5 ms·

> “correct translation” is a matter of taste So, bias no matter what then?

hprotagonist 6y ago
Bias no matter what. It’s unavoidable; languages are ambiguous and you are always, always going to leak out of band information to your audience. (by word choice, accent, font, punctuation, ... )
- csmpltn 6y ago
  Makes sense to me. I wonder what's the point behind this entire Reddit thread then. More fuel for the fire?
  - hprotagonist 6y ago
    Well, just because there's always _a_ bias doesn't mean you can't influence which one your machine translation system favors. I'm reminded of one of the early "oh, wow" moments with word2vec (a word embedding scheme), where the statement "King - Man + Woman" evaluates to "Queen". This is not a measure that the system understands anything, just that there are really strong statistical correlations that are preserved in the embedding that let you see how close vectors are to one another. Necessarily, this sort of vector similarity or word choice probability is a function of (at least) your training data. If you want to stop over-weighting pronoun choice for adjectives like "pretty" and "weak" or "emotional" to be feminine, be more choosy about your training set and don't just snarf up reddit willy-nilly and go brrrrr.
    - csmpltn 6y ago
      > be more choosy about your training set and don't just snarf up reddit willy-nilly and go brrrrr That's a much more interesting (and complex) conversation. Siphoning raw Reddit/Wikipedia is the closest you'll get to an honest snapshot of our internet-based society (for all of its tremendous flaws). One could make the case that we don't need an "honest snapshot of our internet-based society" to build a Translation app. But that's debatable... Being "more choosy" will inherently introduce a different set of biases. Perhaps those biases align more closely with our core beliefs and values - but they serve as bias nontheless.
      - hprotagonist 6y ago
        No disagreement here! One thing that will absolutely raise some people's hackles is pointing out that "intentionally modulating biases in translation techniques" already has a name, and it is "being prescriptive about language usage". I've found that generally, tech/progressive/machine learning ethics folks tend to believe that prescriptivism is the domain of, for example, stuffy old people like Strunk and White, how dare they tell me what is "proper" -- and then turn right around and do just the same with a straight face but don't seem to really realize that that's what they're doing. At any rate, designers of such systems do need to face head on the fact that machine translation systems at scale both 1. need to reflect usage in the world in a descriptive enough way as to be useful and honest and provide coherent approximations of translation, even sometimes if it's reflecting perhaps unpleasant truths about ourselves we might not like to think about and 2. need to realize that their participation and choices of embeddings, etc., contribute to the corpus of texts in the world as well as reflect them, so their prescriptive choices are also real and will have feedback effects.

Bias no matter what. It’s unavoidable; languages are ambiguous and you are always, always going to leak out of band information to your audience. (by word choice, accent, font, punctuation, ... )

csmpltn 6y ago
Makes sense to me. I wonder what's the point behind this entire Reddit thread then. More fuel for the fire?
- hprotagonist 6y ago
  Well, just because there's always _a_ bias doesn't mean you can't influence which one your machine translation system favors. I'm reminded of one of the early "oh, wow" moments with word2vec (a word embedding scheme), where the statement "King - Man + Woman" evaluates to "Queen". This is not a measure that the system understands anything, just that there are really strong statistical correlations that are preserved in the embedding that let you see how close vectors are to one another. Necessarily, this sort of vector similarity or word choice probability is a function of (at least) your training data. If you want to stop over-weighting pronoun choice for adjectives like "pretty" and "weak" or "emotional" to be feminine, be more choosy about your training set and don't just snarf up reddit willy-nilly and go brrrrr.
  - csmpltn 6y ago
    > be more choosy about your training set and don't just snarf up reddit willy-nilly and go brrrrr That's a much more interesting (and complex) conversation. Siphoning raw Reddit/Wikipedia is the closest you'll get to an honest snapshot of our internet-based society (for all of its tremendous flaws). One could make the case that we don't need an "honest snapshot of our internet-based society" to build a Translation app. But that's debatable... Being "more choosy" will inherently introduce a different set of biases. Perhaps those biases align more closely with our core beliefs and values - but they serve as bias nontheless.
    - hprotagonist 6y ago
      No disagreement here! One thing that will absolutely raise some people's hackles is pointing out that "intentionally modulating biases in translation techniques" already has a name, and it is "being prescriptive about language usage". I've found that generally, tech/progressive/machine learning ethics folks tend to believe that prescriptivism is the domain of, for example, stuffy old people like Strunk and White, how dare they tell me what is "proper" -- and then turn right around and do just the same with a straight face but don't seem to really realize that that's what they're doing. At any rate, designers of such systems do need to face head on the fact that machine translation systems at scale both 1. need to reflect usage in the world in a descriptive enough way as to be useful and honest and provide coherent approximations of translation, even sometimes if it's reflecting perhaps unpleasant truths about ourselves we might not like to think about and 2. need to realize that their participation and choices of embeddings, etc., contribute to the corpus of texts in the world as well as reflect them, so their prescriptive choices are also real and will have feedback effects.