Looks like promising research. I will have to read the actual paper later as opposed to just the blog post. One thing I would like to say is my qualm with the obsession with unsupervised learning as the quintessential technique. Unsupervised != no human input. Having the benefit of knowing about word embeddings is something inherently built into this system by the human designers and something that supervised learning does not have the benefit of when given original-translated pairs.
> One thing I would like to say is my qualm with the obsession with unsupervised learning as the quintessential technique. Unsupervised != no human input.
I'm not sure I follow the qualm you are trying to get across. Are you saying you disagree with the term 'unsupervised' because unsupervised algorithms still bake in human assumptions (like a human-designed word embedding model) so that's essentially still supervision?
The obsession with 'unsupervised' learning as the quintessential technique is about getting better results for less money/effort. The premise is that we assume deep models tend to scale up in accuracy as training data size increases, so we always want larger datasets to increase accuracy. But creating labeled data takes a linear amount of human effort ($$$) as the dataset size increases. At a certain point, creating more labeled data to improve a model is not cost effective or maybe even impossible.
Unlabeled data can be acquired nearly for free in nearly unlimited quantities in many cases. So if we can use unlabeled data instead (even if it requires complex pre-processing like CBOW embedding models which essentially turn bits of the unlabeled data into it's own label) your final results-per-dollar-invested goes through the roof over supervised learning. That's the obsession. It's not about literally no supervision being involved in the process. It's about driving down the cost of data acquisition while driving up the percent of the world's available data you can use for training a model.
I apologize in advance if I'm missing the point you are making.
No need to apologize, I like HN precisely because people point out the flaws/confusing parts of my comments/opinions. I agree with everything you say and am happy you said it because this is how unsupervised learning should be viewed. I.e., a better ROI in specific cases. However, I have seen too often the "cake of AI" where the batter is unsupervised, the icing is supervised, and the cherry on top is reinforcement learning. Somehow, this image connotes that unsupervised is at the core of AI and also the most important. For what is a cake consisting of only icing and a cherry?
Where I disagree with you is that the obsession is purely driven from "results-per-dollar-invested", at least in the academic world. That being said, unsupervised learning is a great tool and definitely worthy of research.
To summarize, my comment was completely tangential of this paper (the authors make no such claims). It was more of a stream of consciousness comment that arose because I envisioned someone reading the paper and saying "see!, unsupervised learning leads to real understanding, no humans needed!"
The previous paper they mention explains the core insight that makes unsupervised translation possible: https://arxiv.org/abs/1710.04087
The original paper didn't receive the attention I thought it would, but I continue to think this is a fascinating result which has deep implications for machine learning and for linguistics.
These word embeddings keep on yielding all kinds of amazing benefits. Is there any kind of explainability research to help people understand them better in terms of human psychology?
Word embeddings work because they reflect co-occurrences. I don't know whether that counts as an explanation in terms of psychology, but humans tend to put related things together. In a newspaper the articles aren't jumbled together, but there are sections on different topics, and in each section the articles are clearly delineated instead of mixing their sentences and each sentence represents a single unit instead of giving partial information on a dozen unrelated things.
It might seem obvious that things should be done that way, but if you consider servers hosting lots of different websites on the same physical machine, or data structures spread out over several memory allocations held together by pointers, it's clear that there are other possibilities. So it does seem to be specific to the way humans use language.
And because human language has this property of co-occurrences corresponding to relatedness in meaning, you can represent the meaning of a word by building a model that only predicts the probability that two words occur together.
In linguistics, there's a classic principle "You shall know a word by the company it keeps" (Firth, J. R. 1957) - colocations (a sequence of words or terms that co-occur more often than would be expected by chance) are very informative about what a word means.
Not so much in psych, where language remains a pretty fundamental mystery, but from the philosophy of language side, word embeddings are, I think it's fair to say, fundamentally an implementation of Sassurian structuralism and semiotics. Words (signs) have no intrinsic, native meaning. And so, the only way to figure out which one is which is by measuring it in relation to all other signs in the lexicon; that is to say, words only mean something because they don't mean all the other things, and that the Structure as a whole is what provides meaning.
There has been much in the way of discussion for, and concern about (for example, the Deconstructionist movement), these ideas, for the past 60 years or so. And a bit of practical exploration in the field of child development.
Interestingly, the fact that embeddings between languages seem to share some common shapes (per the linked paper in this thread), would seem to suggest that
A) Fundamentally, most languages have the same deep structure, whether through coincidence or common evolutionary root.
or
B) The brain has a hard-wired structure for language, evolved alongside the development of language itself. The Chomskian Language Acquisition Device. We're not born tabula rasa, we've got some hardcoding indicating how we're going to understand things
Or a little of A, a little of B maybe, as it does end up being a boostrapping problem.
Normally this technique wouldn't be useful, because it's overfitting a specific training set. (If you make space X as similar as possible to space Y, then this mapping from X to Y is only useful for X to Y – it can't generalize to other situations, which is often the goal of an ML model.)
But since the task is "Translate from English to Italian," and since all languages have similar embedding structures (Zipf mystery), overfitting is exactly what we want: we want, for any given English phrase, to find the closest-fitting mapping to a corresponding Italian phrase.
The more I learn about ML and data mining, the more I'm astounded by how clever many of the techniques are, and how much artistry is involved. You have to be clever to make a certain model perform well in a certain domain. If you want to make a stock trading bot, you can't randomly subdivide stock market data into e.g. 70% training data and 30% test data, because the data is ordered by time. You have to use the past 3 years of stock market data as a training set, and validate it against the subsequent 1 year of market data.
I really like ML because the techniques applicable for training a stock market bot seem unrelated to the algorithms for doing unsupervised machine translation, which differ from how to model credit fraud, which are no doubt different from how to build a dota 2 bot. :)
Splitting the time series data into training and test datasets by making a single cut is not some artistry or epiphany following years of research. It is common sense if you work in time series domain.
"We can improve upon this by making local edits using a language model that has been trained on lots of monolingual data to score sequences of words in such a way that fluent sentences score higher than ungrammatical or poorly constructed sentences."
Does anyone knows some references / corpus in English language for this?
I have a feeling the answer is “no”, but can anyone comment on whether or not this could be used to decode the utterances of other animals, such as whales?
Without commenting on whether this method could be applied to whales, the method described here does not require a bilingual dictionary. They learned their "dictionary" unsupervised by aligning monolingual word embeddings.
So, they built translation system using bilingual dictionaries, then asked it to translate from English to Urdu, and then back to English, and minimized loss between original and double-translated English.
In one of my previous companies we used this technique to hire pair of translators: we gave translator pairs such task, and hired pair which reconstructed original text more closely.
No, they didn't start with dictionaries, or any other parallel corpora; they learned the word by word translations as well from monolingual corpora, by finding alignments between monolingual word embeddings in the target languages.
Next step: unsupervised word segmentation. That way they could maybe apply this unsupervised translation system to undeciphered texts, e.g. Linear A, Rongorong, etc. I doubt it will work since most of the undeciphered scripts have a very small corpus, but maybe worth a try.
2. rotate embeddings space of two languages for optimal alignment, assuming frequency and neighborhoods of word embeddings are more-less the same in any language
3. iteratively minimize difference in bidirectional translations
All of this seems to be using word embeddings, but most languages don't have all that many words. You're effectively trying to train with just a few thousand data points, and will quickly overfit.
Wouldn't the method work better with n-gram embeddings, where n=3 or 4?
How does it learn idioms (sequences of words that make no sense when translated into another language word-for-word). For example "Stop beating around the bush!" would result in complete nonsense if translated into any language other than English.
24 comments
[ 2.0 ms ] story [ 58.3 ms ] threadI'm not sure I follow the qualm you are trying to get across. Are you saying you disagree with the term 'unsupervised' because unsupervised algorithms still bake in human assumptions (like a human-designed word embedding model) so that's essentially still supervision?
The obsession with 'unsupervised' learning as the quintessential technique is about getting better results for less money/effort. The premise is that we assume deep models tend to scale up in accuracy as training data size increases, so we always want larger datasets to increase accuracy. But creating labeled data takes a linear amount of human effort ($$$) as the dataset size increases. At a certain point, creating more labeled data to improve a model is not cost effective or maybe even impossible.
Unlabeled data can be acquired nearly for free in nearly unlimited quantities in many cases. So if we can use unlabeled data instead (even if it requires complex pre-processing like CBOW embedding models which essentially turn bits of the unlabeled data into it's own label) your final results-per-dollar-invested goes through the roof over supervised learning. That's the obsession. It's not about literally no supervision being involved in the process. It's about driving down the cost of data acquisition while driving up the percent of the world's available data you can use for training a model.
I apologize in advance if I'm missing the point you are making.
Where I disagree with you is that the obsession is purely driven from "results-per-dollar-invested", at least in the academic world. That being said, unsupervised learning is a great tool and definitely worthy of research.
To summarize, my comment was completely tangential of this paper (the authors make no such claims). It was more of a stream of consciousness comment that arose because I envisioned someone reading the paper and saying "see!, unsupervised learning leads to real understanding, no humans needed!"
The original paper didn't receive the attention I thought it would, but I continue to think this is a fascinating result which has deep implications for machine learning and for linguistics.
It might seem obvious that things should be done that way, but if you consider servers hosting lots of different websites on the same physical machine, or data structures spread out over several memory allocations held together by pointers, it's clear that there are other possibilities. So it does seem to be specific to the way humans use language.
And because human language has this property of co-occurrences corresponding to relatedness in meaning, you can represent the meaning of a word by building a model that only predicts the probability that two words occur together.
There has been much in the way of discussion for, and concern about (for example, the Deconstructionist movement), these ideas, for the past 60 years or so. And a bit of practical exploration in the field of child development.
Interestingly, the fact that embeddings between languages seem to share some common shapes (per the linked paper in this thread), would seem to suggest that A) Fundamentally, most languages have the same deep structure, whether through coincidence or common evolutionary root. or B) The brain has a hard-wired structure for language, evolved alongside the development of language itself. The Chomskian Language Acquisition Device. We're not born tabula rasa, we've got some hardcoding indicating how we're going to understand things
Or a little of A, a little of B maybe, as it does end up being a boostrapping problem.
Normally this technique wouldn't be useful, because it's overfitting a specific training set. (If you make space X as similar as possible to space Y, then this mapping from X to Y is only useful for X to Y – it can't generalize to other situations, which is often the goal of an ML model.)
But since the task is "Translate from English to Italian," and since all languages have similar embedding structures (Zipf mystery), overfitting is exactly what we want: we want, for any given English phrase, to find the closest-fitting mapping to a corresponding Italian phrase.
The more I learn about ML and data mining, the more I'm astounded by how clever many of the techniques are, and how much artistry is involved. You have to be clever to make a certain model perform well in a certain domain. If you want to make a stock trading bot, you can't randomly subdivide stock market data into e.g. 70% training data and 30% test data, because the data is ordered by time. You have to use the past 3 years of stock market data as a training set, and validate it against the subsequent 1 year of market data.
I really like ML because the techniques applicable for training a stock market bot seem unrelated to the algorithms for doing unsupervised machine translation, which differ from how to model credit fraud, which are no doubt different from how to build a dota 2 bot. :)
Does anyone knows some references / corpus in English language for this?
To translate whale song, you need data that corresponds to what the whales are singing about.
In one of my previous companies we used this technique to hire pair of translators: we gave translator pairs such task, and hired pair which reconstructed original text more closely.
2. rotate embeddings space of two languages for optimal alignment, assuming frequency and neighborhoods of word embeddings are more-less the same in any language
3. iteratively minimize difference in bidirectional translations
Wouldn't the method work better with n-gram embeddings, where n=3 or 4?
And since you are simply learning a rotation matrix, there is no risk of overfitting.