Rhyme refinement against Corpus

Rhyme search answers one question: which words sound like the one you typed. Refine answers a second one: which of those words do people actually use.

Press Refine — the button at the top of the results panel, to the left of the two Copy All buttons — and every number on the list changes. Instead of how close each word sounds, you now see how often it turns up in real writing. Press it again and you are back to the rhymes and their Proximity scores; nothing is lost either way.

Refine is for the moment you have twenty rhymes on screen and half of them are words you have never heard anyone say out loud.

What "the corpus" is

A corpus is simply a very large pile of real texts, counted once in advance: books, articles, lyrics, everything. For each word in it we know how many times it appeared.

Refining looks up every word in your results list in that pile and reports back what it found. Nothing is guessed and nothing is scored — it is a lookup, so a word is common because people wrote it a lot, not because the app thinks well of it.

Each language has its own corpus, and the search uses the one for the language your song is set to.

The number on each row

The number is how many times the word appears in every million words of text.

• 1.00000 — about once in a million words. On this scale that is a common, everyday word.
• 0.05000 — about once in twenty million words. Recognisable, but not something you hear every day.
• 0.00100 — rare. Fine as a deliberate choice, easy to trip over by accident.
• A dash (—) — the corpus has never seen this word at all. Usually a name, a coinage or a very unusual form.

Hover a row and the app spells the same thing out: …occurrences per million.

Refining re-orders the list. The rows come back sorted by how common the words are — most common first — so the closest-sounding rhyme is no longer necessarily at the top. Press Refine again and the rhyme order comes back, along with the Proximity scores.

One thing the number does not do is pass a verdict: a beautiful line can be built on a word with a low number. Refine tells you how far off the beaten track a word is, and leaves the decision to you.

Refine needs you to be signed in, because the lookup happens on our server rather than in your browser. The rhyme search itself does not.

How the Quality slider works

While Refine is on, the Quality row takes the place of the Proximity row. It is a single dial with one job: how common a word has to be to stay on screen. It starts at 25%, and the app remembers wherever you leave it.

• At 0% nothing is hidden — every result stays, including the ones with a dash, whose count the corpus does not have.
• Drag it right and the bar rises. The rarest words drop off the list first; the words with a dash are the very first to go, the moment you move off 0%.
• At 100% only genuinely common words survive — roughly one appearance per million words or better.
• The default 25% is a low bar on purpose: it clears out the dashes and the true curiosities and leaves nearly everything else.

Two things are worth knowing about the feel of it. The scale is not a straight line — every tenth of the way you drag it, the app asks for a word getting on for twice as common as before, so most of the real cutting happens in the right-hand half.

And the bar is fixed, not relative to your list: the slider does not keep "the top half of what you found". If every rhyme you found is a rare word, dragging far right can empty the list completely. That is the slider telling you something true about your rhymes. Drag it back down and they all return: the bar sets a threshold rather than running a new search.

The refined list shows at most a thousand rows at a time, purely so the panel stays quick, and more loads the next batch — the words past the thousandth are waiting their turn, not missing. And because refining hands the rows back most common first, raising the bar only ever trims from the bottom: the most common words are already in front of you.

Stress marks disappear when you refine

This is the part that surprises people, so it is worth stating plainly: refined rows show the word without its stress mark, and that is correct.

The corpus counts spellings. It has no idea where the stress falls in any of them — that information simply is not in there. So за́мок and замо́к are one and the same entry to it, and there is exactly one number for the plain word, however you were stressing it a moment ago.

That has two visible consequences:

• Refined rows are shown plain. Keeping a mark would be the app claiming a distinction the count cannot make.
• A word that appeared twice in your rhyme list — once per stress — appears once after refining. It has not been thrown away; the two rows had the same count, so they were the same row.

Turn Refine off and the marks come straight back, along with the separate Proximity scores for each reading. The rhyme list still knows the difference. Only the counting does not.

Next

Rhyme search and the Proximity score covers the other half of the same panel: where the results come from in the first place, what the dictionaries do, and what makes a rhyme score high.

The syllable grid is where the word you picked has to actually land on a beat.

Or just try it: open the realyrics editor, search for a rhyme and press Refine.