Zooming out, the language field breaks into several subfields:
- A large group of Chomsky followers in academia are all about logical rules but very little in the way of algorithmic applicability, or even interest in such.
- A large and well-funded group of ML practitioners, with a lot of algorithmic applicability, but arguably very shallow model of the language fails in cases like attribution. Neural networks might yet show improvement, but apparently didn't in this case.
- A small and poorly funded group of "comp ling", attempting to create formalisms (e.g. HPSG) that are still machine-verifiable, and even generative. My girlfriends is doing PhD in this area, in particular dealing with modeling WH questions, so I get some glimpse into it; it's a pity the field is not seeing more interest (and funding).
https://www.google.com/search?q=%22Shirt+without+Stripes%22&
Yes, the English grammatical rules make it unambiguous where it belongs. This is solvable.
Couldn't you just parse the sentence into a dependency tree and look at the relationships to figure that out? CoreNLP got both of your examples right (try it at http://nlp.stanford.edu:8080/corenlp/process, can't link the result directly).
Well, one could argue, that it belongs exactly where anyone entering the query put it. Before "stripes".
The problem is often, that search engines try to be too clever, while not offering any kind of switch "exactly those words in this order" and that is just a bad user interface.
If it just disregards the word without, well, that's pretty bad.
I will not be surprised millions of dollars are being lost because of this substandard query result per year.
“Shirt -stripes” is unambiguous to a system, yet the first result on Amazon(.ca) is a striped shirt, and the 3rd is sweatpants.
I mean, context is key, right? You're on Amazon and your first search term is "shirts". Unless their is a band called "shirts without stripes", the user wants shirts. The rest of the query is probably some filter of that product. You know shirts sometimes have stripes. It's not a one-size-fits-all algorithm, but it's simple enough that the user should end up with the results they wanted.
> "no evidence of cancer" and "evidence of no cancer" are very different things.
Why is it not as simple "no belongs to the word it precedes" ? like unary operator, ! (not), in typical computer languages.
Does she want ice cream? Answer: No, she doesn't. I added a not, so she's reversing the answer as Japanese people do.
The number of times I've been dumbstruck by this is larger than I'd like to admit, and I'm a coder.
There have been some lengthy discussions on HN about vertical search and how Google doesn't always buy up a small company; they litigate.
I'd be curious to see how many sentences with attribution problems actually have other structural issues. If I want to write clearly and without ambiguity, I rewrite sentences that have these problems. Why wouldn't I do the same for search queries?
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.
Edit: "stripes" not "stripped" ugh
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...
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.
Google has not yet discovered how to automate "is this a quality link?" evaluation or not, since they can't tell the difference between "an amateur who's put in 20 years and just writes haphazardly" and "an SEO professional who uses Markov-generated text to juice links". They have started to select "human-curated" sources of knowledge to promote above search results, which has resulted in various instances of e.g. a political party's search results showing a parody image. They simply cannot evaluate trust without the data they initially harvested to make their billions, and without curation their algorithm will continue to fail.
Google is a dumbass nowadays, and regularly ignores half your search terms to present you with absolutely irrelevant results, that have gotten lots of visits in the past.
Today, it will silently guess at what I want, and rewrite the query. If they have indexed pages that contain the words I put in, but don't meet their freshness/recency/goodness criteria, they will return OTHER pages with content that contains vaguely related words. "Oh, he couldn't have meant that, it's from 6 months ago, and it's niche!"
They'll even show this off by bolding the words I didn't want to search for.
So, if I'm looking for something that isn't popular -- duckduckgo it is. It doesn't do this kind of rewriting, so my queries still work.
This is a case where, while it makes sense to say the sentence, it's not a common use of language, and at the end of the day, the search engine will find what's written down, it's not a natural language processor yet (despite any marketing).
Shirt stores don't advertise "Shirts without stripes - 20% off", they describe them as "Solid shirts" or "Plain shirts". Men's fashion blogs talk about picking "solid shirts" or "plain shirts" for a particular look. If I walked into a clothing store and asked for "shirts without stripes", the sales person would most likely laugh and say "er, you mean you want plain shirts?".
Plain shirts/solid shorts are the most common way to refer to these, and people seem to be searching this way:
https://trends.google.com/trends/explore?date=all&q=solid%20...
Regarding moving towards natural language processing - the "without" part is not as important as knowing the context.
My kids will ask me to get from the bakery things like "the round bread with a hole and seeds", which I know means "sesame bagel", or "the sticky bread", which means "cinnamon twists" - which I understand because I know the context. Sometimes they say "I want the red thingy", and I need to ask a bunch of questions to eventually get at what they want (sometimes it's a red sweater, sometimes it's cranberry juice).
Unless Google starts asking questions back, I don't think there is any way it can give you what you want right away.
Searching "pants" only shows me "trousers", that's a big fail for Google IMO, I'm accessing google.co.uk.
The joke from Zizek: https://www.youtube.com/watch?v=wmJVsaxoQSw
To extend on that, you can think of the human brain as just another (powerful) statistical model.
Doing just that for 10 years, beating hand-coded systems: https://www-nlp.stanford.edu/pubs/SocherLinNgManning_ICML201... [pdf]
> I would guess that most modern NNs from the NLP area (Transformer or LSTM) would be able to correctly differentiate the meaning.
Yes. See demos like: https://demo.allennlp.org/constituency-parsing/MTczNjYyNA== and https://demo.allennlp.org/dependency-parsing/MTczNjYyNg==
> I think there is no fancy NN (yet) behind Google search,
During the deep learning boom, Google made a huge push towards NN-based NLP. SEO's and their PR calls their efforts collectively RankBrain: https://en.wikipedia.org/wiki/RankBrain
I think we are on the cusp of combining symbolical/logical operations over the vectors produced by Neural Networks (or at least, major effort there). Could be by neatly tying up all these different NN-based NLP modules (parsing, semantic distance, knowledge bases, ...) with another set of decision layers stacked on top.
Chomsky: Statistical analysis of snowflakes falling outside the window may predict the next snowflake, but it will do very little for weather prediction, and nothing for climate analysis.
Norvig: Give us enough data and we will get close enough for all practical purposes.
It's easy to say, isn't it? Unfortunately, sticking the word "just" in there doesn't affect the difficulty. I do it all the time, too.
That said, "meaning" is not statistical.
- Shirt Without Stripes: shirts where the description contains both "without" and "stripes". Example: a shirt without collar, with stripes.
- "Shirt Without Stripes": a mess, with and without stripes, suggesting an unusual search query. In fact, the linked article site is the first result in web search.
- Stripeless shirt: sexy women in strapless shirts
- "stripeless shirt": pictures of Invader Zim...
- "stripeless" shirt: mostly shirts without stripes, but there are some shirts with stripes that are described as stripeless...
The last one may give us a hint at the problem. If you have to mention a shirt is without stipes, you are probably comparing is to a shirt with stripes. For example imagine a forum, some guy is posting a picture of a shirt with stripes, I can expect some people to ask questions like "do they sell this shirt without stripes"? Or maybe the seller himself may have a something like "shirt without stripes available here (link)" in the description. So the search engines tie "shirt without stripes" to pictures of shirts with stripes.
I remember an incident where searching for "jew" on Google led to antisemitic websites. The reason was simply that that exact word was rarely used in other contexts. Mainstream and Jewish source tend to use the words "jews" and "jewish" but not "jew". And because Google doesn't look at the dictionary meanings of words but rather what people use them for, you get issues like that.
I had a similar problem when I was trying to convince a friend that homeopathy was a complete and utter fraud with absolutely no basis in reality. She was convinced that the internet's overwhelming consensus was that homeopathy was valuable and regular doctors were control-freaks who make things up when they don't know the answers.
To prove her point, she did an internet search for allopathic medicine and showed me how the majority of the results were negative.
https://en.wikipedia.org/wiki/Allopathic_medicine
Just a humorous anecdote, not trying to start any conversations about the relative value of different medical paradigms.
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 think you're overestimating Google's sophistication.
Knowing how the machine will interpret humans is just as important to finding your results.
I guess because they leave it up to the advertiser to determine the negative match words and that seems to always have priority.
Everything is for the best, in this best of all possible search engines -- the Candide fallacy.
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.
This assumes that AI wants truth. These three companies AI don’t necessarily want truth, they want revenue.
"Humans usually don't intuitively understand the word 'no'. Please imagine a non-pink elephant."
It would be great if I could add negative keywords to a website, or mark text as "don't index" or "index with a negative weight". But probably, people would game this in ways I can't imagine.
There is probably a clever ML solution for this, like having meaning-vectors for distinct ideas, and pushing pages that are close to one meaning away from the other meaning. Classification is easy if you have a keywords like "painting" and "catholic", but if it is "virgin" or "prayer" then it could be either meaning, so there is never a bullet-proof solution.
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?
Likewise, Google says I should log into their website for personalized search results, but after years of always clicking on Python 3 results over Python 2.7 results, it never learned to show me the correct result.
Eventually I realized that personalized recommendations are more or less just a thin cover for collecting vast amounts of data with no benefit to the consumer. I believe we have the technology to do better, but we don't use it. In fact, we seem to be using it less and less.
As humans we know immediately that the search is for documents about shirts where stripes are not present. But the term 'without' doesn't make it through to the term compositor step which is feeding terms in a binary relationship. We might make such a relationship as
Q = "shirt" AND NOT "stripes"
You could onebox it (the Google term for a search short circuit path that recognizes the query pattern and some some specific action, for example calculations are a onebox) and then you get a box of shirts with no stripes and an bunch of query results with.
You can n-gram it, by ranking the without-stripes n-gram higher than the individual terms, but that doesn't help all that much because the English language documents don't call them "shirts without stripes", generally they are referred to as "plain shirts" or "solid shirts" (plain-shirt(s) and solid-shirt(s) respectively). But you might do okay punning without-stripes => plain or to solid.
From a query perspective you get better accuracy with the query "shirts -stripes". This algorithmic query uses unary minus to indicate a term that should not be on the document but it isn't very friendly to non-engineer searchers.
Finally you can build a punning database, which is often done with misspellings like "britney spears" (ok so I'm dating my tenure with that :-)) which takes construction terms like "without", "with", "except", "exactly" and creates an algorithmic query that is most like the original by simple substitution. This would map "<term> without <term>" => "<term> -<term>". The risk there is that "doctors without borders" might not return the organization on the first page (compare results from "doctors without borders" and "doctors -borders", ouch!)
When people get sucked into search it is this kind of problem that they spend a lot of time and debate on :-)
It's a completely artificial construct. Simply the fact that this hacker-news entry is the #1 search result shows that real human people do not perform this search in significant quantity. But we can quantify that with data to backup the assumption [1][2]. When people want to buy a shirt without stripes, they do not describe the shirt by what it doesn't have.
In fact, it's trivial to cherry pick a random selection of words that on the face of it sounds like something a human might search for, but it turns out never occurs in practice. Add to that the fact that the term is being searched without quotes [3], which results in the negation not actually being attached to anything.
Do you go to a store to buy it along with your Pants Without Suspenders, Socks Without Animal Print, and other items defined purely by what they don't have?
[1] https://trends.google.com/trends/explore?geo=US&q=%22white%2... [2] https://trends.google.com/trends/explore?geo=US&q=%22plain%2... [3] https://trends.google.com/trends/explore?geo=US&q=plain%20sh...
Likewise, here, I would search for solid-colored shirts.
And these services are limited to the content/terminology utilized by the cataloged sites/products.
If I am selling a "black shirt" or a "solid black shirt," it is not google's job to catalog it as a "shirt without stripes," unless I advertise it as a "black shirt without stripes."
I would use natural language to test a services' NLP ability.
Unfortunately people don't search for "solid shirts". At best they search for "plain shirts", but there's a lot of taste to clothing that means people often do want a shirt without stripes, but are open to patterned/plain.
I think searching "shirts without stripes" is very legitimate in fashion.
I say this having built a clothing search function for the company I work for, and one that does not support this sort of query.
There's too many products nowadays to be manually attributed (e.g. pattern=stripes), making it hard return good results even with entity resolution for queries. We train classifiers to categorize products, including what something is not, using their images and descriptions.
I'm guessing one of those reflections looks like a turtle? Or maybe a pattern on the floor, wall, or rug?
Although there are examples where I'm unsure if the AI is dumber than my 4yo or smarter than me. This is a result for "truck": https://i.imgur.com/JcgXZAG.jpg
Even (especially?) my 4yo knows those are Brio trains, not trucks. However, trains have components called trucks! https://en.wikipedia.org/wiki/Steam_locomotive_components I'm unsure whether or not any of the wheel assemblies on these toy trains are considered trucks, so either the AI is extremely smart or slightly dumber than a 4yo.
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.
Joking aside, it doesn't surprise me that this isn't being picked up — aren't most of these AI teams more R&D than actual public-facing? Maybe I'm just cynical though.
'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)
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.
Larry: "I saw a red Lamborghini in the parking lot!"
Most people will assume Lisa is driving a red Lamborghini and back from Vacation, meanwhile, all the bots are searching for Lamborghini vacations and trying to figure out what's going on in the conversation.
"shirts -stripes" results: https://www.amazon.com/s?k=shirts+-stripes&ref=nb_sb_noss_2
So basically the AI doesn't convert "without x" to "-x" even though the basic capability needed is there. This is why AI is a hard problem, especially when it meets the real world.
It's 2020 and we're still quibbling about the terminology used in SQL, what did we expect?
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.
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 (!)
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.
"birds without flight"
"cars without wheels"
"cats without tails"
"dogs without hair"
"intersections without lights"
"poems without rhyme"
"shirts without collars" (also "sleeves", "shoulders", "buttons", "logos", "pockets", and more)
What if you walked into a store and asked an associate for a shirt without stripes? What would you get?
Probably some further questions for clarification. What about checked shirts? Floral prints? Plaid? Do you want no pattern at all? T-shirt? Polo shirt? Dress shirt?
Granted, the AI results are particularly bad because they give you the one thing that you specifically didn't ask for, but that's also the only information you provided. Defining a query in terms of what you don't instead of what you do isn't going to go well.
What if you went to google and said "Show me all the webpages that aren't about elephants"? Sure, you'd get something, but would it be anything useful?
Google has gotten better, it's just HNer expectations that have changed as they expect more and more magic.
For example, the subtitle on the repo is "Stupid AI" when this query has never worked in these search engines, and it won't anytime soon.
You'd think the technical HN crowd would be more advanced than to make the same mistakes that (they complain that) stakeholders/users/gamers make when they mistakenly think everything is much easier than it actually is. Things aren't "stupid" just because they can't yet read your mind.
I would expect there to be an e-commerce site or blog post somewhere containing a page with the exact title "shirts without stripes" and I'd expect it to be the first match.
This thread is an excellent example. The author of the linked page didn't have the decency to actually make a substantive point, instead sharing three screenshots and posting the link here, chumming the HN waters with the kind of stuff that brings in the sharks from far and wide.
Bashing on big cos: Check
Vague pronouncements about AI: Check
Generic side-swipes about 'ad revenue': Check
This is why a coherent thesis is required to even initiate a proper discussion, because in the absence of that it invariably devolves to lowest-common-denominator shit-flinging.
Negations sidestep almost all of the algorithms that try to provide an improved result set, and fall through to pure text relevancy. So try searching on amazon for shirt, then search for: shirt -xkxkxkxk. Since xkxkxkxk doesn't match any documents, the negation should have no effect, but it does, the effect it has is to sidestep all the fancy relevancy work and hardcoded query rewrite rules, domcat rules, demand and sales/impression statistics etcetc, and give you basically awful search results. You don't even get shirts.
Anyone with a programming background knows there is an art to forming useful search queries--it is an acquired skill. I'd personally much rather the engine bring back predictable results given mundane rules and keywords than attempt to understand sentences using an opaque method of understanding.
That said, this seems like an obvious place for improvement where both groups can be made happy.
Given exact query to human, they create environment thus context themselves.
It may also depend on whom you are asking to. For example, myself, entering this site to find out news about software & tech. Also since 'Stripe' is a company name, I assumed link will get the list of shirt shops who do not accept Stripe as a payment method/provider. (Thus some kind of protest related thing)
I literally thought about that yesterday and did not see the page thinking "That's too much for tonight".
Now seeing topic is somewhat very different.
The former returns lots of mixed race couples, mostly not white couples. However the latter returns black couples.
What is going on here? Similar phenomenon perhaps?
What is the expected result, can we agree?
It is a shirt with anything except stripes.
And it's only that smart assistant that automates coping with the deficiencies of a one-size-fits-all central solution, finding me shirts with no stripes by using a rather dumb search engine. (Or "a pizza I would like", etc.)
https://medium.com/pinterest-engineering/pinsage-a-new-graph...
"Don't think of a cow !"
What did you just think of ? A cow, of cowrse.
If you want a shirt w/o stripes, just google "plain shirt" or "dress shirt -stripes.
There are several reasons for this, including the following:
1) Natural language understanding for search has gotten a lot better, but it is still not as robust as keyword matching. The upside of delighting some users with natural language understanding doesn't yet justify the downside of making the experience worse for everyone else.
2) Most users today don't use natural language search queries. That is surely a chicken-and-egg problem: perhaps users would love to use natural language search if it worked as well or better than keyword search. But that's where we are today. So, until there's a breakthrough, most search engine developers see more incremental gain from optimizing some form of keyword search than from trying to support natural language search.
3) Even if the search engine understands the search query perfectly, it still has to match that interpretation against the documentation representation. In general, it's a lot easier to understand a query like "shirt with stripes" than to reliably know which of the shirts in the catalog do or don't have stripes. No one has perfectly clean, complete, or consistent data. We need not just query understanding, but item understanding too.
4) Negation is especially hard. A search index tends to focus on including accurate content rather than exhaustive content. That makes it impossible to distinguish negation from not knowing. It's the classic problem of absence of evidence is not being evidence of absence. This is also a problem for keyword and boolean search -- negating a word generally won't negate synonyms or other variations of that word.
5) The people maintaining search indexes and searchers co-evolve to address -- or at least work around -- many of these issues. For example, most shoppers don't search for a "dress without sleeves"; they search for a "sleeveless dress". Everyone is motivated to drive towards a shared vocabulary, and that at least addresses the common cases.
None of this is to say that we shouldn't be striving to improve the way people and search engines communicate. But I'm not convinced that an example like this one sheds much light on the problem.
If you're curious to learn more about query understanding, I suggest you check out https://queryunderstanding.com/introduction-c98740502103
Just my theory.
Nobody would describe a plain shirt as a shirt without stripes unless it’s within that context.
The result, of course, show shirt with some kind of stripe, albeit not prominent like the English one.
Google can do this now, for example in a prototype. The tough thing is to get it to consumer-grade quality without messing up other searches. The QA process is utterly brutal because one weird search can be a scandal.
Or does input need to have basic filters applied before handing to ML? "without X" or "no X" = "-X"? Can be foiled with "shirt without having stripes".
But make a script that scrapes the top X results for these sites. Get your own AI / humans to rate it.
Make it competitive for these large sites <==> give them an incentive.
The real question is “shirts without stripes” really a query people enter? Or representative of a real pattern in the data?
Citation needed.
As far as my personal observations go, Google is NOT optimized for long tail at all. It is always trying to return most popular results from cache of most popular results. Once the cache is exhausted, Google starts to return completely irrelevant trash (anything after first two pages of search is pure spam and meaningless keyword soup).
If you try to look up some obscure keyword and find nothing, try again after couple of months. There is a very high likehood, that you will see dozens of "new" results — most of them being from several years old pages. Perhaps, the actual long-tail searches still happen somewhere in background, but you are not going to see their output right away — instead you need to wait until they get committed to the nearby cache.
Another alarming change, that happened relatively recently (4-5 years ago), is tendency to increase number of results at expense of match precision. A long time ago Google actually returned exact results when you quoted search phrase. Then they started to ignore quotes. Then they started to ignore some of search terms, if doing so results in greater number of results. Finally, Google gained horrifying ability to ignore MOST of search terms. OP's example probably has the same cause — Google's NLP knows the meaning of word "without". But Alphabet Inc. can't afford to hose all those websites, that use AdWords to sell you STRIPED SHIRTS. This would mean a loss of money! THE LOSS OF MONEY!!!
And since most search applications are basically just finding you the results with the most keyword matches, with a little bit of extra magic thrown in, the above is basically what you see.
These queries are basically the equivalent of optical illusions in cognitive psychology when studying the visual system -- seeing how the systems break tells you a lot about how they work.
Have you ever seen shirt dresses in your dress shirt queries, or vice versa? The search application isn't caring enough about bigrams and compound words.
Have you ever seen bowls in your bowling queries or fish in your fishing queries? The search application is over-stemming.
Natural language search is a real pain on any general purpose search application, particularly ones that have to deal with titles. The obvious simple fix to this query is to rewrite [x without y] to [x -y], but then when someone goes to search for [a day without rain] or [a year without summer], you are going to totally break those queries.
So it's not such a big deal that negation doesn't work.
Also, "shirts -stripes" does seem to work in both Amazon and Google. Or at least, I see no striped shirts.
> in particular, it shows clear insensitivity to the contextual impacts of negation.
As in, "X without Y" sounds like a common enough use case to have it's own little parser branch in places as big as Google or Amazon
So it's essentially the same input, and essentially the same expected output, but there must be quite a knot between understanding the word "without" and literally just using the - operator.
https://www.amazon.ca/s?k=shirt+without+stripes&ref=nb_sb_no...
shirt -stripes
Fishbone - Drunk Skitzo https://youtu.be/SaPGH4Yd_zc?t=231
(Apologies for the snarky low-content flip reply.)
You have to know to search for "solid colored shirt", but when you can't think of this variation of search, or maybe there isn't one, exclusion is your only option, and it's broken.
On Amazon's side of things I would also include the obnoxious "Hey you just bought a pair of sneakers so now I will change all your recommendations to sneakers".
If it's meaningful for some reason, then it works:
https://www.google.com/search?q=woman+without+makeup&tbm=isc...
If it's an user error (like a dumb query) it fails and it shouldn't be a surprise:
But there probably aren’t many images labelled as stripeless.
I’m not sure why BERT doesn’t try shirt -stripes.
=> "shirt -stripes" works pretty well on google at least
We are in a funny place with UIs.
"shirt without sleeves"
That something that someone may actually search for. (At least the guys at my gym would!) And Amazon gets it mostly wrong.
Search results could be better? Sure.
Can we find adversarial examples? Almost always.
shirt -stripes
Achilles: Who could deny it?
Tortoise: Good. Likewise, “Cast-iron sinks” is a valid utterance, isn’t it?
Achilles: Indubitably.
Tortoise: Then, putting them together, we get “Politicians lie in cast iron sinks”. Now that’s not the case, is it?
---- Douglas Hofstadter, Gödel, Escher, Bach: An Eternal Golden Braid. Basic Books, 1979
"Vaguely similar to a joke from _the movie_ Ninotchka that _the Slovenian philosopher_ Zizek often uses...."
Give people context. Don't assume people know what you know.
If Google wants to group words by semantics, they should have a semantical grouping operator. For example "shirts (without stripes)". What if I am looking for a song text with these exact words in random positions?
If what author wants was implemented, it would make my experience with Google even worse, unless it could think for me also. But then why would it need me in the first place?