Expansive translations: considerations and possibilities

Interesting post! This comment is going to be more some random thoughts I had while reading than a proper response.

First, translations of Shakespeare is a super interesting arena. No Fear Shakespeare 'translates' the original plays into modern English, which I admit is a helpful idea, but there's a problem with these beyond just the feeling of being juvenile: the 'translations' are often wrong, sometimes blatantly so. One such line I remember well is from Hamlet. The original is:

Horatio, or I do forget myself!

which became

Nice to see you again, Horatio—that is your name, right?

The No Fear Shakespeare version has clearly translated this as if it were, "Horatio - or do I forget myself?" (all punctuation in Shakespeare is arbitrary, so the important thing here is "I do" vs "do I"). The NFS version makes absolutely no sense in the context of the play, because Horatio is Hamlet's best friend! At the time of this meeting they haven't seen each other in a few months, but the reason Horatio is in Denmark at all is to be a companion to Hamlet, because they're BFFs. Hamlet definitely fucking knows his name. My interpretation of the original line is something like "Welcome Horatio, who I would as soon forget as my own self," i.e. basically the exact opposite of what NFS said.

---

Why are modern translations so narrow? What level of nuance would you like them to capture? A lot of the beauty - and a lot of the meaning - in Shakespeare is in the specific use of language (rhythm, imagery, sound, antithesis, repetition, et cetera). Arguably you just can't get every iota of the meaning out of Shakespeare without the exact original words. So it seems like it necessarily has to be a spectrum. (Incidentally this reminds me of the question of whole brain emulation: at what level of resolution does the emulation have sufficient fidelity to 'be' you, if you can't replicate the entire brain quark for quark?)

In translations of poetry - something I have amateur experience with - you have a lot of decisions to make. Do you try to preserve both the meter and the rhyme scheme? That already severely restricts your choice of words, making it harder to get across the explicit meaning. And then what about alliteration, assonance, in-rhymes? What about double (or triple) meanings? (when these latter take a backseat the omitted meaning is often added as a footnote or annotation). What if there's cultural information that doesn't carry over well to your intended audience? And if you've taken all the foregoing considerations into account, what's the likelihood that it's even possible for you to end up with something that feels true to the original (in the sense that poetry evokes certain images and emotions)?

There's a great scene in Henry IV Part 2 that's almost entirely sexual puns*. If you had to write a Spanish version of this scene, would it be better to preserve the literal (primary) meaning of each line? Or should you write a new scene that has the same basic plot, but where you're at liberty to change the literal meanings of the lines so you can fit in more sexual puns that actually make sense in Spanish? From what I've seen most translators choose the former (though perhaps this is specific to Shakespeare due to the crazy reverence in which people hold his works). I think this is a mistake artistically, but also there's clearly a trade-off here. Without incredibly intensive labor, you can't have it both ways.

I don't know much about ML so perhaps it would be a lot easier for it to solve these problems than it is for a human. But given the current strategy of training on corpora, and the fact that I doubt there are more than a handful of good examples of how to balance all of these considerations, I'm not sure how an ML system would learn to do it. But again, I'm not a computers person.

---

Although I was talking about poetry, I think a lot of the same broad considerations apply to the idea of personalized translations (e.g. of Wikipedia). I can see an ML system learning my idiolect just by listening to me 24/7, but learning where my inferential distances are (for lack of better phrasing) seems quite a lot harder? Though... perhaps I am just underestimating GPT-N?

---

*The link doesn't do it justice at all, but just for example look at the words 'stab', 'enter', 'lusty', and 'fist', and the delightful line 'do me, do me, do me your offices'.

As the post notes, inferential distance relates to differing worldviews and life experiences. This was written to an audience that mostly understands what inferential distance has to do with different worldviews - how would you explain it to a different audience?

Well, a typical translation doesn't try to bridge the gap between languages, it just picks something on the far side of the gap that seems similar to the one on the near side. But that leaves something out.

An example of this is in translations of Harry Potter, where Dumbledore's password is translated into a local sweet. The UK versions has "Sherbet Lemon" while the US version has "Lemon drop." Are these the same? I assumed so, but actually it seems the UK version has a "fizzy sweet powder" on the inside. In Danish and Swedish, it's (mis?) translated as lemon ice cream - which isn't the same at all. And in Hebrew, it's translated as Krembo, which doesn't even get close to translating the meaning correctly - it's supposed to be an "equivalent children’s dessert" - but the translation simply doesn't work, because you can't carry a Krembo around in your pocket, since it would melt. Does this matter? Well, the difference between a kindly old wizard who carries around a sucking candy, and one who carries around a kind-of-big marshmallow dessert. But that's beside the point - you don't translate the life experience that growing up eating sherbert lemons gives you, you find an analogue. 

The only way to translate a specific word or term accurately could be to provide so much background that the original point is buried, and the only way to translate an idea is to find an analogue that the reader already understands. And that's why translation is impossible - but we do it anyways, and just accept that the results are fuzzy equivalents, and accept that worldviews are different enough that bridging the gap is impossible.

For me the idea of expansive translations is fantastic. Every time I read a new post in Lesswrong that brings important information to the table, I think about translating it into Portuguese and bringing the information to the members of my tribe. But obviously I don't think about translating literally, word for word, because I can see the loss of information that this would bring. I know exactly how I could write in Portuguese that would bring the sensations desired by the original author of the post, considering all the cultural nuances and inferential distances. When you really know more than one language you can see why and when it is a bad idea to translate literally.

So how could we improve an expansive translation system? Suppose I took this post from Lesswrong and translated it into Portuguese. Then I would post the translation of the post in a software or expansive translation platform for arbitrary sites. Our new expansive-translations dot com, ou our new chrome extension.

Translators in the platform could give a score (from 0 to 10) of how good that translation looked for different translation formats: translation for children, translation for people with little or no math background, literal translations, focused translations for people with visual, auditory weaknesses, etc. Also people who would come into contact with those translations could give a grade of how easy it was to understand the subject matter.

Thus, we could create a market for expansive translations focused on people of different styles. For example, the system could consider that translations by people with similar mathematical/computational background to mine would probably please me more than expansive translations focused on a lay audience. Obviously this would depend on the type of subject matter, because I am a complete layman when it comes to various subjects, but in general the similarities of my profile with the translator's profile could be a proxy for me to find good expansive translations. Also, the score I assign for each expansive translation can be used to understand what kind of expansive translation fits me more.

It would be interesting if I could even select an expansive translation of each category. Today I want to explain what bitcoin is to my grandmother, what would be the best way to do that? Surely expert translators for this kind of audience would know how to do it much better than me. I would select a specific category and see several expansive translations sorted by relevance (a metric that considered inferential distances, similar characteristics between the one who wrote and the one who reads, etc).

Each person reading an expansive translation could also assign a score to the post. I can imagine the many problems that such platforms could introduce, but having a diversity of expansive translations would help a lot and I would certainly use it often. For example, a market I would certainly pay to be part of is one of expansive translations of scientific articles. By hovering the mouse over a paragraph of an article a pop-up could appear indicating that there were 8 translators with 8 different expansive translations for the same paragraph. I could click on a (+) and then select the expansive translations I would like to read.

Certainly each translator can elaborate the ideas of that paragraph in different styles, considering differential inferential distances from the reader, etc. Suppose I read three expansive translations among the eight. I could select which one pleased me the most. Then we would use machine learning to train a system that could predict what kind of expansive translation I would identify myself with the most in a set of expansive translations.

Maybe we could still do optional microtransactions for those good expansive translators. E.g., I select the best expansive translation and pay a few cents or microcents, as simple as a like button in the corner of each expansive translation. This way we could ensure benefits and incentives for expansive translators to produce the best translations as they could be rewarded in status and financially for anyone.

I can see a lot of ways in which we could monetize this system, so we could get more money to put on research and improve the system even more. Rewarding directly good translators is an idea to ensure that we don't lose the best candidates. I will stop my babble here, but there are lot more I can talk about this topic. Very interesting this topic, ozziegooen. Also, I believe I could program this system myself. But let me know what you think.

Maybe I am misreading, but in case not everone is not aware of them, in many ways concept-wise the "expansive translation" sounds quite similar to critical editions / translations [1] and other similar scholarly annotated editions [2]. This kind of work usually includes more or less extensive commentaries (often presented as footnotes or endnotes) by later scholars that attempt to explain meaning and context of the original text that may be lost on their contemporary audience. (Critical editions also attempt to deal with cases where there are several differing or fragmented versions of the source text or the original language is very archaic or dead which ofc leads to issues with faitful translations.)

Sometimes the commentaries become influential on their own right. (Weirder things have also happened, like when according to one textual history theory, one day in China somebody shoehorned in a historical novel to terse ancient administrative logbook as "a commentary" to make it more interesting as the original chronicle was attributed to most famous of all scholars, thus creating a renowned classic [3].)

The point being, in scholar traditions it is traditional to distinguish between the text / translation and the commentaries, which at the same time deals with impossibility of perfect translations (by acknowledging that explanations of translation choices at minimum and in general, also other annotations are warranted) but also tries to retain a text with considerable fidelity to its source (such as, has no large insertions to improve it for contemporary audience) which is also useful.

[1] https://en.m.wikipedia.org/wiki/Textual_criticism

[2] https://en.m.wikipedia.org/wiki/Annotated_edition

[3] https://en.m.wikipedia.org/wiki/Zuo_zhuan

Imagine a system where when you land on a Wikipedia page, it translates it into a version optimized for you at that time. The examples change to things in your life, and any concepts difficult for you get explained in detail. It would be like a highly cognitively empathetic personal teacher.

Hmm, something about this bothers me, but I'm not entirely sure what. At first I thought it was something about filter bubbles, but of course that can be fixed; just tune the algorithm so that it frames things in a way that is just optimally outside your intellectual filter bubble/comfort zone.

Now I think it's something more like: it can be valuable to have read the same thing as other people; if everyone gets their own personalized version of Shakespeare, then people lose some of the connection they could have had with others over reading Shakespeare, since they didn't really read the same thing. And also, it can be valuable for different people to read the same thing for another reason: different people may interpret a text in different ways, which can generate new insights. If everyone gets their own personalized version, we lose out on some of the insights people might have had by bouncing their minds off of the original text.

I guess this isn't really a knockdown argument against making this sort of "personal translator" technology, since there's no reason people couldn't turn it off sometimes and read the originals, but nevertheless, we don't have a great track record of using technology like this wisely and not overusing it (I'm thinking of social media here).

The fact that people have different understanding of the same texts and have to "translate" them through an inferential distance is a necessary evil. Just because something is a necessary evil doesn't mean it's good, and certainly doesn't mean that we should be fine with deliberately creating more of it.