Contexto: Find the Secret Word
contexto.meSeems similar to https://semantle.com/ but backed by a different "similarity" calculation.
The "bar" visualization is also nicer to track how close/far you are.
I might as well take this opportunity to plug my clone of Semantle which uses a tSNE visualization. I've heard many people say this helps them visualize the chains of logic in their guesses.
Yours has the best interface I've seen so far, but the words still feel kind of obscure when I'm clicking the hint button. Ideally those should be limited to maybe the top 10,000 english words.
I think there's a really approachable game somewhere in this space, but it needs to implement something along the lines of an auto hinting system.
I imagine that for every guess, you could get a word or two that is similar (to help you understand what part of the word's context is important), and maybe words that are further away to help understand what isn't important?
My partner and I were consumed by this for a while. There's something very satisfying about hopping from cluster to cluster and recognizing the various meanings of a word.
Makes me wonder what 3D (or even 4D) version would be like.
At least a 2D visualization is at https://word2vec.xyz/
Your visualization makes the game a lot more fun.
I made this in Jan 2022. It's prior art, but I was learning Unity at the time so it is a little janky. In my oh so very humble opinion as its creator, a better guessing game because of the initial hint tuning.
Interesting. Took me way too long to realize that "close" is measured semantically, not lexically. A good strategy is to go up and down the ladder of abstraction, e.g. if "car" is close, try "engine" and "vehicle."
The semantic similarity is not always sensible, either. The underlying model needs some work.
(spoilers for today below)
.
.
.
.
.
.
.
These are some of my guesses for today, in order of similarity:
Ugh! I feel like I'm blindly stumbling around a maze of puns, or a Saturday NY Times crossword full of eye-rolling clues. It just feels like unsystematic guessing at what the semantic similarity model might contain.kilometer 3501 authority 3149 tomato 2960 exercise 2357 fruit 2008 healthy 1715 run 972 input 927 power 909 foot 598 meter 300 control 159Since I had the same issue while looking at the app for a minute, and your comment resolved it in 3 seconds, I would argue the explanation could be better. Thanks.
Also: Interesting game!
Few months back, I also tried to create a short puzzle game called Diffudle. The game was basically AI Art + Wordle. You were given an AI generated image and you have to guess the prompt for it.
Initially, I got some traction on the site and even some strangers started promoting it.
But I soon realized that there is very little marketing spend reward. The customers spend no money and ad revenue per user is abysmal. This means you can't use conventional channels and have to rely on people actually finding it fun and recommending it to others.
Although it was getting recurring traffic with zero promotion, it still couldn't reach a self sustaining level and I had to shut it down.
You can see Youtube video of it here: https://youtu.be/uLoMd0TkGKk
The point is: Getting the right puzzle formula is extremely hard and every puzzle will not grow like "Wordle".
>it still couldn't reach a self sustaining level
This site sounds like you could get away with free static hosting after pregenerating a few hundred puzzles locally.
Something is missing, I feel like.
The basic idea is sound, but it's painful to both start and end a game, and winning isn't very satisfying.
One of wordle's biggest flaws is "the opening". If you're going to have a daily puzzle, it shouldn't be exactly the same board to start every day. Games are so fast that this idea doesn't totally work there, but I think it would be better if they gave you a starting word every day.
The same rule applies here, I think. Your opening move set is basically fixed, and no one is really interested in learning the theory of that. Make the opening less tedious by starting with a couple of random words already played. This has two added benefits. The simplest is that it serves as a tutorial, in all likelihood, if you gave a list of prechosen words and their scores, and a textbox, people would already completely understand the rules of the game without any explanation at all. This is very valuable to a game like this, as it lowers friction.
The second benefit is more specific to this game, but it gives players a lot of hypotheses to test new words against from the outset. You want to reward play that deduces the next word not only from similarity to chosen close words, but also dissimilarity from chosen far words. In order to do that though, you have to choose a pretty large disparate set of far words. This hurts your score and is not particularly dependent on the days puzzle (ie you should pick about the same set every time). The game should probably do this for you, at the beginning.
Finally, you should reveal how many letters are in the target word. I don't want to just probe the model with synonyms. That's extremely tedious. give me enough information to make a plausible guess.
You could even consider a variant of the game where you're only allowed to guess words with the target number of letters, which might actually be mroe fun.
Hi you are describing my game, Lexicode.
https://greatfilter.itch.io/lexicode
I am currently remaking it as a web app rather than a unity app.
Got today’s word in 77 guesses. According to the game that’s not a good score, but I was impressed I was able to figure it out at all. Definitely much harder than Wordle.
I got it in 31. All it says for me is “Congrats!” What did it say for you?
I don't think it works very well. Trying not to give too much of a spoiler, today's Portuguese word is a noun. When I tried the "verb form" of that noun, I got a score of 207. This led me away from the correct word. I finally got it right in 30 guesses.
That's a tough problem when generating word embeddings. Bodies of text will rarely write "walking walk" or "the dancing dance" consequently leading naïve implementations to consider them unrelated.
I find redactle and its clones to be better games than the context ones because the similarity measures are inherently without any sense, and so even if you guess the correct word it often feels like pure luck.
In redactle winning by pure luck still happens, but at least in hindsight you can tell where you went astray, in contexto the answer often turns out to be "because you aren't a high dimensional word embedding trained by Open AI".
In french: https://cemantix.certitudes.org/
We used to play the french version a few months back. It's a nice game!
I got red=12, blue=20 and magenta=9283. Also, yellow=33 and orange=93, green=95 and pink=389 if that might help. The word was "flag" but good luck if you are from Nigeria, Pakistan, Brazil, Libya or Mexico.
What does the error "this word doesn't count" mean? Should the word be countable? Is it known (because "I don't know this word" also exists as an error) but not on the top N list?
I also got "this word doesn't count because it's too common" as well as the separate "this word doesn't count" - seem to have a few different exclusion reasons, some of which don't have an accompanying error detail.
I've gotten that on words that might've been slurs or something, but sometimes on words where I can't imagine that being the case.
Oh, right I can see how that could be the reason for the word I tried.
After getting good scores with <zvpebjnir naq bira, V gubhtug vg zvtug or cbg>. It's a strange world where we can't speak of <pbbxjner> to AIs anymore because they think you're speaking of a-drug-which-must-apparently-not-be-named
https://rot13.com/ (side fun fact: DeepL detects this text as Danish; Google and Bing just give up and default to English)
I find these to be less-than-ideal puzzles. For example, (no spoilers) today's puzzle answer is an abstract concept X, and two very common examples of X score both higher and lower (83 vs 7) than words that are much further away IMO (29).
It's interesting that it works. (179->150->13->1 for me. [I think it is more interesting to discuss paths instead of x and neighbors. A graph plotting all successful paths so far would be ++interesting to see. I wonder if a certain common final sub-path (like above) occur. And which one is it?])
Can we actually assert a 'natural semantic order'? I've always thought this was a fundamental aspect of our cognition, that we have 2 distinct ways of ordering: names and numbers. Outside of lexical ordering (which is ordering a representation not the things themselves), names can at best be ordered in a DAG. Pretty much half of computer science is converting names into numbers so we can compute over them.
Semantle (of which this appears to be a clone) uses cosine similarity with word2vec from my understanding and it felt much more intuitive.
I suspect like Wordle, or crosswords for that matter, it will take a certain amount of practice to get the hang of it. But! 43 guesses and I feel good about it.
44 here
I think i lucked out by landing a 'green' word i.e., a close word as first guess today
I played semantle for a few days and it used ri take hundreds of guesses sometimes
so not sure if we can do much better than 10-20 guesses
some word -> a high score
another word with a meaning close to the previous one -> this word doesn't count
OK, thanks for making this but I don't why I can't say the_word_which_shall_not_be_named - it is a perfectly valid word after all even if some feel offended by the mere mentioning of it.
First couple of words it converted to normalized forms, which I guess was intended to make the game easier? The third word I tried ("restlessnesses") it just said it doesn't know at all.
It's weird that Celsius, Kelvin and Fahrenheit were so far apart.
It's because it's not based on a simple measure of the similarity of meaning of the words. It's based on how often the words are used next to each other, which means that since celsius is used more often with the word temperature in the corpus, it will be closer to the word temperature.
Try game #160 (25/02).
Input as guesses:
digit
number
car
compare
Once you got the solution, please explain to me how it works again.
I don't know what game #160 is supposed to teach me, lol. I do admit that my explanation isn't right. The game's measure of similarity is based on context in which words are used, not necessarily their meaning. Some people probably argue that context is what gives meaning to words. I think it's not just context.
What I meant was that the "nearness" is somehow "random", I don't doubt that the results come from a clever analysis of a zillion documents or websites, still the final result makes little sense.
Your explanation is likely to be accurate in the case of celsius, but it seemingly doesn't fit on this other game/answer.
I don't want to spoil the answer to that game, but to me it is hard to believe that both "car" and "compare" are nearer than "digit" or "number" to the answer if it is based on number of occurences in context.
The way I explained it to a friend, is that "bastard" is closer to "ol" and "dirty" than to "illegitimate". Because it appears near those words more often.
This was fun! I wish it would give me two words as a hint, one in the middle of the stack and one just a bit north of it. Something to get me running.
I played contexto.me #168 and got it in 29 guesses and 6 tips.
13 8 14
When I select give up, I dont see an option to conrifm when it asks me if I really want to.. im using my phone to play.
I had to turn Safari content blocking off to get the confirmation buttons to appear if that helps.
Strange, I did get a confirm on my android on Firefox
The first word I tried was: "fish" it got 44. Nearly getting there...
Bonus points if you can do some prompt engineering to instruct ChatGPT how to play the game.
I just did a random puzzle, got #12, and got the word in 5 guesses.
Pure luck.
my word was "temperature" and my "voltage" attempt gave a distance of 46, so I was biased and finally gave up. I will try again another day
Neat! Must just be compring the distance of embeddings?
I got #2 as Humidity, #22 as Humid, Hint gave me #3 as Celcius, so I guessed Farenheit which was #484. Give up, answer was Temperature. Really? Celcius #3 and Farneheit #484? smh.
Just curious, did you spell it Fahrenheit or Farenheit?
Yeah, apparently "Farenheit" is a valid word for the game, but it's a typo. Celsius and Fahrenheit are actually very close.
68 guesses, 1 tip. green: 16, yellow: 9, red: 44
I played contexto.me #168 and got it in 6 guesses.
3 1 2
My best guess was Water too today. But Temperature as a solution was a way far imo. Can you align the similarity index somehow? Does it evolve? Is it GPT?
For non-historic games, it might be good to use https://rot13.com to not spoil others
It’s a little janky. My word was “temperature”. I got to “water” (no. 29) and then got stuck. Words leading up to it were rain, cloud, storm. Hard to take the leap from there to “temperature”
You might want to remove the word for spoiler's sake. It is the same for everyone, like Wordle
Keeping it generic I started with some identical words. It took me a little bit to realize I was getting some conflicting things (i.e. opposites that were similar) so needed to change the kind of word I was looking for. Wouldn't really call it janky because of that though, trying to figure out what could be related to the words so far is the point not a side effect. Looking at the final closest words I might have been upset having put in something like #4 and getting lost though.
I had it easier, won on 9 guesses:
temperature 1
heat 5
hot 83
light 292
sun 496
burn 1066
car 2663
girlfriend 16293
blacksmith 20611