I don't see how machine learning can be of any help here. Were would the
training examples come from?
To train a machine learning algorithm, presumably a machine translation
system, you'd need examples of already translated khipu. It's clear from the
artcile that we don't have anything like that. We know how to interpret
numbers, we know where on a khipu they used to represent someone's clan, but
we still can't read the things. What "data" will the machine learning algorithms
analyse?
A lot of the work the article talks about seems to involve manually recording features like the materials and colors used, the position and type of knots and so on, followed by searching for similar patterns in other khipus. Machine learning could conceivably speed up that encoding work.
Given a large enough collection of transcriptions, one could use "sophisticated" techniques like counting ngrams to try and identify common Quechua words. Depending on the content of a single khipu, i.e. whether it's more like a sentence or a longer text, and how similar modern Quechua languages are, it might even be possible to leverage that into a somewhat usable (i.e. sometimes correct) machine translation system.
Counting n-grams will tell you nothing useful, unless the language is Quecuha, and a dialect that is well known, at that. But if it is such a dialect, why are we not any closer to understanding it? You don't need machine learning for that, or even a computer; it helps, but the work can be done by hand (see The Golden Bug: https://en.wikipedia.org/wiki/The_Gold-Bug).
For khipu that are clearly not narrative, i.e. include mostly numerical, plus some other, possibly literary information, it would be even harder to pull off. You'd have to understand words out of context, including names etc.
At best you might end up modelling the structure of an unintelligible language- like I think was done for the Voynich manuscript. You'd still not be any closer to undestanding it.
I just happen to be reading Clive Cussler's Inca Gold (1994) where they use computers to decode a quipu cord from scratch in about an hour and a half. I'd love to see the spreadsheet used for the initial insight.
I actually love these kind of trashy archeology-themed books, where the brilliant linguist decyphers the ancient writing system over his/her lunchbreak, but it's rather spoiled by the inconvenient fact that the actual process - mountains of data, years of slog - is so unsexy by comparison. Plus, there is never a Rosetta Stone (that would make it too easy!), whereas almost every real-world decipherment has needed something like one. The only exception I can think of is Linear B; I've always thought it'd be a fascinating exercise to repeat Ventris' painstaking cataloguing and cross-referencing on a computer to see if his breakthrough could be repeated more quickly, or automatically.
Comments
This data is begging to be analyzed by machine learning algorithms.
I don't see how machine learning can be of any help here. Were would the training examples come from?
To train a machine learning algorithm, presumably a machine translation system, you'd need examples of already translated khipu. It's clear from the artcile that we don't have anything like that. We know how to interpret numbers, we know where on a khipu they used to represent someone's clan, but we still can't read the things. What "data" will the machine learning algorithms analyse?
A lot of the work the article talks about seems to involve manually recording features like the materials and colors used, the position and type of knots and so on, followed by searching for similar patterns in other khipus. Machine learning could conceivably speed up that encoding work.
Given a large enough collection of transcriptions, one could use "sophisticated" techniques like counting ngrams to try and identify common Quechua words. Depending on the content of a single khipu, i.e. whether it's more like a sentence or a longer text, and how similar modern Quechua languages are, it might even be possible to leverage that into a somewhat usable (i.e. sometimes correct) machine translation system.
So you'll model the features. Then what?
Counting n-grams will tell you nothing useful, unless the language is Quecuha, and a dialect that is well known, at that. But if it is such a dialect, why are we not any closer to understanding it? You don't need machine learning for that, or even a computer; it helps, but the work can be done by hand (see The Golden Bug: https://en.wikipedia.org/wiki/The_Gold-Bug).
For khipu that are clearly not narrative, i.e. include mostly numerical, plus some other, possibly literary information, it would be even harder to pull off. You'd have to understand words out of context, including names etc.
At best you might end up modelling the structure of an unintelligible language- like I think was done for the Voynich manuscript. You'd still not be any closer to undestanding it.
I just happen to be reading Clive Cussler's Inca Gold (1994) where they use computers to decode a quipu cord from scratch in about an hour and a half. I'd love to see the spreadsheet used for the initial insight.
I actually love these kind of trashy archeology-themed books, where the brilliant linguist decyphers the ancient writing system over his/her lunchbreak, but it's rather spoiled by the inconvenient fact that the actual process - mountains of data, years of slog - is so unsexy by comparison. Plus, there is never a Rosetta Stone (that would make it too easy!), whereas almost every real-world decipherment has needed something like one. The only exception I can think of is Linear B; I've always thought it'd be a fascinating exercise to repeat Ventris' painstaking cataloguing and cross-referencing on a computer to see if his breakthrough could be repeated more quickly, or automatically.
This is truly brilliant work. There must be other clever potential rosetta stone potential as the smart people in the article figured out.