I'm glad to have kept reading to the author's conclusion:
> As a hybrid approach, you could produce a large number of inferred sentiments for words, and have a human annotator patiently look through them, making a list of exceptions whose sentiment should be set to 0. The downside of this is that it’s extra work; the upside is that you take the time to actually see what your data is doing. And that’s something that I think should happen more often in machine learning anyway.
Couldn't agree more. Annotating ML data for quality control seems essential both for making it work, and building human trust.
This approach only works if you use OP's assumption that a text's sentiment is the average of it's word's sentiment. That assumption is obviously flawed (e.g. "The movie was not boring at all" would have negative sentiment).
Making this assumption is fine in some cases (for example if you don't have training data for your domain), but if you build a classifier based on this assumption why don't you just use an off-the-shelf sentiment lexicon? Do you really need to assign a sentiment to every noun known to mankind? I doubt that this improves the classification results regardless of the bias problem.
Heck, it's so important that it needs people with detail-orientation and solid judgement, because crowdsourcing (ie populism) may not be the best source of Godwin's law ethical mooring.
> There is no trade-off. Note that the accuracy of sentiment prediction went up when we switched to ConceptNet Numberbatch. Some people expect that fighting algorithmic racism is going to come with some sort of trade-off. There’s no trade-off here. You can have data that’s better and less racist. You can have data that’s better because it’s less racist. There was never anything “accurate” about the overt racism that word2vec and GloVe learned.
The big conclusion here after all that code buildup does not logically follow. All it shows is that one new word embedding, trained by completely different people for different purposes with different methods on different data using much fancier semantic structures, outperforms (by a small and likely non-statistically-significant degree) an older word embedding (which is not even the best such word embedding from its batch, apparently, given the choice to not use 840B). It is entirely possible that the new word embedding, trained the same minus the anti-bias tweaks, would have had still superior results.
I also disagree with the conclusion, but for a different reason. I think it's unlikely that the word embeddings were just lower quality. That should result in noise, not bias.
I that there is a real statistical pattern in the training data that names associated with certain ethnicities are more likely to appear close to words with negative sentiment. I just don't think this necessarily means that the news is racist. I think more analysis is needed to see where this pattern comes from.
However, if it is true that the news is biased and racist in a quantifiable way, that would be a bigger problem than biased word vectors. I would genuinely be interested in seeing that type of analysis.
Note though that "the news is racist" is different from "the model we learned (from the news) is racist". Maybe the first can be false while the second is true.
I think you're reading this statement as more general than it's meant to be? I interpret it as meaning that there is not necessarily any tradeoff, as there wasn't in this case. "You can have data" -> there exists.
Is there anyone who thinks that the current level of racism is required for the current accuracy? I can't imagine people that racist to be common in the data community
> Is there anyone who thinks that the current level of racism is required for the current accuracy? I can't imagine people that racist to be common in the data community
It depends on two things. The first is how you're defining racism. If the algorithm is predicting that 10% of white people and 30% of black people will do X, because that is what actually happens, some people will still call that racism but there is no possible way to change it without reducing accuracy.
If the algorithm is predicting that 8% of white people and 35% of black people will do X even though the actual numbers are 10% and 30%, then the algorithm has a racial bias and it is possible to both reduce racism and increase accuracy. But it's also still possible to do the opposite.
One way to get the algorithm to predict closer to 10% and 30% is to get better data, e.g. take into account more factors that represent the actual cause of the disparity and just happen to correlate with race, so factoring them out reduces the bias and improves accuracy in general.
The other way is to anchor a pivot on race and push on it until you get the results you want, which will significantly harm accuracy in various subtle and not so subtle ways all over the spectrum because what you're really doing is fudging the numbers.
"If the algorithm is predicting that 10% of white people and 30% of black people will do X, because that is what actually happens, some people will still call that racism but there is no possible way to change it without reducing accuracy."
What is actually happening? Does it tell you if they are they doing X precisely because they are black or white? The racist part might not be the numbers per se, but in the conclusion that the color of their skin has anything to do with their respective choices.
ML is spitting out correlations, not an explicit causal model. If, in reality, X is only indirectly and accidentally correlated with race, but I look at the ML result and conclude the skin color has something to do with X, then the only racist element in the whole system is me.
> Some people expect that fighting algorithmic racism is going to come with some sort of trade-off.
Um, that's because we know it comes with trade-offs once you have the most optimal algorithm. See for instance https://arxiv.org/pdf/1610.02413.pdf. If your best performing algorithm is "racist" (for some definition of racist") you are mathematically forced to make tradeoffs if you want to eliminate that "racism".
Of course, defining "racism" itself gets extremely tricky because many definitions of racism are mutually contradictory (https://arxiv.org/pdf/1609.05807.pdf).
Not necessarily. In the case of word vectors we are using unsupervised learning to identify patterns in a large corpus of data to improve the learning. This is a completely different issue than your credit score example, which is supervised learning.
Not all patterns are equally useful. By removing those unuseful patterns we might make less mistakes (for example giving negative sentiment to a Mexican restaurant review) and free up capacity in the word vectors to store more useful patterns. I would expect baking other real-world assumptions into your word vectors unrelated to bias could also be helpful.
1. Racism makes the algorithm good so we should make the algorithm less racist (at a cost to its performance) or decide we want to allow systematic racism.
2. The metric for how good the algorithm is (ie training data) encourages it to be racist and therefore correcting the bias in the algorithm may decrease its performance on the training data but may not affect its performance in the real world, or decrease its performance in the “performance + meets legal requirements” metric.
It seems to me that if you wanted to root out sentiment bias in this type of algorithm, then you would need to adjust your baseline word embeddings dataset until you have sentiment scores for the words "Italian", "British", "Chinese", "Mexican", "African", etc that are roughly equal, without changing the sentiment scores for all other words. That being said, I have no idea how you'd approach such a task...
I don't think you could ever get equal sentiment scores for "black" and "white" without biasing the dataset in such a manner that it would be rendered invalid for other scenarios (e.g., giving a "dark black alley" a higher sentiment than it would otherwise have). "Black" and "white" is a more difficult situation because the words have different meanings outside of race/ethnicity.
I think I would agree. You otherwise run the risk of having fixed the metric ("Italian" vs. "Mexican", "Chad" vs. "Shaniqua", etc.) without actually fixing the underlying issue.
Also, regarding black/white etc., there might legitimately be words which have so many different meanings (whether race-related or not) that you should just exclude them from sentiment analysis. "Right" can mean like "human rights", "right thing to do", or "not left". Probably plenty of other words like that. You might do better to have a list of 100-200 words that are just excluded because of issues like that.
> there might legitimately be words which have so many different meanings
I haven't studied word embeddings past the pop-sci level but wouldn't such words form multiple clusters in the embedding space? I would have thought it would be relatively easy to get different 'words' for 'right (entitlement)', 'right (direction)', etc?
Would it be worth trying to think of words with different meanings as entirely new words? So, "white" in one sentence may be a different word than "white" in another?
There's a long list of papers on that - 'multi-sense word embeddings'. But more recently we have found that passing the raw character embeddings through a two layer BiLSTM will resolve the ambiguity of meaning from context - 'ElMO'.
Does “a dark black alley” have a sentiment at all?
I would argue that it’s pragmatically associated with bad things (e.g., being mugged, overcrowded areas) but it’s not intrinsically bad (or good) itself.
My intuition is that word-level sentiment is rather pointless. “The Disaster Artist was not bad” has a positive sentiment overall, but each of the individual words, except possibly ‘artist’, have are usually thought to be negative. Moreover, you can totally flip the overall sentiment by adding another neutralish word “The Disaster Artist was not even bad.”
Similarly, my guess is that alley is rarely found in a positive context, but the actual sentiment comes from elsewhere in the utterance.
Word-level sentiment is like spherical cows in a vacuum in physics. Everyone knows its an extremely flawed model, but it produces good results in a lot of scenarios, so it will inevitably be used because it also has the enormous benefit of simplicity.
It certainly gets a sentiment score, but whether that score is in any way meaningful or corresponds to actual human sentiment is important. Otherwise, you’re just playing stupid games, and winning stupid prizes...though I suppose just stupid is a step up from stupid and racist.
Yes, and it's pretty scary how many technologists seem to be surprised by this. If we train bots using data derived from humans, the expectation is that they will inherit biases from humans. There's nothing about a silicon brain that automatically bestows perfect objectivity, only perfect obedience.
I struggle to think of a single person with the faintest understanding of what machine learning algorithms are being surprised by this. Who are these "technologists" you're speaking of?
Almost everyone I know with at least a faint understanding of ML is surprised by models picking up racism etc when there was zero intent to do so, because of systemic racism etc in available data. Or at least surprised by how much can be picked up. You're bubbled if no one you know is surprised.
Sometimes data might be 'racist' (i.e. human written corpus text)... but sometimes data is just data.
Are facts racist?
I would seem the world is rather diverse, i.e. 'people are different' and as we are different, AI is going to pick up on that. That's the whole point.
Now ... some bad examples like in this example taking positive/negative inferences the wrong way. OR actual systematic racisms showing up in bad ways i.e. maybe some groups are more likely to be monitored than others, thereby showing up more frequently in mad terms etc..
Why is this surprising? ML models are just recognizers and bias on the basis of ancestry is observable in all human cultures at all times.
If we nobly insist that the models describe the world as we wish it were and ought to be, then we won't be describing the data accurately. Maybe that trade-off is worthwhile if it somehow reforms human attitudes along lines we find more agreeable?
Conversely, almost everyone I know with at least a faint understanding of ML is entirely unsurprised about this.
Then again, my personal social bubble leans heavily liberal and hard left. And I think that has a lot more to do with it than with how much people understand ML. When you explain this sort of thing to people who have no idea about ML, in very simple terms ("we give the robot the text that humans wrote, so that it can pick up the patterns" etc), they see why it does that very quickly, as well - if their politics makes them aware of bias in general.
Hmmm...I'm no expert, but my master's thesis topic in the 90's was on neural networks that use R-squared (a measure of correlation), and when I saw the news about Microsoft's chatbot going Nazi, I was not at all surprised. Not saying no one you knew was surprised, but I had "at least a faint understanding of ML", and the primary thing I learned about it was that it learns what's in the data, whether that's the part of the data that you intended it to learn or not.
Tay was trolled hard by 4chan, that's why she went hardcore Nazi almost immediately. It was amusing, but not a fair & controlled experiment by any means.
It’s because most tech people have the default position that racism is not really a big deal, an edge case in modern society. That certainly is the message the political center and right is pushing.
Given that the data showed a massive range by name within the same race and a much smaller skew between different races, couldn't this data be said to support that conclusion?
Disclaimer: I don't know enough about the data or the algorithm to determine this mathematically but I think worth pointing out. Would have been nice to see some statistical analysis instead of just assuming the charts speak for themselves.
This thread reminds me of the nature of political discourse at the moment, for example with regards to political correctness. The loudest and most popular voices are pushing simple fixes to intractable problems, and more sensible voices mentioning the truth of the matter are buried.
Indeed. Before there were computers, in fact. In his 1864 autobiography, Charles Babbage wrote:
"On two occasions I have been asked, — 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' In one case a member of the Upper, and in the other a member of the Lower, House put this question. I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question."
The data is correct and a-biased. If you ask 100 people around you, they are, on average, more likely to have had a negative burrito experience, than a negative pasta experience.
The learning algorithms are crude and dumb. They will simply fit to any data you provide it (you choose how many Mexican restaurant food reviews you train your sentiment classifier on). Then they count how many times the words "mexican" and "man" and "mexican man" appear with a positive or negative label in the train set. And objectively try to give the best probability for that.
Current sentiment analyzers are not AI: no common sense, no understanding, no reasoning. We are just rushing to replace looking a job candidate in the eyes with running some 1960's logistic regression over their cover letter. Let's hope for their sake they did not manage a Mexican restaurant.
On May 3rd, 1997 La Costeña of Mountain View, California created the world's largest burrito. The burrito weighed in at 4,456.3 pounds and was measured at 3,578 feet long. It was created at Rengstorff Park in Mountain View.
Well, not so bad if we use the mirror to reflect on ourselves and our biases, and work to negate them. Fairly bad if they're used for recommendations and for rankings.
That was my thought too. You can't manage what you can't measure. This is a tool for measuring the amount of racism in our society. It's a good thing, not a bad thing :)
I think that the bias problem they are highlighting is very important. That said, I'm wondering if they really didn't try (like the title suggests) or if they choose this approach on purpose because it highlights the problem.
To explain what happened here: They trained a classifier to predict word sentiment based on a sentiment lexicon. The lexicon would mostly contain words such as adjectives (like awesome, great, ...). They use this to generalize to all words using word vectors.
The way word vectors work is that words that frequently occur together are going to
be closer in vector space. So what they have essentially shown is that in common crawl and google news names of people with certain ethnicities are more likely to occur near words with negative sentiment.
However, the sentiment analysis approach they are using amplifies the problem in the worst possible way. They are asking their machine learning model to generalize from training data with emotional words to people's names.
I think the point is that they did what's commonly done in real world machine learning. It's no surprise that it's flawed - but that flawed stuff is actually being used all over the place.
They could have tried to have a dataset of bias triples (A in relation to C is like B in relation to C), and minimise the score on that by adding it to the loss function, so the model trains with minimal bias.
It would be interesting to use the Uber/Lyft dataset of driver and passenger ratings to do an analysis like this.
For any such analysis there are a great many confounds, both blatant and subtle. Finding racism everywhere could be because overt racism is everywhere, or it could be confirmation bias. It could even be both! That's the tricky thing about confirmation bias—one never knows when one is experiencing it, at least not at the time.
I've heard a lot about racism in AI, but looking at the distributions of sentiment score by name, a member of any race would rationally be more worried about simply having the wrong name. Has there been any work done on that?
I mean name within the same race. On the range of racial averages was 3, but the range of names within a race was around 10. I don't know how significant the results for individual names are, but I was very surprised by that result.
> Reminds me of how Google Photos couldn't differentiate between a black person & a monkey, so they've excluded that term from search altogether.
Technically that is what happened, but it paints an incorrect picture in people's minds. Out of the billions of images that Google Photos had auto-tagged, it tagged one picture of two black people as "gorillas".[1] This was probably the first time this had ever happened. (If it had happened before, it surely would have been spread far and wide by social media & the press.)
So Google's classifier was inaccurate 0.0000001% of the time, but the PR was so bad that Google "fixed" the issue by blacklisting certain tags (monkey, gorilla, etc). If you take photos of monkeys, you'll have to tag them yourself.
I'm sure Google could do better, but the standard required to avoid a PR disaster is impossible to meet. If the classifier isn't perfect forever, they're guaranteed to draw outrage.
Our expectation of our algorithms are based on human performance. A human would never tag a black person as a gorilla, or vice versa, and if someone did it even once we could pretty safely conclude they're either extraordinarily incompetent, or racist, and in either case we wouldn't trust any tagging done by such a human.
> This was probably the first time this had ever happened. (If it had happened before, it surely would have been spread far and wide by social media & the press.)
That is a very big leap. Social media might be widespread, but almost everything in the world goes unremarked upon. Think of all the news stories that turn up an old tweet or Facebook post that, if anyone had paid attention at the time, would have stopped events from progressing.
One of the things that amuses me is trying to find racist/sexist google search results. Here's a few:
I remember a while back Google got flack because the image search for "scientist" was almost entirely famous African American scientists. That's now changed and shows stock images of (mostly white) people in lab coats.
"Three black teenagers" shows mostly groups of mugshots.
The word "Brazilian" shows hot, almost nude women. "German" shows the flag. "Portuguese" shows maps, flags, and a lot of normal looking people. "Hispanic" all pictures are normal looking people.
Seeing images that would be 'racist' or 'sexist' is reflective of you, not the results. For instance if you search for 'white man and white woman' you'll find almost exclusively pictures of interracial couples. Is it some conspiracy to push interracial relations onto people? People of a different bias would say so, and it's equally ridiculous. In reality the simple matter is that Google's search is still extremely primitive and the results are mediocre at best. So you can easily break the search when searching for anything that cannot be trivially mapped to a direct text mapping such as e.g. Justin Bieber or Abraham Lincoln.
For instance search for 'green circle' - okay you get mostly green circles. Now search for 'green circle with red line' and the results are completely nonsensical. The huge leap forward in search engines was being able to avoid returning hardcore porn when searching for Abraham Lincoln. But in spite of tens of thousands of engineers, hundreds of billions of dollars in revenue, and all sorts of fancy declarations of ultra sophisticated AI solving every problem under the sun, we really haven't moved that far beyond that early milestone.
Yeah, I didn't mean to suggest I think the AI/search results are actually racist/sexist. If I really believed that, I wouldn't find it amusing. As you suggest, it's an amusing anecdote which shows how much farther we have to go with regards to getting ML/AI/search right.
'Brazilian' has other meanings you may be unaware of... namely it being the name for a bikini wax. Almost nude women is beyond expected in this case. Just another example of how complex these things are linguistic and cuturally.
Well, to be fair they excluded high tens / low hundreds of potentially offensive terms from search before even launching and when this came out they just extended the list a little.
Sometimes having product vision requires recognizing that the products you build come with limitations and potential for very real emotional reactions of very real human being users.
I believe it was gorilla, not monkey and I understand Google for not wanting its product to randomly call people animal names, especially when they are part of a group where it is far too common.
To me the article brought great insight - I realized that humans do not just pattern match. They also seek understanding, which I would define as an ability to give a representative example.
It is possible to give somebody a set described by arbitrarily complex conditions while the set itself is empty. Take any satisfiability problem (SAT) with no solution - this is a set of conditions on variables, yet there is no global solution to these.
So if you were a Chinese room and I would train you on SAT problems, by pure pattern matching, you would be willing to give solutions to unsolvable instances. It is only when you actually understand the meaning behind conditions you can recognize that these arbitrary complex inputs are in fact just empty sets.
So perhaps that's the flaw with our algorithms. There is no notion of I understand the input. Perhaps it is understandable, because understanding (per above) might as well be NP-hard.
Humans can do more than pattern-match. But they often just pattern-match anyway, because it's far easier and quicker, and doing more than that for all the brief day-to-day interactions is virtually impossible.
So at some point you need to decide when you pattern-match and accept the result for granted, and when you decide to dig into it further to understand why the pattern matched the way it did, and whether it's relevant. But that is itself a choice, and it's also going to be biased (for example, towards people you personally know, and against random strangers).
That is not my argument at all. What I argue is that brains attempt to resolve the problem, while computers (when they pattern match in typical ML algorithm) do not.
It is possible that brain has specialized circuits to solve small instances of SAT, and it just gives up on large enough instance. I am sure you know the feeling that you get when you understand something - it's very much like the pieces of the puzzle that suddenly perfectly fit to each other.
> Note that the accuracy of sentiment prediction went up when we switched to ConceptNet Numberbatch.
> Some people expect that fighting algorithmic racism is going to come with some sort of trade-off. There’s no trade-off here. You can have data that’s better and less racist. You can have data that’s better because it’s less racist. There was never anything “accurate” about the overt racism that word2vec and GloVe learned.
I wonder if this could be extended to individual names that have strong connotations with people because of the fame of some particular person, like "Barack", "Hillary", "Donald", "Vladimir", or "Adolf", or if removing that sort of bias is just too much to expect from a sentiment analysis algorithm.
Where I grew up, there is a majority group with fair skin, later(possibly incorrectly) attributed to the fact that they worked in the fields less. The minority group is darker skinned. If you train any reasonable machine learning model on any financial data, it will pick up on the discrepancy. If it did not I would say it is a flawed model. But that is more a sign that people should avoid such models.
How to make a program that does what you asked it to do, and then add arbitrary fudge factors as the notion strikes you to "correct" for the bogeyman of bias.
Suppose sentiment for the name Tyrel was better than for Adolf. Would that indicate anti-white bias? Suppose the name Osama has really poor sentiment. What fudge factor do you add there to correct for possible anti-Muslim bias? Suppose Little Richard and Elton John don't have equal sentiment. Is the lower one because Little Richard is black, or because Elton John is gay?
What we have been seeing lately is an effort to replace unmeasurable bias that is simply assumed to exist and to be unjust and replace it with real bias, encoded in our laws and practices, or in this case, in actual code.
Setting aside blatant shock behaviors... If the other side, the audience, were less sensitive and not looking for the next micro-outrage, wouldn't ML chatbots evolve more pro-social values by positive reinforcement?
It takes two to Tango .. the average audience behavior isn't blameless for the impact of its response. Also, how an AI decides interprets an ambiguous response as being desirable or not is really interesting.
> As a hybrid approach, you could produce a large number of inferred sentiments for words, and have a human annotator patiently look through them, making a list of exceptions whose sentiment should be set to 0. The downside of this is that it’s extra work; the upside is that you take the time to actually see what your data is doing. And that’s something that I think should happen more often in machine learning anyway.
Couldn't agree more. Annotating ML data for quality control seems essential both for making it work, and building human trust.