Semantle Method
Semantle scores each guess by cosine similarity to the secret word using a word2vec model. Unlike Contexto's rank, you see a raw similarity number and, when you get close, a "getting warm" temperature. The skill is following the gradient of meaning uphill.
What the similarity number means
The score is how aligned your guess is with the secret word in vector space. A high positive number means closely related meaning; near zero means unrelated. The scale is not linear, so early gains feel small and the last few points are the hard part.
The word2vec model learned meaning from context, so it groups words that appear in similar sentences. DOCTOR and NURSE score close; DOCTOR and GRAVEL do not.
Following the gradient
Climb toward the answer
- Open with a spread of unrelated common words to find any thread.
- When one scores higher, guess its synonyms and associated words.
- Watch the number: rising means warmer, so keep pushing that direction.
- Near the top, test word forms and very close synonyms one at a time.
Escaping a dead end
The daily Semantle answer page confirms the word, but the game is most satisfying when you follow the similarity gradient patiently from a cold start to the solve.
A high score can still mislead
Words can be close in the model for surprising reasons. If synonyms stop helping, jump to a related but different concept to see if the gradient points elsewhere.
What the score measures
Semantle compares your guess with the secret word using a word-embedding model, which represents every word as a position in a high-dimensional meaning space. The score you see is how closely aligned those two positions are.
The endpoints of the scale are definitional rather than arbitrary. The secret word compared with itself sits at the top of the scale, and a word with no relationship to the target sits at the bottom, near zero. Everything you guess lands somewhere in between.
Unlike Contexto, Semantle reports the score itself rather than a rank among all candidates. That is the more useful of the two readings for following a gradient, because the score of a single guess tells you something on its own — a rank only means something in comparison.
The two ends of the scale are fixed by definition. Every real guess lands between them, and the movement between guesses is the useful signal.
Reading a score that will not move
The frustrating part of Semantle is not the low scores, it is the scores that stay put. A guess that scores almost exactly the same as the previous one is telling you something specific: you have found the right area of meaning and the wrong word within it.
That is a different problem from being in the wrong area, and it needs a different response. Changing topic when the score is stuck wastes the position you have built; changing the sense of the word when the score is stuck is what breaks the plateau.
The table below maps each pattern of movement to the change it calls for. Read it as a decision table applied after every guess, not as a description of the scale.
| Reading | What to change next |
|---|---|
| Score rises | You are moving the right way. Stay in this area of meaning and refine the word rather than the topic. |
| Score falls | Wrong direction. Move back toward the previous guess and try a different neighbour instead of pushing further. |
| Score barely moves | Same concept, wrong sense. Change the part of speech or the specific meaning, not the subject. |
| Score close to zero | Unrelated domain. Abandon the current topic entirely and pick a different starting word. |
A decision table for the step after each guess. The direction of movement is more informative than the absolute value.
Following the gradient uphill
Semantle rewards a hill-climbing loop: keep the guess that scored better and vary one thing about it, then keep whichever variation improves the score again. Because the score is a number rather than a rank, each step is directly comparable with the last.
The steps below are that loop written out. The discipline that makes it work is varying one property at a time — a synonym swap, a change of register, a move from a general term to a specific one.
The trap to avoid is guessing a word because it looks like the target. The score does not read letters, so a near-miss spelling scores no better than an unrelated word of the same length.
-
Open with a common, high-frequency word
A frequent word usually has a well-populated neighbourhood, so its score gives you a usable starting direction rather than a flat zero.
-
Keep the better-scoring word as your new base
After two guesses, discard the worse one and treat the better one as the position to improve from. Do not average the two or split the difference.
-
Vary one property at a time
Change a single dimension — synonym, specificity, part of speech, register — and compare the new score with the base. One change at a time is what makes the score difference readable.
-
When the score stalls, change the sense
A plateau means the topic is right and the meaning is wrong. Look for a different sense of the same idea rather than a different idea.
-
Guess the answer once the score is very high
At the top of the scale the remaining candidates are near-synonyms. Name the most likely one instead of continuing to climb.
Semantle questions
How does Semantle measure closeness?
It uses cosine similarity between your guess and the secret word in a word2vec model. Higher numbers mean the meanings are more closely related.
How is Semantle different from Contexto?
Semantle shows a raw similarity score and a warmth indicator near the top. Contexto shows only a rank from 1 upward. Both reward following meaning rather than spelling.
Why does a Semantle guess that looks similar score low?
The score compares meaning, not spelling. A word that looks like the target but means something different sits far away in the embedding space, so it scores near the bottom of the scale no matter how similar the letters are.
What does it mean when two different guesses score almost the same?
That they occupy nearly the same place in meaning space. You have found the right topic and the wrong word within it, so the next guess should change the sense of the word rather than the subject.