THAT’s the contrast that interests me. What’s in those human language texts is an encounter between mind and world. Old style MT attempted to model the mind’s half of the interaction, and failed. These new systems attempt to model nothing, just the statistical regularities in the encounter between mind and world.

But just what is the mind’s component of the interaction between mind and world? Where does it come from? Much of it is the product of prior interaction between mind and world and I suppose we can trace that back through evolution as the nervous system itself reflects millions of years of interaction between living creatures and the world.

One thing in the back of my mind when thinking about these problems is this: What’s the world have to be like if it is to be intelligible? I’ve concluded that the world must be “lumpy.” A world that’s “smooth and continuous” would be very difficult to “parse”.

What do I mean?

Consider all the mammals. While a mouse is quite different from a rhinoceros, and both are different from a giraffe, not to mention a bobcat, in contrast to, say, a pine tree, they’re all rather like one another. That is to say, they’re mammals, not trees. Similarly, pine trees, oak trees, and palm trees are rather different from one another, but, in contrast to, say, a bobcat, they’re rather like one another. They’re trees, not mammals.

That’s what I mean by lumpy. And it’s not simply that mammals and trees look different, they also have very different modes of temporal existence, if you will. Mammals move around under their own power; trees stay put. And so forth. We’ve got lumpiness in the temporal domain.

A lumpy world has scattered regions each of which is internally dense with activity. But the space between these regions is all but empty.

Whatever it is that the DLT system is doing, it depends on the lumpiness of the world. The statistical regularities it's learning are the regularities characteristic of the densely populated lumps in the world. So, given some new text in, say, Japanese, it’s able to produce a corresponding French text that’s more or less in the appropriate region of the world of French language texts. It’s going to be deficient in some ways, which may or may not matter. You probably wouldn’t want to translate legal documents with such a system.

Still, what these systems can do seems remarkable to me and worth thinking about. Do I think they’re ‘intelligent’ in any interesting sense? No. Do I think we know what intelligence is? No. Do I think they’re doing ‘real translating’? Of course not. Sure, we can say they’re fake translations, but does that labeling tell us anything about real translations? No. It’s just a line, and that line doesn’t erase the interesting performance of these systems. What about those classical MT systems, the ones based on laboriously hand-crafted rules? Real translations or fake translations?


Bill B

> Dear Bill,
> In the second of your two recent Blog posts you point us to --
> Borges redux: Computing Babel -- you say your thoughts are
> driven by a wondering like this.
> "How come, for example, we can create this computer system
>  that crunches through a two HUGE parallel piles of texts in
>  two languages and produce a system than can then make
>  passable translations from one of those languages to the
>  other WITHOUT, however, UNDERSTANDING any language
>  whatsoever?  Surely the fact that THAT – and similar things
>  -- is possible tells us something about something, but
>  what?" [1]
> Yes, this is something to wonder about.  And I agree with
> where your thinking on this takes you.  Here is how I would
> Bable on about your wondering.
> First, because this seems unclear from you what you write,
> these (so called) Deep Learning Translators (DLTs) work
> through two huge piles of paired texts: each text, in one
> language, is paired with a human translation into the other
> language.  DLTs need to be fed with human translations of
> texts.  They don't "learn" to become (artificial) translators
> if they are fed with their own translations, unlike DL systems
> which can "learn" to play games, such as Go for example, by
> playing against themselves.  (That's because in the case of
> these games, success and good play are externally definable:
> not so for translation.  Good translation is manifested by
> examples, not defined by rules.)
> If you feed a (suitable configured) DLT with enough examples
> of good enough (human made) text translations, then it is not
> a surprise that it will come to encode meaning-free
> text-to-text translation mappings that work well enough for
> the translated text to be a "passable translations."  If this
> DL approach works at all, it can hardly end up doing anything
> else.  The surprise would be if, after much good training, a
> DLT couldn't cough up at least passable translations.  There
> are regularities in text translations; probabilistic
> regularities.
> But, we should notice, as you do!  It takes a human to tell if
> the translation is passable or not, and it takes a skilled
> human translator to turn this passable translation in to a
> professionally acceptable one, which is often what is really
> needed, and certainly needed to produce more training fodder
> for the DLT.
> So, first humans have to do all the translating of texts to
> produce the huge piles of stuff we feed to our DLT, then it
> takes a human to say if the resulting translation is passable
> or not.  The DLT cannot do this.  It'll give you certain
> values to do with the statistical probabilities of its
> translation efforts, and perhaps even include other less
> probable possibilities, according to its calculations.  But it
> cannot tell if it's most probable version is a passable
> translation or not.  It probably will be passable, if the DLT
> has been trained with enough good human translation pairs.
> All this is to get to this question.  Why do we think these
> DLTs are doing any translating, when it requires humans to
> first do loads and loads and loads of good translations for it
> to have anything to "learn" from?  And why do we think these
> DLTs are doing any translation when they can't tell you if
> what they come out with is a passable translation or not?
> Only we humans can do this.  All the translating here is being
> done by humans, no?
> I'd suggest that a fairer way to talk about this is to say
> that these (so called) DLTs produce passable translation
> look-a-likes, not real translations.  How they do this is
> clever, certainly, ingenious, even.  But they do not do it by
> translating texts.  They do it by inducing probabilistic
> pattern mapping rules from huge numbers of human made text
> translation pairs.  It's a clever trick.  
> We should call this Artificial Cleverness (AC), not AI
> (Artificial Intelligence), I think.  Or, if you insist upon
> calling it AI, we should be clear what kind of 'artificial' is
> in action here: the artificial in 'artificial light,' or the
> artificial in 'artificial flower'?  Artificial light is real
> light made by artificial means.  Artificial flowers are not
> real flowers, just artificially formed things made to look
> like real flowers.  DLTs are, I'd say, examples of A(flower)I,
> not A(light)I. Or, better said, I think, examples of AC.
> I may be muddled in my thinking, and only able to write
> passable English, but I'd bet some that Google Translate won't
> make a more passable version in any language you ask it to
> translate this text into.  (At least, I hope not!)
> Best regards,
> Tim
> PS: Are we, I wonder, in a Trumpian world, allowed to call
>    what DLTs give us, "fake translations"?

Bill Benzon
bbenzon at mindspring.com


http://new-savanna.blogspot.com/  http://new-savanna.blogspot.com/
http://www.facebook.com/bill.benzon  http://www.facebook.com/bill.benzon
http://www.flickr.com/photos/stc4blues/  http://www.flickr.com/photos/stc4blues/
https://independent.academia.edu/BillBenzon <https://independent.academia.edu/BillBenzon>
http://www.bergenarches.com/#image1  http://www.bergenarches.com/#image1

More information about the Humanist mailing list