Word Frequency Counter

A frequency table is the fastest way to find out what a piece of writing is actually about, and the fastest way to catch the word you have leaned on eleven times without noticing. This counter reads whatever you paste and ranks every term in it, with the count, the share of the text it occupies, and how far in it first shows up.

The thing that makes one frequency count differ from another is the definition of a word, so that definition is yours here rather than buried in a regular expression. Capitals can be folded together or kept apart. Hyphenated compounds can be one term or four. Apostrophes can stay inside contractions or be treated as breaks. Numbers can join the tally or sit it out. Markup and links can be stripped before anything is counted, and English plurals, possessives and -ed and -ing endings can optionally be folded onto a common stem so that "design", "designs" and "designed" stop competing with each other for the top of the list.

Words are not the only unit. Switch to phrases and it counts runs of two to five adjacent words — the pairs and triples that reveal a stock sentence or a habit of phrasing — and it refuses to build a phrase across a full stop, a line break or a bracket, so nothing in the table is an accident of two neighbouring sentences. Switch to letters and it counts single characters, which is what a cryptogram, a word game or a keyboard-layout argument needs.

The list of words to leave out is visible and editable in English, Spanish, French and German, because a filter that silently decides most of the table should not be a mystery. Click any row and every occurrence appears with the words on either side, so a count turns into an edit you can actually make. And a second box lets you count another text with identical settings and see, term by term, which words the two use differently.

Phrase counting is where a repetition problem usually becomes visible, because writers repeat constructions more often than they repeat single words. Phrases are assembled from the untouched sequence of words and only then judged against the ignore list, which is deliberate: dropping filler first would splice non-adjacent words together and report pairs that never appeared on the page. You choose whether phrases containing an ignored word are kept, dropped when one sits at either end, or dropped outright.

Letter mode counts single characters, either letters alone or every character that is not a space. It is the mode for solving a substitution cipher, checking the letter distribution of a puzzle, or settling an argument about which keys a language actually uses.

Clicking a row opens its concordance — up to two hundred occurrences, each with about fifty characters either side and the term itself marked, prefixed by how far through the text it sits. Seeing the same word in twelve contexts at once is what tells you whether it is a tic worth cutting or a technical term doing its job. The word cloud is the same data drawn loosely: the top sixty terms sized by the square root of their counts, so the most common one does not swallow the panel.

Every share in the table is measured against the terms that were actually counted, and the figure it is measured against is printed above the table: so many found, so many counted, so many set aside. That distinction matters the moment you switch an ignore list on. If a hundred-word passage contains forty function words and you ask for them to be left out, the remaining sixty are the population, and the column adds up to a hundred percent of that sixty rather than to sixty percent of something you can no longer see.

Where the filters have removed anything, a second column appears giving each term's share of everything found, filters included. The two readings answer different questions — one is "how much of the writing that carries meaning is this word", the other is "how much of the document is this word" — and neither is more correct, so both are on offer rather than one being chosen for you.

Shares are printed with enough precision to stay useful in a long document. In a hundred thousand words, a term used once and a term used four times would both round away to nothing at one decimal place, so small values are given three.

A tally works on the characters in front of it. Unless you ask otherwise, "walk", "walks", "walked" and "walking" are four entries describing one idea, and a word that carries two meanings — a company and a fruit sharing the spelling "apple" — is one entry describing two. This is a limitation of counting rather than a fault in any particular counter, and it is worth holding in mind before treating a ranking as a summary.

Two controls narrow the gap. Grouping word forms applies English suffix rules, folding plurals, possessives and -ed and -ing endings onto a stem and undoing a doubled consonant so that "running" reaches "run". It is rules, not a dictionary, which means it will never turn "ran" into "run" or "better" into "good"; the control says so, it is off unless you switch it on, and every spelling that was merged is listed on the row so you can see what happened. Turning capitals back on goes the other way, separating a proper noun from the ordinary word it shares its letters with.

There is one more habit worth resisting: reading frequency as importance. A term at the centre of a document may appear rarely because the writer worked to avoid repeating it, while a term that recurs constantly may be scaffolding. The standard fix weighs a word by how unusual it is across a whole collection of documents, which needs a collection. The comparison box is the version of that idea you can run without one — count a second text and the change column shows which words this text leans on that the other does not.