Following on from the meteoric success of Wordle where you have to guess a word from five letters, there is now a whole collection of daily guessing games. Quordle, where you have to find four words at once, Nerdle, where you have to guess the sum, Worldle, where you have to find a country, and Heardle, where you have to identify a song. Absolute favorite is indicationwhere you have to guess a word and you are told how grammatically similar the word you entered is to the word you want to guess – so how close you are to the meaning.
Semantle is very challenging and frustrating, but you seem to get better at it slowly as you’ve been playing it for a while. The victory if I find the word relatively quickly (and within thirty turns I find it quickly) is much greater than in all the simpler games.
Every day there is a new problem that starts with something like: “The nearest word got a score of 66.53. The 10th closest word scored 47.90 and the thousandth word scored 27.59. Then you enter the words and you’ll be told how they score. I usually start with an animal, something edible, an adjective and a verb.”
Monday went like this: Cat: 4.47 (cold), Banana: 1.06 (cold), Large: 9.72 (cold), Life: 16.84 (cold). Anything not in the top A is ‘cool’, but since ‘life’ was a little less cool than the rest, I tried another verb: go:31.67 – then a nice green bar with 592/1000. So the word “go” had 592 similarities in meaning! After this stroke of luck, I proceeded with actions. Through “understanding,” “jumping,” “seeing,” and a lot of stuttering, I finally came up with the password: shine. But why did Sun get a score of only 27.53 when she tried it next? And how does the game on the ground calculate the number of two words with the same meaning?
The Dutch Semantle is based on the English Semantle of David Turner. Turner answers a lot of questions on his site. Behind his game lies modern science: techniques for converting human texts into data that a computer can use. One way is to convert any word into a vector arrow, a multidimensional arrow. Semantle uses word2vec, which as the name hints a bit, actually translates words into vectors. By looking at a lot of existing text – and doing all kinds of cool math, word2vec learns words that people use the same way.
This means that the word “sun” will not be very similar to “brightness” because you use it differently in a sentence. The closest words to “shine” are “shine” and “appear”. Here you immediately notice the subtlety that often makes the game difficult: “shine” has different meanings and both the synonyms of “radiate” and “appear” score well. The next word in the list of closest words is “hit” and that seems surprising until you think about how often people talk about the weather.
Sometimes the algorithm does something inexplicably weird and then sits behind the daily Semantle game. David Turner also has a solution for this: for 15 euros per month, you can buy the right to send him complaints via e-mail. Not only did he create my new favorite game but he also came up with my new favorite revenue model for online activities.
“Travel enthusiast. Alcohol lover. Friendly entrepreneur. Coffeeaholic. Award-winning writer.”