Yes. I expect to get results that contain the keywords "shirt", "stripes", and "without", but perhaps with the last term ignored. Why, what would you expect? Would you search for "shirts that don't have stripes"? Or "shirts lacking stripes"?
It sounds like you expect the search engines to understand natural language, because your expectations are influenced by what is "touted". Do you think that's a realistic expectation?
Google started doing that with its regular search queries several years ago. There are thousands of complains on HN about how it's almost impossible now to find anything without "quoting" "every" "word" "of" "your" "query".
This is the only interface you really get into the amazon catalog. What if you really are looking for shirts that don't have stripes? How would you formulate the query?
Like Google, Amazon is supposed to support the - operator, but, also like Google, it's not clear whether it does anything at the moment. As others here have suggested, maybe "solid shirts"?
Of the first ten hits:
Amazon: 7/10 aren't shirts (but all are stripeless)
Google: All shirts, one is checked
Bing: All shirts, all contain stripes in description.
Not exactly glowing for Amazon, clearly unsupported by Bing.
I would prefer results containing words "shirt", "stripes" and "without". I don't want search engines to freely interpret my query and guess what I probably wanted to see while burying exact results in hundreds of pages of useless garbage.
Exactly. Those who want the search engine to "understand" what they are asking for when they use words like "without" may not have thought through the unintended consequences. For there is no combination of search terms that might not be interpreted in a variety of ways that the searcher never imagined, because of the ambiguity of natural language. Better to learn and use search operators, and the strategies of careful choice of keywords.
How is interpreting "shirts without stripes" as "I wish to see shirts that don't have any stripes" guessing?
I would venture to say that most people (meaning your use case is in the minority) who type "shirts without stripes" want to see results showing shirts without stripes not results "containing words "shirt", "stripes" and "without".
I think what is happening here is that you know how search engines work and so you are conditioned to expect them to do what they're doing.
«How is interpreting "shirts without stripes" as "I wish to see shirts that don't have any stripes" guessing?"»
Because I was looking for pages about the band called "Shirts without Stripes". Because I wanted pages with shirts that have stripes but where the page featured the word "without", because their shirts are without something else. Because I want to see striped shirts from the company called "Without". I don't want the machine to guess what I mean. It can never know.
I think it depends on the application. If I’m searching for products in a store I do want “shirts without stripes” to give me shirts without stripes. I have a hard time thinking of a case when including results with the word “without” would be useful when it comes to online shopping.
When I’m looking up information, like an article, especially technical information, then yes I want the search engine to do as little interpreting as possible. That’s because any interpretation rules it uses won’t always be relevant so I’d rather have more control.
But for Amazon? I don’t see how someone can see the average user typing “shirts without stripes” and getting almost nothing but shirts with stripes, and going “yup, works as expected”.
Precisely, as there is a better way to do it. Search for "solid-colored shirt" or "black shirt" and/or further define your search to exclude the term "stripes."
the icab link from yesterday showed a screenshot of amazon when their categories listings were still meaningful and helpful. nowadays, it's just a pain.
i recently looked for dough scrapers. i wanted to see what's selling best and what's most rated. they are everywhere. in dessert & decoration, in utensils, in bakeware, and many other categories. i mean i get it...
it's not just search that's hard. categorization is also an issue here.
Although search within apps can be even worse.
I was looking through Google Movies the other day for the film 2001 and instead it swamps the results with those from the year 2001 - one could argue that there are lots of people who are massively keen to buy films based on the year of production, but I suspect it's better to satisfy those looking for years in the title first and then after that brief interruption list the year based results, rather than the other way around.
Similarly looking for "The Book of Why" on Audible is dismal: even when it's in quotes it isn't until the 42 result that the exact match shows up, with a load of useless not-obviously connected results turning up first.
Both these failures interest me as they have a tangible financial implication (I clearly had money to spend) and yet they remain unfixed.
I realize the obvious fix for the Audible problem is to interpret double quotes as "only search for this phrase," but I wonder if an alternative solution would be do TFIDF for combinations of wordsm, not just individual words. For example, the search crawler could watch for the phrase "the book", "the book of", "the book of why", "book of", "book of why", and "of why", and weight the search results accordingly.
Yes, that would be a better user experience as:
1. not everyone will think to use quotes
2. it would probably be more resilient to minor user errors too (eg if it had been misused misheard as The Book of Way you'd still have a fair chance of it returning the desired result fairly high in the list)
None of them make an attempt at filtering this. While I found it amusing because they advertise with AI magic in other areas, those search engines all advertise to be keyword searches and one of them being "the worst" means it's either random (since none of them are trying) or they are the best (since it matches the stripes keyword).
To be honest I wish they actually were keyword searches and that the machine doesn't try to be smarter than you. Many times when I carefully specify which keywords must appear on the page, it'll ignore parts of the query or add unrelated synonyms. Usually one can work around it with operators but it's tedious and doesn't work reliably.
They want to always return results even when they don't have any. So they will remove some of your query. This last week Google even returned months old results when I asked for the last 24 hours. It's so frustrating. I wish the Google team that decided this nonsense have to migrate Python 2 to Python 3 projects until they retire.
Yeah that works, but it's not the point. The author isn't saying there's no way to search for shirts that don't have stripes. They're saying search engines still don't understand how to find shirts without stripes in the same way humans do.
I think that this is actually really encouraging in showing that we still have a ways to go in improving search engines. A lot of people treat search engines as a solved problem, at least for non-question answering aspects.
I am amazed at the amount of things people are willing to treat as "solved problem".
At a point there was a TED talk explaining social networks were a solved problem now that facebook was dominant. Recycling was seen as solved problem until it wasn't etc.
I wonder how many actual "sovled problems" we have.
Google has gotten significantly _less_ useful for finding technical information over the last decade or so. It used to be the case that when searching for some tech-related item (say, how to use functions in bash), the results would take you to someone's personal website or a wiki. Ostensibly, the more people linked to it and the more people clicked on it in the results, the better it ranked for that query and similar queries.
Today, entering in any tech-related query at all takes you to StackOverflow, end of story. Not only are SO answers quite often outdated (or even terrible advice in general), most of the time I'm not looking for a "here's how you do X", I'm looking for background information on a topic.
Most non-tech queries I put into google are even _more_ useless as the results tend to fall into these categories:
* Wikipedia (okay for _very_ general things, useless for domain-specific knowledge)
* SEO-enhanced blogspam, (a.k.a. "8 Weird Ways to Earn Millions Through Gaming The System!")
* Tweets on twitter (!)
The dev/tech industry desperately needs a search engine that somehow prioritizes _quality_ content, not one-off answers, blogspam, and tweets.
That is, learn to structure your query in a way that Google understands what you're trying to say. This used to be what yuo had to do, but now that Google tries to understand the intent of what you're trying to say and advertises as such, it's clearly a hack.
The point of this isn't asking how to apply boolean search operators, it's showing that the largest AI-focused companies in the world absolutely suck at NLP.
Not that but maybe someone can explain if the technology is "AI" why it still takes tens of thousands of hours to implement a project? Shouldn't we call it AI when the frameworks can be implemented with relative ease? My point is that the level of customization required to achieve AI, in my opinion, doesn't make it very intelligent.
One of the primary methods to join multiple learned models is a decision tree (or combination of decision trees making a decision forest), which can be simplified as a series of if statements/conditionals. So if you join a 'is it a shirt' model with a 'is it striped' model you get two sets of things, and with how big data approaches this it is something you can do quickly. As other people have pointed out the issue here is that the NLP of the actual search is not creating a negation of two sets, it is returning the intersect of sets 'is it a shirt', 'is it striped' and shrugging its shoulders and either intersecting it with 'things with the text without', or throwing up its hands entirely based on the context because it wasn't programmed to do something smarter.
The term has been bastardized. AI is now any app that uses if statements, so long as it's obscured to the user. So now "AGI" exists, because apparently the word "general" now means "intelligent", unlike "intelligence".
At some point in the future marketers will learn about AGI, and we'll have to make yet another term, maybe artificial general practical intelligence?
Right, because when I think of "intelligence" I assume it's specialized and otherwise is dumb as rocks. Naturally. Which is why marketers are really just advertising specialized tools, not agi. Obviously.
I know what agi is; I just find the terms backward.
When you think of intelligence, you think of the human kind. As we create different things that have some of the properties we usually identify as "intelligence", but lack others that were completely correlated until that time, we need to create better words that can express the new situation.
It's perfectly ok to prefer the qualifier to go the other way around, keep intelligence general and change the name of the specialized form. But that's just not how our language evolved.
Or maybe there's just not a lot of AI in their main search product (for whatever reason). They seem to be pretty good in other areas (Translate, Cloud speech-to-text, Alexa etc.)
For the matter of translation they actually aren't pretty good. DeepL outsmarted all of them (at least for the supported languages). Given the resources thrown at this problem by the different companies I would even say the results of Google, MS and such are actually quite disappointing. (And I don't think one could say "but this companies did the upfront work" as the basic ideas are something 60 years old).
Generally, yes. But you can go very far and cover a lot of cases when you stick to a narrow domain and common patterns that customers of your product will typically use.
Siri has lots of known problems. My biggest gripes are the lack of context and lack of a "discussion" state.
But Siri is a general domain problem, which is really really hard. Siri set the expectation you can ask it anything, and it works terribly and for most questions just gives up and runs a web search.
If you are an e-commerce company though, that's a narrow enough domain, because you know that for most people, they're looking for products to buy or compare. It's not an unbounded Q&A service.
Yeah the examples would probably do better if they had a front layer model determining the general topic "clothing" then shipping the query off to the clothing and fashion specific model for details.
Lots of focus on a general purpose mono-model, but
I think a collection of specialized subsystems is a better representation and would produce better results, faster.
Why would you really apply NLP to a search engine though? Generally speaking a weighted keyword search is good enough 95% of the time and requires significantly less resources to perform.
I work specifically in this field with clients, and deliver training on applying NLP to search.
You’d be surprised how effective NLP is for use when identifying query intent, and pulling out modifiers that should apply as metadata filters.
Weighted keyword search works a lot, but it fails hard for many long tail queries (especially in e-commerce and other attribute heavy domains).
IMO there really isn’t a good excuse for these firms to fail at queries like this. The query itself isn’t particularly difficult when using a decent NLP stack and following well known practices.
If it's technically possible then presumably it's a deliberate product choice to not have better search results for "shirt without stripes". And that seems entirely plausible.
Google is already by far the most widely used search engine, so they don't really need to innovate or improve the search product very much in order to attract and retain users. Presumably capturing more advertising spending from the companies paying for ads is a bigger priority.
Microsoft under Satya Nadella has been all about enterprise and cloud, and I doubt Bing is a strategic priority any more, so it's not surprising that they wouldn't put a lot of resources into making it better.
Amazon is a little surprising. You'd think they'd have a lot to gain from making it easier for people to find what they're looking for. But maybe less than perfect search results are deliberate? Maybe it's like how supermarkets put basic items in the back of the store and high-margin impulse buys in the front - so you have to walk past chocolates and chips if you want to buy a carton of milk.
If Amazon is deliberately nerfing search results then maybe Google would stand to benefit from having better shopping-related results - people would get frustrated trying to find a shirt without stripes on Amazon and just use Google instead, letting Google profit from advertising in the process. But maybe people selling shirts aren't willing to pay much for ads, so there isn't much money for Google to make by getting better at finding specific types of shirts.
I dunno if any of these conjectures are anywhere near accurate, but it's interesting to think about.
I feel like that's stretching the definition of NLP though. Technically you are processing natural language but it seems like you've found that doing essentially a keyword match but using certain keywords as more advanced filters rather than just search terms.
Why would you really apply NLP to a search engine though?
If you go to Google's homepage and click the microphone at the end of the search input box you can search by speaking. All it does is convert to speech to text, but it implies you might be able to search in a more "natural language" way.
Depends on your audience, but I imagine many find the answer boxes on Google search pretty useful. Getting the population of a city without having to click any links is probably good for your perceived value. For this you need some NLP tech to extract intent from the query and match it to the right entity in their knowledge graph (in addition to something to help you build the graph in the first place).
I think that is probably not a common enough use case to optimize for. Additionally it would be easy for a user in that position to just search "shirt" and ignore the occasional striped one, or to search "polka dot" and "paisley" seperately
I understand the point, but there are plenty of bigger fish that I would want Amazon and Google to fry before spending their engineers' time on a triviality like this. I just don't think that having to make three queries instead of one in this occasional situation is such a big deal.
You're arguing the search is "good enough". Because we can adapt to the machines. But the company who will not force us to do so will get our business. The company who creates the best digital butler will win. They know this. They try hard. And they still fail at simple stuff when judged by humans.
Another possibility is that maybe it really is "good enough", and they get a bigger advantage by competing with each other on more important aspects of usability than these kind of trivial issues
Assumptions like this is what makes search so terrible for many many companies.
You can't assume that customers would type one thing or another - you need to gather lots of query log data and see what you find. You'd be surprised how much variation there is, but once you do have this data you can then find patterns to cover lots of (but not all) cases.
Of course, they should look at the data before making such an assumption. I am not suggesting otherwise, I'm just making a guess as to why it's been done this way. I suspect they have looked at the data and identified that it is not a common use case.
Apple probably expected most users genuinely wanting to make a phone call would hold their iphone 4s a certain way. Turns out expectations don't always match reality.
From a product perspective, I would say there is a reasonable expectation that a customer will provide that query and expect the results to come back with plain shirts. Anything different is a degraded customer experience. Sure, a technical user will understand which queries to provide better, but 90% of customers won't have that skillset. Its our job as engineers to serve those people, and the queries they provide.
So NLP is totally a thing you want to have in search. Arguably, its the whole point of search as it exists now.
Certain query styles which the NLP operators help clarify. Recent personal examples: "is butyl rubber an organic or inorganic compound?" and "which gloves are best for both acetone and nitric acid?"
I agree, I think it's a better idea to improve the knowledge management (in terms of a relational hypergraph) before trying to create the bridge between the natural/digital knowledge representations. I have a hunch a good knowledge management system would be more amenable to being queried using natural language since relations are probably constructed similarly to the patterns we see in natural language syntax.
all three of these search engines offer "voice assistant" platforms that encourage you to speak to them in natural language, and send your query to the search engine verbatim under a broad set of circumstances.
It's not as simple as NLP. The shirts them selves have to labelled as having stripes or not. If striped is not a attribute of the shirt, it doesn't matter of the NLP can parse the meaning of striped.
IIRC if you use quotation marks in your search query, it would search for the exact match of that phrase without any search operators processed inside of it. So those scenarios you listed would still work fine.
Though I somewhat agree, I don't think that an average user even knows about existence of search operators in the first place, let alone being aware of this specific one and when to use it.
What happens when someone asks you about one of those items? You consult your memory reserves, and you find that those are proper names of specific entities. So your brain returns those entities. Now there very well be a high school band called "Shirts without Stripes", you most likely would call up plain shirts or shirts with non-striped patterns. No reason that a search engine wouldn't follow the same rules (i.e., Google must have millions of entries for Doctors without Borders and Men without Hats).
Search engines are indexing content based on [edit]STATIC [/edit] keywords, though.
The author's intent exceeds both the capabilities and intended use case of search engines.
The query "shirts without stripes" if interpreted by human would require any search system to not only analyse the keywords and tags (of the products/images), but also the content, which is an infeasible task given its dynamic nature.
So the author wants: select all shirts where content analysis of returned images yields no stripes.
This is a context-sensitive image/product search based on arbitrary, dynamically created criteria and shows that user isn't aware of what the search functionality does as opposed to exposing weak "AI". [edit]To clarify: you cannot add all possible keywords/criteria in advance[/edit]
This sounds great, right? But consider the recommendation engines that are part of Amazon, Stack Overflow, whatever. You've just filtered out pages that have a shirt without stripes on it ... but have a striped shirt somewhere in a recommendation bar on the page.
I've run into so many variations of this. You can search for something only to have the recommendation/related results embedded on whatever page to throw off your results one way or another.
I genuinely think that whatever standard HTML/XTHML is at ought to have, either as an attribute or a semantic tag, some kind of "related" or "recommended" ability to set that content apart. My cynical thought is that even if it were adopted, it would probably get abused in some fashion.
This is expected, not wanted though, I would expect some semantic analysis translated into "shirt -stripes", but what you really mean is "solid color shirt". This is a tough one but surely something that can be tackled with research
shirt -stripes does not mean "solid color shirt", as the - operator looks for that text on the page, instead of performing a semantic "I don't want stripes".
The point that the author is making, in a very understated way, is that all three companies have PR websites that breathlessly describe their advanced AI capabilities, yet they cannot understand a very simple query that young children can.
He probably meant "more spicier". Second image for "now one with stripped this time please" yielded image that is linked to article about deep-fake nudes.
Does Amazon pretends to do AI? They are just offering a platform to do your own Machine Learning. I don't think they ever said their search engine was doing anything smart.
EDIT: scrap that, I didn't mean Alexa, which is doing AI obviously, but the search engine of Amazon's retail website.
Anyway, NLP is hard and everyone sucks at it. Think about it: just building something that could work with any <N1> <preposition> <N2> or any other way to express the same requests would mean understanding the relationships of every possible combinations of N1 and N2. It means building a generalized world model that is quite different from simply applying ML to a narrow use case. Cracking that would more or less mean solving general AI which probably won't happen soon.
Not actually true. ML is one area of study within the field of AI. Thanks to marketing departments and slightly shoddy journalism these two things are now casually treated as equivalents, but they're really not: ML is still very much a subset of AI.
I disagree, "shirt without stripes" is an unusual word choice, not one that our ML models would be optimized for. Try "solid color shirt" and you'll see how much better the results are - at least on Google.
"Shirt without stripes" may (or may not) be an unusual word choice to enter into a search engine, but it's definitely one that a child would understand.
Additionally, "shirt without stripes" is not the same as "solid color shirt"; as an example, take a look at:
Quite so. "Shirt without stripes" can include shirts in plenty of patterns other than solid colours (paisley, polka dot, checked, battenberg, floral print, etc.).
You have to realize that search is not AI. It's pattern matching. And the string "shirt without stripes" matches really good with "shirt with stripes". Levenshtein distance is 3.
If vendors would use the term "shirt without stripes" than it would match great, but they call it "plain shirt".
Yes, that exact sequence of words isn't particularly common. And yet a child, even if they have hey have never been exposed to it, has no problem understanding what it means.
Whereas all these services seem to be processing the input in such a superficial way that they give the searcher results that aren't just inaccurate but are the opposite of what was asked for.
“Chicken without head”, “men without pants”, “sky without clouds” only work because the users uploading the images tended to tag them as such... (in that case the users do the hard coding of meaning)
I would go in a store and search there. Not everything has to be solved by tech and AI. This is a prime example of a problem that requires insane amount of work and yet provide absolutely no value to the world.
People still think we will have self driving cars "in two years" yet here we are talking about dumb shirts. AI winter is coming
Just searched same on google and first non-plain shirt was in the second hundred. Duckduckgo was similar. Considering that they classify images according to surrounding text, it seems like pretty good result.
I couldn't quite believe your comment when I read it so I did a Google image search for "person" and the results weren't a lot better than you'd suggested. Mostly white men, a few white women, a very few black women, a handful of Asians, and multiple instances of Terry Crews.
The net result of that Google search, combined with the "Shirt Without Stripes" repo, leaves me even more unimpressed with the capabilities of our AI overlords.
I just did a google image search for "person". The first 5 images were of Greta Thunberg. She must be the most representative person ever.
The next few images contained Donald Trump, Terry Crews, Bill Gates and a French politician named Pierre Person.
After that it was actually quite a varied mix of men/women and color/white people.
I am still not very impressed with Google's search engine in this aspect, but it is not biased in the way you suggest.
At least it is not biased that way for me. As far as I am aware, and I might be completely wrong here, Google, in part, bases its search results on your prior search history and other stored profile information. It is entirely possible that your search results say more about your online profile than about Google engine :)
> The first 5 images were of Greta Thunberg. She must be the most representative person ever.
Well, she was the 2019 Time Person of the Year.
Likewise, Trump was the 2016 choice, and Crews and Gates have been featured as part of a group Person of the Year (“The Silence Breakers” and “The Good Samaritans” respectively).
I think the skewing of results lessening your impressed-ness is the wrong takeaway. If anything, the AI is a more perfect mirror of the society it learned from than you expected. Perhaps the right way to look at it is that we are capable of producing things that we don't understand, that are more sophisticated than we realize.
You may be right. It's been bugging me since I posted earlier on so I fired up a VPN with an endpoint in Japan, along with a private browsing session in Firefox, to see if I got different results. As it happens the results were interesting:
- If I entered "person" I'd see a mix of images substantially similar to what I saw using google.co.uk up to and including Terry Crews, which was frankly a little weird, and otherwise mostly white
- If I entered "人", which Google Translate reliably informs me is Japanese for "person", I'd see a few white faces, but a substantial majority of Japanese people
So it seems possible that Google's trying to be smart in showing me images that reflect the ethnic makeup I might expect based on my language and location. I mean, it's doing a pretty imperfect job of it (men are overrepresented, for one) but viewed charitably it's possible that's what's going on.
Is the case for woke outrage against Google Image Search overstated? Possibly; possibly not. After these experiments I honestly don't feel like I have enough data to come to a conclusion either way, although it does seem like they may at least be trying to do a half decent job.
This seems like you're attributing motive to google here, but I don't believe that's right. For example, Terry Crews appears in the query "person" because his "TIME Person of the Year 2017 Interview" article was very popular online. I get a lot of Greta Thunberg because she was TIME Person of the Year 2019 and received similar online attention because of Donald Trump.
The TL;DR of it is that google crawls the internet for photos, associates those photos with text content pulled from the caption or from the surrounding page, and gives them a popularity score based on the popularity of the page/image. There are some cleverer bits trying to label objects in the images, but it's primarily a reflection of how frequently that image is accessed and how well the text content on the page matches your query. There's some additional localization, anti-spam, and freshness rating that influences the results too.
The majority of pages with "人" and a photo on it that has a machine labeled person image would be a photo of a japanese/chinese person, and if you're being localized to japan with a vpn, that would be even more true.
Google doesn't "know" what you're trying to search. It's a giant pattern matching game that slices and dices and rearranges text to find the closest match.
> Google doesn't "know" what you're trying to search. It's a giant pattern matching game that slices and dices and rearranges text to find the closest match.
I'm not disputing that, and it certainly explains why it's "good enough" for somes search queries whilst being totally gimpy for others.
My understanding was that Google does prioritise what it's classified as local search results though, on the basis that they're likely to be more relevant.
This is the problem though, all those companies are advertising fantastical results. They aren't saying "Hey! We spent billions of dollars so our algorithm could be as racist as your uncle Steve!". Oh and by the way, Steve is now right - because all the crimes he ever finds out about are by black people, because that's what Google has decided he wants to see. So it's no longer him seeking out ways of justifying his latent racist tendencies, no, he's outsourced that to Google.
Interestingly, duckduckgo shows me, as second result, an albino tiger with, you guessed it, no stripes. The page title has "[...] with NO stripes [...]" in it, so I assume that helped the algo a bit.
EDIT: I also got the painted horse (it looks spray-painted, if you ask me) and I must admit it's quite funny to look at
> The AI has no morality. It simply reflects and amplifies the morality of the data it was given.
Key point right there. Unless Google is deliberately injecting racial and/or gender bias into their code, which seems extremely far fetched (to put it kindly), the real fault lies with us humans and what we choose to publish on the web.
What bias? Who is biased? Quick duckduckgoing indicates there are far more male than female doctors in the US. So statistically, it would be correct to return mostly male doctors in an image search. If you want a photo of a specifically gendered doctor, it's not hard to specify. Not really seeing a problem here.
I would contend that society is biased. There is no evidence that says men are better doctors than women, and in fact what little this has been studied says that women make better doctors than men (and is reflected in the more recent med school graduation classes which are majority women).
So it's a question of what you are asking for when you search for [doctor]. Are you asking for a statistical sampling or are you asking for a set of exemplars?
> So statistically, it would be correct to return mostly male doctors in an image search.
And that's exactly it. The AI has no morality. It's doing exactly what it should, and is amplifying our existing biases.
Honestly, I don't think morality is the issue here; it is objectively inaccurate to show only white men for the search string "doctor" when not all doctors in the U.S. are white men, and most doctors in the world are not white men. This would be like showing only canoes if someone searched "boat"--we would rightly consider that an error to be corrected.
IMO, wrapping it in a concept like "morality" because the pictures have people in them just serves to excuse the problem and obscure its (otherwise obvious) solution.
Those look almost entirely like stock photos or part of advertisements. It's probably just reflecting the biases of what photos other businesses like, which get the label of "doctor" or "nurse".
Any sort of image search is going to tend to be biased toward stock photos, because those images are well labeled, and often created to match things people search for.
> So really you can't blame anyone but society for having such deeply engrained biases.
You can blame statistics for that. Beyond that, you can blame genetics for slightly skewing the gender ratios of certain fields and human social behavior to amplify this gap to an extreme degree.
What does the color of people's skin in search results have to do with morality? I was raised not to see color, now we have this "progressive" movement hell bent on manipulating search results to disproportionately represent minorities. If you want to filter your search results based on the color of skin you can do that easily.
I tried this as well in an incognito window on Firefox and got the results you mentioned. I notice, however, that virtually all of the results have associated text containing the word person. It seems likely that Google image search featurizes photographs to include surrounding document context.
(That's how I would do it if I wanted more accurate rather than more general results.)
4 of my top 7 images (the top line) are Greta Thunberg in a search for "person". First viewport is 11 men, 11 women, 1 stick person, of which there are 4 Thunbergs, 4 Trumps, 2 Crews. People seem to be if they got major "person" awards like "most powerful person" or "person of the year".
There's not much diversity, assuming Terry Crews is from USA, then all the first viewport full of images are Western people; except Ms Thunberg they're all from USA AFAICT [I'm in UK].
The first non-Western person would be a Polish dude called Andrzej Person (the second Person called Person in my list after a USA dancer/actress), then Xi Jinping a few lines down. The population in my UK city is such that about 5/30 of my kids primary and secondary school, respectively, classmates have recent Asian (Indian/Pakistani) heritage. So, relative to our population, there are more black people, far fewer Indian-subcontinent no obviously local people.
Interesting for me is there are no boys. I see girls, men and women of various ages but no boys. 7 viewports down there's an anonymous boy in an image for "national short person day". The only other boys in the top 10 pages are [sexual and violent] crime victims.
The adjectives with thumbnails across the top are interesting too - beautiful, fake, anime, attractive, kawaii are women; short, skinny, obese, big [a hugely obese person on a scooter], cute, business are men.
I don’t understand why AI or a search engine had to meet your or anyone’s expectations for diversity. If I searched for “shirt” and didn’t get shirt pictures in the color I wanted I would just tune my query instead.
Jokes on you. Not having diversity is now considered incorrect, even if it wasn't stated. AI needs to learn to keep up with the craving for relevance the rest of Silicon Valley has by ensuring all results comply with whatever equal opportunity mantra is now in vogue. The next time I search for "CSS color chart" I expect the preselected color to be black.
Most of the person results appear to be 'Time Person of the Year' related. Another result is a guy with the last name Person. The results don't seem to be related to the definition of the word 'person'.
For me it shows all newsworthy people and articles. It shows the titles of the pages and they are all stuff like "11 signs you are a good person" So it seems clear that there is no kind of AI bias here but simply that high ranking articles with the word person more often than not choose white men as their stock image.
Most of the very top results seem to be of trump and greta thunberg.
Yeah, that's one plausible explanation. (I don't remember the nature of the letter.)
Relatedly, one time I picked up a prescription for a cat. The cat's name was listed as CatFirstName MyLastName. They had another (human) client with that same first and name. It turned out that on my previous visit they had "corrected" that client's record to indicate that he was a cat.
I think search algorithms still have a long way to go to really understand the intention. Try your image search results for
"white person" "black person" "asian person"
"white inventors" "black inventors" "asian inventors"
Doesn't quite deliver what would be expected.
One can play this game a lot and most results will return expected cultural biased results. A "kind person" is apparently a white girl. A "good person", a white woman. A "bad person", white men. A "evil person", white men. A "honest person", equal mix of white women and white men. "Dishonest person", white men in suits. "Generous person", hands of white women. "Happy person", women of color. "Unhappy person", old white men. "Criminal person", Hispanic men. "Insane person", white men. "Sane person", white women.
Is it surprising that very few of the result surprises me?
Yes. I thought about words people use in priming studies, usually in order to trigger a behavior, and just typed the word with space and "person" appended.
I did use images.google.se in order to tell google which country I wanted my bias from since that is the culture and demographics I am most familiar with. I also only looked at photos of a person and ignored emojis.
I have also seen here on HN links to websites that have captured screen shots of word association from google images and published them so you could click a word see the screen shot. They tend to follow the same line as above, but with some subtle differences, and I suspect that is the country culture being just a bit different to mine.
You really should link to screenshots of your results so people can judge for themselves.
I just submitted all your searches to google.com from Australia, and the results were nothing like what you described; all the results were very diverse.
This is to be expected, as Google has been criticised for years for reinforcing stereotypes in image search results, and has gone to great effort to adjust the algorithms to reduce this effect.
I usually don't spend time producing evidence since no one else does it, nor did the parent comment, or you for that matter. It also tend to derail discussions onto details and arguments over word definitions.
First is happy person. Out of 20 we have 14 women, 4 guys, 2 children.
Second is criminal person. The contrast to the first image should be obvious enough that I don't need to type it.
If I type in "person" only I get the following persons in the first row in following order:
Pierre Person (male)
Greta Thunberg (female)
Greta Thunberg (female)
Unnamed man (male)
Unnamed woman (female)
Mark zuckerberg (male)
Keanu Reeves (male)
Greta Thunberg (female)
Trump (male)
Read Terry (male)
Unnamed man (male)
Greta Thunberg (female)
Greta Thunberg (female)
Unnamed woman (female)
Unnamed woman (female)
Resulting in 8 pictures of females, 8 males, which I must say is very balanced (I don't care to take a screenshot, format and upload, so if you don't trust the result then don't).
Typing in doctor as someone suggested in a other thread I get in order (f=female, m=male): fffmffmmmmfmmfffmfmfmmmff
and Nurse: fffmffmfmmffmffmfffmffmffff
Interestingly the first 5 images have the same order of gender and are both primarily female, through doctor tend to equalize a bit more later while nurse tend to remain a bit more female dominated.
Thanks for the screenshot. It helps (and by the way, yes the onus is on you to provide evidence as you're the one making the original claim).
Your initial comment said "Happy person", women of color.
But your screenshot showed several white people, several men, and a diversity of ages. Yes, more women, which is probably reflective of the frequency of photos with that search term/description in stock photo libraries and articles/blog posts featuring them. No big deal.
You also said "Criminal person", Hispanic men
But the screenshot contains more photos of India's prime minister than it does of Hispanic men. In fact I can't see any obviously-Hispanic men, and the biggest category in that set seems to be white men (though some are ambiguous).
The doctor and nurse searches suggest Google is making some effort to de-bias the results against the stereotype.
To me the biggest takeaway is that image search results still aren't very good at all, for generic searches like this.
Indeed it's likely that they can't be, as it's so hard to discern the user's true intent (for something as broad as "happy person"), compared to something more specific like "roger federer" or "eiffel tower".
Yes. If I had the energy and time to do a proper researched data set I would have a bot search through the top 100 common words associated with either warmth (sociability and morality) or competence, and then use a facial recognition system go through the first 100 images of each to determine the distribution of gender, age and skin color.
Following the stereotype content model theory I would likely get a pretty decent prediction of what kind of culture and group perspective produced the data. You could also rerun the experiment in different locations to see if it differ.
FWIW, this is most likely not a bias of the search engine, but just a reflection of its sources (mostly stock image platforms I suppose). So if most stock images of blue trolls would be labelled with "politician", you'd eventually find blue trolls when searching for "politician".
You've raised an entirely unrelated problem. Showing shirts with stripes when you search for "shirts without stripes" is just plain wrong. Showing only a single demographic of person when you search for "person" is correct, it just doesn't have the level of diversity you seem to want. Nothing about diversity is implied in the query, and so your observation is completely unrelated to a plainly incorrect query.
On the other hand, the bias in the results means they're somewhat incorrect: there is more than one demographic of person, showing only one in response to a query that doesn't ask for a particular one is incorrect.
If you were unfamiliar with them and searched "widgets" to find out more and got widgets of a single colour and form, it would not be an unreasonable assumption that widgets are mostly (if not entirely) that shape and colour, especially if there was nothing to indicate that this was a subset of potential widgets.
It's not so much "demand for diversity" as it is "more accurate and correct representation".
The inverse: a favorite trope of the American far right is that GIS for "american family" will show you photos of... mixed race families. (Something the far right has strong opinions on, and is a tiny minority of all marriages in the US)
Something of a corollary to Brooksian egg-manning: with an infinite number of possible searches, you can find at least one whose results do not exactly match the current demographics of the state from which you place the search.
More than racism on the part of google[1] I would attribute that to it being an hard problem with too many dimensions. About three years ago if you searched "white actors" google would give two full pages of only black people (I have no idea whether the actor part was correct).
Many interpreted this along tribal lines, but likely it is that there is constant tuning and lots of complex constraints.
[1] not to say that you implied the reason was racism, but often it is attributed to something along those lines
That point is akin to stating: These three companies have not solved the hard problem of common sense [1], so are not allowed to advertise their AI without looking silly.
Nobody has solved the common sense knowledge problem yet. A solution for that would qualify as Artificial General Intelligence and pass the Turing Test.
But search engines have come a long way. I even suspect that when search engines place too much logical - or embedding relevance to stop words such as "without", that, on average, the relevant metrics would go down. It is not completely ignored as "shirt with stripes" surfaces more striped shirts than "shirt without stripes". "shirt -stripes" does what you want it to do.
Searching for "white family USA" shows a lot of interracial families. Here "white" is likely not ignored as much, and thus it surfaces pages with images where that word is explicitly mentioned, which is likely happening when describing race.
You can use Google to find Tori Amos when searching for "redhead female singer sings about rape". Bing surfaces porn sites. DDG surfaces lists (top 100 female singers) type results. The Wikipedia page that Google surfaces does not even contain the word "redhead", yet it falls back to list style results when removing "redhead" from your query, suggesting "redhead" and "Tori Amos" are close in their semantic space. That's impressive progress over 10-20 years back.
I think a huge Chinese room type parser with a bunch of heuristics bolted on probably provides much better bang for the buck than trying to implement actual NLP (in every possible language, or even just in English). So that's probably what nearly everybody is doing.
Google has for years put out puff pieces talking about high accuracy on image tagging. It’s only within the last few months that searching my Photos library for “cat” returned something other than pictures of my dog.
There’s a nuanced argument that practitioners know how ML is so dependent on training data and accuracy tails off sharply, but that nuance tends to removed from anything selling to potential customers — which has not been a great way to keep them in my experience.
I'd assume pets are hard as there are so many varieties (potentially even harder than humans). For the last few years Google Photos has correctly returned photos for a search of "Lamborghini" in my albums. I'd expect "shirt stripes" to fall into that category.
Sure, I’m not saying it’s an easy problem — just that the marketing is once again setting the field up for failure by giving the impression of human-level performance but delivering results only in very narrow scenarios.
I have a PhD in NLP (which is what we often call it on the CS side, but is almost synonymous with CL="computational linguistics" on the cognitive/linguistics side of the field). I remember a talk at our annual conference, well-attended, perhaps around 2003 or so. The speaker was from one of the labs that was really leaning into "big data", which was only just becoming possible at that point, and argued persuasively that we should all just throw out our parsers and formalisms—ditch the computational linguistics side, basically—because we were on the edge of functionally infinite (unsupervised) data, and supervised and partially supervised systems would never ever be able to keep up. He presented performance numbers and how the unsupervised systems needed a lot more data to compete with the supervised systems, but that data was available, and he threw more and more and more data at the system and it got better and better. (I no longer remember the specific task he was using to illustrate his point.)
There were gasps in the room and a kind of depressed acquiescence: geez, he might be right. And the pendulum indeed swung in that direction, hard, and the field has been overwhelmingly dominated by the statistical machine learning folks on the CS side of the field, while the linguists kind of quietly keep the flame alive in their corner.
But I thought then, and I still think now, that it really just was another swing of the pendulum (which has gone back and forth a few times since the birth of the field in the 1960s). Perhaps it's now time again for someone to ring up the linguists and let them apply their expertise again?
To me, the most interesting implication here is that this must not adversely affect Google's ad revenue. If it did, they would surely fix it. This, in turn, means that apparently we have been trained to interface with search engines such that this is not a problem.
Sometimes I wonder how much my brain has changed to use search engines / how much of it is dedicated to effective googling. Makes me feel like a cyborg.
The other day I found that googling something like “mesothelioma -lawyer” will exclude results with “lawyer” but all the ads will still be for lawyers and contain the word “lawyer”.
I guess because they leave it up to the advertiser to determine the negative match words and that seems to always have priority.
Similar thing happens on Twitter. If you mute a phrase it will block organic content including the phrase from appearing in your timeline, except ads including the muted terms still appear in your timeline.
That sounds like an ad business version of the efficient market theory. E.g, that can't possibly be a hundred dollar bill on the ground, because if it was someone else would surely already have picked it up by now.
I think you're overestimating Google's sophistication.
What exactly would happen if it was a material negative impact on their revenue and they didn't fix it?
If Google isn't under survival pressure to get better (and they aren't) the incentives aren't aligned for them to improve or even to not get worse every year.
If Google is failing first gradually then suddenly it might not even be within the institutional power to notice how bad it's become before it's too late.
That statement was true of every past search engine, too. AltaVista must have been great and its severe limitations must not have affected revenue, or else they surely would have improved it.
Everything is for the best, in this best of all possible search engines -- the Candide fallacy.
You're joking, but perhaps you're right. It's unnatural to search in the negative -- "solid shirt" or something like that would be more likely from a human with that actual intent.
This just suggests to me that real humans haven't issued that type of search query enough for the AI to know what to do with it. Which wouldn't be so big of a problem.
Vaguely similar to a joke from Ninotchka that Zizek often uses about the difference between 'coffee without cream' and 'coffee without milk'. He usually uses it to reference the concept of negation in the Hegelian dialectic, but he's also mentioned the difficulty of computers understanding negation in the context of the coffee/cream example.
It's not enough to say "Oh, we should add a rule that 'without' means negate the next word" because that only applies to this one situation, in this one language. Let's generalize the problem: We aren't correctly translating from English (or other spoken languages) to Computer/Logic.
The state of the art in machine translation (from what I've read at least) is translating from language-A to a language-less "concept space" and then from there to language-B. Could that be done where the output language is something a search engine can use to find what you want correctly?
Given that pattern, I suspect we could see much better results in cases like this.
Only slightly related but a couple of years back I got an alexa as a gift. When you open the alexa app, they had the option to add list of todos as a reminder. The first thing I did is to say something like - Alexa, add a reminder to get milk and eggs and paper. The app literally added a single item like this - milkANDeggsANDpaper.
Every once in a while I try the voice recognition by trying to speak normally to it. Normally, as is saying things like: "please set a reminder for five... umm... no I mean 6 o'clock".
Normal humans do this all the time, and if I can't do it speaking to it becomes incredibly frustrating to the point that I never want to do it again. I don't want to plan ahead what to say before I say it.
Granted, it's been a couple of years since I last tried so maybe they're better now.
This is also confusing what you search for vs. what the vendor thinks you will buy. Product catalog searches often intentionally return items outside your search parameters.
because google does understand that no and without are interchangeable. But, understandably, it does not correlate "shirt without stripes" as being the same thing as "solid-colored shirts." Why, because no one advertises or describes a solid-colored shirt as a shirt without stripes and no one searches that way. It's an irrelevant point, in my opinion.
my point was that this very web page on HN got to top of Google search in 42 minutes or less on this search term
'''
Shirt Without Stripes | Hacker Newsnews.ycombinator.com › item
42 mins ago - The point that the author is making, in a very understated way, is that all three companies have PR websites that breathlessly describe their ...
So the better search would be "DisplayPort adapter for a Mac."
As with including the word "stripes" in a search where you want to omit results with stripes, including the word "mini" is only causing unnecessary confusion. The adapter that works for a Mac Mini will also work for a Macbook, as an example.
Point being that the more narrowly-defined search does not require Google or Amazon to infer any meaning beyond what the object of the search is actually defined as.
A shirt lacking stripes would never be described or labeled as a "shirt without stripes."
In the absence of that actual description, you are asking google to assume what you mean.
It seems quite doable to handle "shirt without stripes" in the following way:
1) Gather all items labeled as "shirts" (among other labels)
2) Filter out any labels that includes "stripes"
A shirt doesn't have to be labeled "shirt without stripes" for this to work. A shirt labeled "shirt with stripes" or "striped shirt" would not match, and lots of other shirts (solid shirts, shirts with prints, whatever) would match just fine.
639 comments
[ 3.0 ms ] story [ 363 ms ] threadI'd expect to see some of the much touted "search intelligence", NLP, term inference, vector term analysis, and AI in action...
Of the first ten hits: Amazon: 7/10 aren't shirts (but all are stripeless) Google: All shirts, one is checked Bing: All shirts, all contain stripes in description.
Not exactly glowing for Amazon, clearly unsupported by Bing.
How is interpreting "shirts without stripes" as "I wish to see shirts that don't have any stripes" guessing?
I would venture to say that most people (meaning your use case is in the minority) who type "shirts without stripes" want to see results showing shirts without stripes not results "containing words "shirt", "stripes" and "without".
I think what is happening here is that you know how search engines work and so you are conditioned to expect them to do what they're doing.
Because I was looking for pages about the band called "Shirts without Stripes". Because I wanted pages with shirts that have stripes but where the page featured the word "without", because their shirts are without something else. Because I want to see striped shirts from the company called "Without". I don't want the machine to guess what I mean. It can never know.
When I’m looking up information, like an article, especially technical information, then yes I want the search engine to do as little interpreting as possible. That’s because any interpretation rules it uses won’t always be relevant so I’d rather have more control.
But for Amazon? I don’t see how someone can see the average user typing “shirts without stripes” and getting almost nothing but shirts with stripes, and going “yup, works as expected”.
i recently looked for dough scrapers. i wanted to see what's selling best and what's most rated. they are everywhere. in dessert & decoration, in utensils, in bakeware, and many other categories. i mean i get it...
it's not just search that's hard. categorization is also an issue here.
To be honest I wish they actually were keyword searches and that the machine doesn't try to be smarter than you. Many times when I carefully specify which keywords must appear on the page, it'll ignore parts of the query or add unrelated synonyms. Usually one can work around it with operators but it's tedious and doesn't work reliably.
At a point there was a TED talk explaining social networks were a solved problem now that facebook was dominant. Recycling was seen as solved problem until it wasn't etc.
I wonder how many actual "sovled problems" we have.
Today, entering in any tech-related query at all takes you to StackOverflow, end of story. Not only are SO answers quite often outdated (or even terrible advice in general), most of the time I'm not looking for a "here's how you do X", I'm looking for background information on a topic.
Most non-tech queries I put into google are even _more_ useless as the results tend to fall into these categories:
The dev/tech industry desperately needs a search engine that somehow prioritizes _quality_ content, not one-off answers, blogspam, and tweets.That is, learn to structure your query in a way that Google understands what you're trying to say. This used to be what yuo had to do, but now that Google tries to understand the intent of what you're trying to say and advertises as such, it's clearly a hack.
AND keyword LIKE '%searchStr%'
At some point in the future marketers will learn about AGI, and we'll have to make yet another term, maybe artificial general practical intelligence?
There is nothing on the word "intelligence" to imply it's not specialized.
I know what agi is; I just find the terms backward.
It's perfectly ok to prefer the qualifier to go the other way around, keep intelligence general and change the name of the specialized form. But that's just not how our language evolved.
But Siri is a general domain problem, which is really really hard. Siri set the expectation you can ask it anything, and it works terribly and for most questions just gives up and runs a web search.
If you are an e-commerce company though, that's a narrow enough domain, because you know that for most people, they're looking for products to buy or compare. It's not an unbounded Q&A service.
Lots of focus on a general purpose mono-model, but I think a collection of specialized subsystems is a better representation and would produce better results, faster.
You’d be surprised how effective NLP is for use when identifying query intent, and pulling out modifiers that should apply as metadata filters.
Weighted keyword search works a lot, but it fails hard for many long tail queries (especially in e-commerce and other attribute heavy domains).
IMO there really isn’t a good excuse for these firms to fail at queries like this. The query itself isn’t particularly difficult when using a decent NLP stack and following well known practices.
Google is already by far the most widely used search engine, so they don't really need to innovate or improve the search product very much in order to attract and retain users. Presumably capturing more advertising spending from the companies paying for ads is a bigger priority.
Microsoft under Satya Nadella has been all about enterprise and cloud, and I doubt Bing is a strategic priority any more, so it's not surprising that they wouldn't put a lot of resources into making it better.
Amazon is a little surprising. You'd think they'd have a lot to gain from making it easier for people to find what they're looking for. But maybe less than perfect search results are deliberate? Maybe it's like how supermarkets put basic items in the back of the store and high-margin impulse buys in the front - so you have to walk past chocolates and chips if you want to buy a carton of milk.
If Amazon is deliberately nerfing search results then maybe Google would stand to benefit from having better shopping-related results - people would get frustrated trying to find a shirt without stripes on Amazon and just use Google instead, letting Google profit from advertising in the process. But maybe people selling shirts aren't willing to pay much for ads, so there isn't much money for Google to make by getting better at finding specific types of shirts.
I dunno if any of these conjectures are anywhere near accurate, but it's interesting to think about.
If you go to Google's homepage and click the microphone at the end of the search input box you can search by speaking. All it does is convert to speech to text, but it implies you might be able to search in a more "natural language" way.
Google have a blog post from October last year with some more complex examples of where more sophisticated NLP helps https://www.blog.google/products/search/search-language-unde...
You can't assume that customers would type one thing or another - you need to gather lots of query log data and see what you find. You'd be surprised how much variation there is, but once you do have this data you can then find patterns to cover lots of (but not all) cases.
So NLP is totally a thing you want to have in search. Arguably, its the whole point of search as it exists now.
"Shirts"
"Polka dot shirts"
"Floral shirts"
"Wikipedia list of clothing patterns"
"Houndstooth shirts"
shirt -stripes
What about a search query for “Doctors without Borders“ or “Men Without Hats“?
Surely interpreting “without“ as the negative operator would ruin those searches.
Though I somewhat agree, I don't think that an average user even knows about existence of search operators in the first place, let alone being aware of this specific one and when to use it.
It would show pictures of shirts in pages that don't mention the word "stripes", whether the shirts have stripes on them or not.
In other words, it has little to do with what the article wants to show...
The author's intent exceeds both the capabilities and intended use case of search engines.
The query "shirts without stripes" if interpreted by human would require any search system to not only analyse the keywords and tags (of the products/images), but also the content, which is an infeasible task given its dynamic nature.
So the author wants: select all shirts where content analysis of returned images yields no stripes.
This is a context-sensitive image/product search based on arbitrary, dynamically created criteria and shows that user isn't aware of what the search functionality does as opposed to exposing weak "AI". [edit]To clarify: you cannot add all possible keywords/criteria in advance[/edit]
I've run into so many variations of this. You can search for something only to have the recommendation/related results embedded on whatever page to throw off your results one way or another.
I genuinely think that whatever standard HTML/XTHML is at ought to have, either as an attribute or a semantic tag, some kind of "related" or "recommended" ability to set that content apart. My cynical thought is that even if it were adopted, it would probably get abused in some fashion.
Edit: "stripes" not "stripped" ugh
EDIT: scrap that, I didn't mean Alexa, which is doing AI obviously, but the search engine of Amazon's retail website.
Anyway, NLP is hard and everyone sucks at it. Think about it: just building something that could work with any <N1> <preposition> <N2> or any other way to express the same requests would mean understanding the relationships of every possible combinations of N1 and N2. It means building a generalized world model that is quite different from simply applying ML to a narrow use case. Cracking that would more or less mean solving general AI which probably won't happen soon.
You're right the NLP is hard, but not everyone sucks at it.
Not actually true. ML is one area of study within the field of AI. Thanks to marketing departments and slightly shoddy journalism these two things are now casually treated as equivalents, but they're really not: ML is still very much a subset of AI.
If your "ML algorithm" doesn't understand straightforward language, how is it any better than a couple if-then statements?
Beyond that, I'm unsure how you think "<something> without <something>" is at all unusual or difficult to decipher.
Additionally, "shirt without stripes" is not the same as "solid color shirt"; as an example, take a look at:
https://www.google.com/search?q=tie+dye+shirt
If vendors would use the term "shirt without stripes" than it would match great, but they call it "plain shirt".
Google advertises using BERT natural language models
https://blog.google/products/search/search-language-understa...
> ... but they call it "plain shirt".
Or polka dotted :)
Whereas all these services seem to be processing the input in such a superficial way that they give the searcher results that aren't just inaccurate but are the opposite of what was asked for.
Lol what? These are words a toddler would understand.
How am I supposed to explicitly search for a shirt without stripes, then?
People still think we will have self driving cars "in two years" yet here we are talking about dumb shirts. AI winter is coming
The net result of that Google search, combined with the "Shirt Without Stripes" repo, leaves me even more unimpressed with the capabilities of our AI overlords.
The next few images contained Donald Trump, Terry Crews, Bill Gates and a French politician named Pierre Person.
After that it was actually quite a varied mix of men/women and color/white people.
I am still not very impressed with Google's search engine in this aspect, but it is not biased in the way you suggest.
At least it is not biased that way for me. As far as I am aware, and I might be completely wrong here, Google, in part, bases its search results on your prior search history and other stored profile information. It is entirely possible that your search results say more about your online profile than about Google engine :)
Well, she was the 2019 Time Person of the Year.
Likewise, Trump was the 2016 choice, and Crews and Gates have been featured as part of a group Person of the Year (“The Silence Breakers” and “The Good Samaritans” respectively).
- If I entered "person" I'd see a mix of images substantially similar to what I saw using google.co.uk up to and including Terry Crews, which was frankly a little weird, and otherwise mostly white
- If I entered "人", which Google Translate reliably informs me is Japanese for "person", I'd see a few white faces, but a substantial majority of Japanese people
So it seems possible that Google's trying to be smart in showing me images that reflect the ethnic makeup I might expect based on my language and location. I mean, it's doing a pretty imperfect job of it (men are overrepresented, for one) but viewed charitably it's possible that's what's going on.
Is the case for woke outrage against Google Image Search overstated? Possibly; possibly not. After these experiments I honestly don't feel like I have enough data to come to a conclusion either way, although it does seem like they may at least be trying to do a half decent job.
The TL;DR of it is that google crawls the internet for photos, associates those photos with text content pulled from the caption or from the surrounding page, and gives them a popularity score based on the popularity of the page/image. There are some cleverer bits trying to label objects in the images, but it's primarily a reflection of how frequently that image is accessed and how well the text content on the page matches your query. There's some additional localization, anti-spam, and freshness rating that influences the results too.
The majority of pages with "人" and a photo on it that has a machine labeled person image would be a photo of a japanese/chinese person, and if you're being localized to japan with a vpn, that would be even more true.
Google doesn't "know" what you're trying to search. It's a giant pattern matching game that slices and dices and rearranges text to find the closest match.
I'm not disputing that, and it certainly explains why it's "good enough" for somes search queries whilst being totally gimpy for others.
My understanding was that Google does prioritise what it's classified as local search results though, on the basis that they're likely to be more relevant.
"Person without stripes" shows several zebras, tigers, a horse painted like a zebra, and a bunch of people with stripes.
Interestingly, duckduckgo shows me, as second result, an albino tiger with, you guessed it, no stripes. The page title has "[...] with NO stripes [...]" in it, so I assume that helped the algo a bit.
EDIT: I also got the painted horse (it looks spray-painted, if you ask me) and I must admit it's quite funny to look at
Unless things have really changed, [doctor] will be mostly white men and [nurse] will be mostly white and Filipino women.
But don't blame the AI. The AI has no morality. It simply reflects and amplifies the morality of the data it was given.
And in this case the data is the entirety of human knowledge that Google knows about.
So really you can't blame anyone but society for having such deeply engrained biases.
The question to ask is does the programmer of the AI have a moral obligation to change the answer, and if so, guided by whose morality?
Key point right there. Unless Google is deliberately injecting racial and/or gender bias into their code, which seems extremely far fetched (to put it kindly), the real fault lies with us humans and what we choose to publish on the web.
I would contend that society is biased. There is no evidence that says men are better doctors than women, and in fact what little this has been studied says that women make better doctors than men (and is reflected in the more recent med school graduation classes which are majority women).
So it's a question of what you are asking for when you search for [doctor]. Are you asking for a statistical sampling or are you asking for a set of exemplars?
> So statistically, it would be correct to return mostly male doctors in an image search.
And that's exactly it. The AI has no morality. It's doing exactly what it should, and is amplifying our existing biases.
IMO, wrapping it in a concept like "morality" because the pictures have people in them just serves to excuse the problem and obscure its (otherwise obvious) solution.
Any sort of image search is going to tend to be biased toward stock photos, because those images are well labeled, and often created to match things people search for.
Nurses it's 34 women to 5 men. Proportions of skin tones are what I'd expect to see in a city in my country.
You can blame statistics for that. Beyond that, you can blame genetics for slightly skewing the gender ratios of certain fields and human social behavior to amplify this gap to an extreme degree.
(That's how I would do it if I wanted more accurate rather than more general results.)
There's not much diversity, assuming Terry Crews is from USA, then all the first viewport full of images are Western people; except Ms Thunberg they're all from USA AFAICT [I'm in UK].
The first non-Western person would be a Polish dude called Andrzej Person (the second Person called Person in my list after a USA dancer/actress), then Xi Jinping a few lines down. The population in my UK city is such that about 5/30 of my kids primary and secondary school, respectively, classmates have recent Asian (Indian/Pakistani) heritage. So, relative to our population, there are more black people, far fewer Indian-subcontinent no obviously local people.
Interesting for me is there are no boys. I see girls, men and women of various ages but no boys. 7 viewports down there's an anonymous boy in an image for "national short person day". The only other boys in the top 10 pages are [sexual and violent] crime victims.
The adjectives with thumbnails across the top are interesting too - beautiful, fake, anime, attractive, kawaii are women; short, skinny, obese, big [a hugely obese person on a scooter], cute, business are men.
The "White American family" and "White American mother" results are especially offensive to White Americans.
If I search for 'person' it's a mixed-race woman, then a white woman (Greta Thurnberg), then a white man.
The google image search you did -- did not provide incorrect answers, unlike the OP's
Most of the very top results seem to be of trump and greta thunberg.
I never figured out what kind of mistake could have led to that.
Relatedly, one time I picked up a prescription for a cat. The cat's name was listed as CatFirstName MyLastName. They had another (human) client with that same first and name. It turned out that on my previous visit they had "corrected" that client's record to indicate that he was a cat.
Is it surprising that very few of the result surprises me?
I did use images.google.se in order to tell google which country I wanted my bias from since that is the culture and demographics I am most familiar with. I also only looked at photos of a person and ignored emojis.
I have also seen here on HN links to websites that have captured screen shots of word association from google images and published them so you could click a word see the screen shot. They tend to follow the same line as above, but with some subtle differences, and I suspect that is the country culture being just a bit different to mine.
I just submitted all your searches to google.com from Australia, and the results were nothing like what you described; all the results were very diverse.
This is to be expected, as Google has been criticised for years for reinforcing stereotypes in image search results, and has gone to great effort to adjust the algorithms to reduce this effect.
But here, not that I think it will help: https://www.recompile.se/~belorn/happyvscriminal.png
First is happy person. Out of 20 we have 14 women, 4 guys, 2 children.
Second is criminal person. The contrast to the first image should be obvious enough that I don't need to type it.
If I type in "person" only I get the following persons in the first row in following order: Pierre Person (male) Greta Thunberg (female) Greta Thunberg (female) Unnamed man (male) Unnamed woman (female) Mark zuckerberg (male) Keanu Reeves (male) Greta Thunberg (female) Trump (male) Read Terry (male) Unnamed man (male) Greta Thunberg (female) Greta Thunberg (female) Unnamed woman (female) Unnamed woman (female)
Resulting in 8 pictures of females, 8 males, which I must say is very balanced (I don't care to take a screenshot, format and upload, so if you don't trust the result then don't).
Typing in doctor as someone suggested in a other thread I get in order (f=female, m=male): fffmffmmmmfmmfffmfmfmmmff
and Nurse: fffmffmfmmffmffmfffmffmffff
Interestingly the first 5 images have the same order of gender and are both primarily female, through doctor tend to equalize a bit more later while nurse tend to remain a bit more female dominated.
Your initial comment said "Happy person", women of color.
But your screenshot showed several white people, several men, and a diversity of ages. Yes, more women, which is probably reflective of the frequency of photos with that search term/description in stock photo libraries and articles/blog posts featuring them. No big deal.
You also said "Criminal person", Hispanic men
But the screenshot contains more photos of India's prime minister than it does of Hispanic men. In fact I can't see any obviously-Hispanic men, and the biggest category in that set seems to be white men (though some are ambiguous).
The doctor and nurse searches suggest Google is making some effort to de-bias the results against the stereotype.
To me the biggest takeaway is that image search results still aren't very good at all, for generic searches like this.
Indeed it's likely that they can't be, as it's so hard to discern the user's true intent (for something as broad as "happy person"), compared to something more specific like "roger federer" or "eiffel tower".
"Kind person" - pictures of men women, children, of all ages and colors.
"good person" - Mostly pictures of two hands holding. No clear bias towards women at all. If anything, more of the hands look "male".
"Bad person" - Nearly 100% cartoon characters
Absolutely ridiculous that you would take the time to write up such fake nonsense.
Following the stereotype content model theory I would likely get a pretty decent prediction of what kind of culture and group perspective produced the data. You could also rerun the experiment in different locations to see if it differ.
If you were unfamiliar with them and searched "widgets" to find out more and got widgets of a single colour and form, it would not be an unreasonable assumption that widgets are mostly (if not entirely) that shape and colour, especially if there was nothing to indicate that this was a subset of potential widgets.
It's not so much "demand for diversity" as it is "more accurate and correct representation".
Something of a corollary to Brooksian egg-manning: with an infinite number of possible searches, you can find at least one whose results do not exactly match the current demographics of the state from which you place the search.
Many interpreted this along tribal lines, but likely it is that there is constant tuning and lots of complex constraints.
[1] not to say that you implied the reason was racism, but often it is attributed to something along those lines
Nobody has solved the common sense knowledge problem yet. A solution for that would qualify as Artificial General Intelligence and pass the Turing Test.
But search engines have come a long way. I even suspect that when search engines place too much logical - or embedding relevance to stop words such as "without", that, on average, the relevant metrics would go down. It is not completely ignored as "shirt with stripes" surfaces more striped shirts than "shirt without stripes". "shirt -stripes" does what you want it to do.
Searching for "white family USA" shows a lot of interracial families. Here "white" is likely not ignored as much, and thus it surfaces pages with images where that word is explicitly mentioned, which is likely happening when describing race.
You can use Google to find Tori Amos when searching for "redhead female singer sings about rape". Bing surfaces porn sites. DDG surfaces lists (top 100 female singers) type results. The Wikipedia page that Google surfaces does not even contain the word "redhead", yet it falls back to list style results when removing "redhead" from your query, suggesting "redhead" and "Tori Amos" are close in their semantic space. That's impressive progress over 10-20 years back.
[1] https://en.wikipedia.org/wiki/Commonsense_knowledge_(artific...
There’s a nuanced argument that practitioners know how ML is so dependent on training data and accuracy tails off sharply, but that nuance tends to removed from anything selling to potential customers — which has not been a great way to keep them in my experience.
The theme of this talk was how they did a study that showed prepositions and articles do have meaning. A big deal was made out of the results.
I think things like this happens when people consider engineering approximations such as bag of words to be the truth over time.
There were gasps in the room and a kind of depressed acquiescence: geez, he might be right. And the pendulum indeed swung in that direction, hard, and the field has been overwhelmingly dominated by the statistical machine learning folks on the CS side of the field, while the linguists kind of quietly keep the flame alive in their corner.
But I thought then, and I still think now, that it really just was another swing of the pendulum (which has gone back and forth a few times since the birth of the field in the 1960s). Perhaps it's now time again for someone to ring up the linguists and let them apply their expertise again?
Sometimes I wonder how much my brain has changed to use search engines / how much of it is dedicated to effective googling. Makes me feel like a cyborg.
I guess because they leave it up to the advertiser to determine the negative match words and that seems to always have priority.
https://help.twitter.com/en/using-twitter/advanced-twitter-m...
Knowing how the machine will interpret humans is just as important to finding your results.
This assumes that AI wants truth. These three companies AI don’t necessarily want truth, they want revenue.
I think you're overestimating Google's sophistication.
If Google isn't under survival pressure to get better (and they aren't) the incentives aren't aligned for them to improve or even to not get worse every year.
If Google is failing first gradually then suddenly it might not even be within the institutional power to notice how bad it's become before it's too late.
Everything is for the best, in this best of all possible search engines -- the Candide fallacy.
https://www.nationalgeographic.com/animals/2019/09/zebra-pse...
This just suggests to me that real humans haven't issued that type of search query enough for the AI to know what to do with it. Which wouldn't be so big of a problem.
The joke from Zizek: https://www.youtube.com/watch?v=wmJVsaxoQSw
The state of the art in machine translation (from what I've read at least) is translating from language-A to a language-less "concept space" and then from there to language-B. Could that be done where the output language is something a search engine can use to find what you want correctly?
Given that pattern, I suspect we could see much better results in cases like this.
After that I facepalmed myself and turned it off.
Normal humans do this all the time, and if I can't do it speaking to it becomes incredibly frustrating to the point that I never want to do it again. I don't want to plan ahead what to say before I say it.
Granted, it's been a couple of years since I last tried so maybe they're better now.
'shirt no stripes'
on Google returned this web page at top of the organic results.
So at some point, searching for a shirt online will involve this conversation. Even more confusing.
(Although I expect my filter bubble will play a part in that)
''' Shirt Without Stripes | Hacker Newsnews.ycombinator.com › item 42 mins ago - The point that the author is making, in a very understated way, is that all three companies have PR websites that breathlessly describe their ...
'''
As with including the word "stripes" in a search where you want to omit results with stripes, including the word "mini" is only causing unnecessary confusion. The adapter that works for a Mac Mini will also work for a Macbook, as an example.
A shirt lacking stripes would never be described or labeled as a "shirt without stripes."
In the absence of that actual description, you are asking google to assume what you mean.
I would just never expect that to work very well.
1) Gather all items labeled as "shirts" (among other labels) 2) Filter out any labels that includes "stripes"
A shirt doesn't have to be labeled "shirt without stripes" for this to work. A shirt labeled "shirt with stripes" or "striped shirt" would not match, and lots of other shirts (solid shirts, shirts with prints, whatever) would match just fine.