Suggestions & prediction

The suggestion strip is the row of tappable words above the keys. It offers completions while you’re mid-word, corrections when what you typed doesn’t look like a real word, and next-word predictions once you’ve finished one. All three come out of the same on-device engine, which also learns your own vocabulary and phrasing the more you type.

Word candidates in the suggestion strip, mid-word.

This page covers the strip itself: its layout, the completion/correction/prediction engine, and how ranking adapts to you. A few things that also live in the strip have their own pages: autocorrect’s confidence math, the personal dictionary screen, the suggestion blacklist, smart chips (calculator, currency, unit conversion), password-manager chips and smart replies, contact and app-name suggestions, and emoji suggestions. Each is linked from the relevant section below rather than repeated here. Everything on this page is about the Latin/English-style strip. A conversion IME like pinyin or kana→kanji draws a different, scrolling candidate row that isn’t covered here.

What’s on the strip

Word candidates fill three slots by default. Number of suggestions (see Options) takes that as low as two or as high as six. Turn on More slots for short words and the strip grows past that number, by up to three slots, whenever the next words are short enough that every one still fits its share at full size: a, is, it, in can get five slots on a phone where international keeps three. Long words get the usual number again, so nothing is ever shrunk to make room. It does nothing while Scroll the suggestions or Keep suggestions in place is on. The strongest candidate, the one autocorrect would pick, sits in the middle with the runner-up to its left. That’s the layout most keyboards use. Turn off Best suggestion in the middle and the candidates rank left to right instead.

Swipe down for more suggestions (off by default) is for the words that don’t fit. Swipe down on the strip while it has words and a grid of up to 24 candidates covers the keys, best first, the strip’s own words leading. Between words, the predictions for the next word page the same way. Tap one to use it as if you’d tapped it on the strip. Swipe the strip again, or tap Back to the keys at the top of the grid, to close it. With this on, a downward swipe over suggestions opens the grid rather than hiding the keyboard; over an empty strip it still hides it if that setting is on.

After the words the grid has two more groups, each under its own switch below that one. Emoji in the grid (on) adds up to 24 emoji that match the word, the same search the emoji panel runs; tapping one works like an emoji on the strip. Fancy text in the grid (off) adds the word written in every fancy text style, so you can pick a look by seeing it; tapping one puts it in place of the word, and nothing about it is learned. The fancy group only appears while the Fancy text tool is on, and never in a password field.

Swipe up for more suggestions (off by default) works the way HeliBoard’s does. Swipe up on the strip while it has words and up to 12 more candidates open just above it, over the app, without covering the keys. Keep your finger down and slide it onto the word you want; the word under your finger lights up. Lift on it to use it, or lift anywhere away from the words (back down on the strip, say) to close them without picking anything. The best of the extra words sit in the row nearest the strip, so they’re the shortest slide away.

More suggestions in a grid (off by default, shown once either swipe is on) lines the extra words up in columns, one per suggestion slot, the way the strip itself is laid out, for both the swipe-down grid and the swipe-up words. Off, each word takes only the width it needs.

Slots are equal width, so a long word gets no more room than a short one. Rather than cut it off, the strip shrinks it to fit: the point size drops first, down to about two thirds of normal, and past that the letters themselves narrow a little. Most words that would once have read “internationali…” now fit whole, set smaller and tighter. The shrinking itself has no settings, though the size it shrinks from does: Suggestion text size under WM Keyboard › Appearance › Toolbar runs from 80% to 160%. Short words are never scaled, and a word that already fits is drawn at full size. Suggestion spacing, on the same screen, is the gap either side of each word inside its slot, 6 dp out of the box. Lower it and a long word has a little more room before it shrinks.

If you’d rather the strip never shrank anything, turn on Scroll the suggestions. Every word is then drawn at its natural size and the row scrolls sideways when the words overrun the strip. The row also runs on past Number of suggestions, through every word the keyboard found for what you typed, up to two dozen, so a word that ranked fourth (an Avro reading, say) is a swipe away instead of nowhere. A slot still gets at least its equal share of the width, so a set of short words looks exactly as it does with the setting off. It’s the setting for a narrow phone that wants five or six candidates: at fixed widths those slots are too tight to read. A new set of candidates always starts back at the left edge. The strip a panel’s own search box gets (the clipboard, emoji, notes and dictionary searches, the to-do box) scrolls the same way and runs just as deep. Turn on More suggestions arrow, under the same setting, and when there are more words than slots a ▾ arrow at the end of the strip opens every word the keyboard ranked in a grid over the keys, the same grid Chinese Pinyin’s arrow opens for its characters and Swipe down for more suggestions opens. Tap the arrow again to close it. It ships off, so the words keep that space; the swipe opens the grid either way. Neither the arrow nor the swipe is there under a panel’s search box, where the panel is in the grid’s place.

Word candidates followed by predicted emoji.

Up to four predicted emoji can ride after the word chips (see Emoji in the strip), and a row of punctuation marks can ride after that (see Punctuation chips), but never both at once. The emoji tail always wins the shared space. Punctuation only appears when no emoji prediction is offered. The tail sits on top of Number of suggestions: three words and an emoji is four chips. Turn on Emoji takes a slot (see Options) and the best emoji takes the last slot instead, the same width as a word, with one word fewer, so the strip never shows more chips than the number asks for. Suggested emoji sets how long the tail can get, from one emoji to four.

The words share the strip between them by default: two words split it in half, and an emoji tail, the punctuation marks, a tool’s keyword chip or a rewrite chip (join, revision, undo) takes room from them, so the best word can drift from one keystroke to the next. Turn on Keep suggestions in place and the strip is always cut into Number of suggestions slots of one width. A slot with nothing to show stays empty, and the best word keeps its slot. With Best suggestion in the middle on that is the second slot, the runner-up on its left and the third word on its right, exactly where they sit on a full strip. Anything that is not a word takes a slot of its own rather than pushing the words aside: a suggested emoji or the punctuation marks the last slot, a tool’s keyword chip or a rewrite chip the first. It has no effect while Scroll the suggestions is on, since that strip sizes each word to fit.

Coloured suggestions puts each word on a rounded pill in the theme’s accent colour instead of separating the words with thin lines, with the best word’s pill the deepest. With a Material You theme the colour comes from your wallpaper, the way SwiftKey shades its suggestion bar. A theme can give the pills a colour of their own instead of the accent: Suggestion pills and Main suggestion pill in the theme editor, on the suggestion strip’s sheet.

Two more colours mark single words rather than slots. Both sit under WM Keyboard › Appearance › Toolbar and both stay off until you pick one. Autocorrect word color paints the word autocorrect has decided to put in for a space, and no other word, so the colour is a promise about what the next commit types. New word color paints a word that no word list, your personal dictionary, your contacts or the spell checker knows. That is usually the word you are typing, which lets you see it would be new to the dictionary before you add it from its hold menu or let the keyboard learn it; the colour goes the moment either happens. Where both apply the autocorrect colour wins, and neither is used on a layout that turns your keys into another script, where the strip shows readings rather than the word as typed.

When one of the keyboard’s own text boxes has the keys (a panel’s search bar, the AI prompt, Find and replace, a plugin’s field), the strip moves to the row directly above the keys, since the panel has taken the toolbar’s row. It shows the same words for the box’s text that it would for your field, and tapping one replaces the word at the box’s caret. The exception is the emoji panel’s search box, which is querying the emoji catalogue rather than writing prose: its strip offers emoji search terms instead of dictionary words. See Panels and captured keys.

The strip shares its row with the toolbar, and several other features can claim that row instead of word candidates. A smart chip (calculator, currency, unit conversion) takes the full row when it has an answer. Autofill chips take it over completely on a login field. The system’s smart replies take the row when nothing is typed, then ride the tail once you start typing. A “paste your last copy” or one-time-code chip can appear from the clipboard tool. Typing : turns the row into inline emoji search. Any of these counts as “the strip has something to show,” so the row rests on suggestions rather than falling back to the toolbar. Suggestion strip always visible (see Options) keeps the strip as the resting view even before any of them has anything to offer.

Completions, corrections, and predictions

One engine produces everything in the word-candidate slots, drawing from several sources depending on what you’ve typed:

• Completions. While you’re mid-word, every source that has a word starting with what you’ve typed contributes a candidate: the active dictionary, any downloaded or custom-imported dictionary, a secondary language’s dictionary if you type bilingually, your own personal dictionary of learned words, and (if you’ve turned them on) contacts, contact emails, and app names. If the only thing wrong with what you typed is a missing space, as in “ofthe”, an “of the” candidate is offered too. A word joined with a hyphen is completed a part at a time: auto-sp offers auto-space and seventy-ni offers seventy-nine, as long as the parts before the last hyphen are words, even when the lists don’t hold the whole compound.
• Corrections. If what you’ve typed matches no known word anywhere, the engine walks the dictionaries for words a short edit away: a letter dropped, added, doubled, swapped with its neighbour, or substituted. A word of five letters or more may be reached in two such edits; anything shorter gets one, since a two-edit fix on a four-letter word explains almost anything. Each candidate is scored by how plausible that slip is on your active keyboard layout, and by where your finger landed. The best-scoring one appears as an ordinary tappable candidate, not only as a silent autocorrect target. So an ambiguous typo like “helo”, evenly split between “hello” and “help”, still gets you both real words to pick from, even when autocorrect itself stays quiet because neither candidate clearly wins. A word that is already known keeps its two nearest one-letter slips on the strip, ranked below it unless the words before it favour the slip: “is” also offers “us”, one key over. Up to 0.6.3 a known word got no corrections at all, which hid the word you meant whenever the slip happened to spell another word, as one often does in a big downloaded list such as Arabic’s, and that included words from a secondary language’s list.
• Predictions. With nothing composing, right after you finish a word, the strip predicts what comes next instead. Your own learned pairs come first, the three-word context ahead of the two-word one. Next comes the following part of a contact’s name, if the previous word matched one. Then your own gappy habits: what has tended to follow the word two back one word later, and the word three back two words later, so “how can someone” can still offer “help”. Then a downloaded n-gram pack for your language, if you have one installed. Last is a small bundled set of common English pairs, so a fresh install has something to offer before you’ve typed anything of your own. If nothing knows the word you just typed, a name or a typo, the strip falls back to what tends to follow the word before it rather than going quiet.
• Openers. An empty field, the space just after a full stop, and the start of a new line are all the start of a sentence, and the strip predicts from that the same way it predicts from a word: the openers you use come first, then a bundled set of common English ones, so there is something to tap before you have typed a letter. A new line counts because it starts a new item or paragraph, so the strip no longer offers words that would follow the last word of the line above. The word you pick is learned as an opener too, so the list becomes yours. Openers only appear with Suggestion strip always visible on. With it off, an empty field, a new line and the space after a full stop rest on the toolbar with no arrow back to the strip, and the candidates come up with the first letter you type.

An ambiguous typo offered as two tappable candidates instead of being auto-corrected.

Candidates from every source merge into one ranked list, and the sources don’t count equally. Each is weighted well above plain dictionary frequency, so a source that’s more clearly you (or more clearly what you’re reaching for) dominates once it has anything to offer:

Source	Weight
Bundled/downloaded dictionary	baseline. A downloaded list counts words on a far bigger scale than the built-in one, so it’s scaled down until its commonest word matches the built-in list’s, keeping its own order
Custom-imported dictionary	×100, or scaled down the same way if the list carries real word counts
Secondary-language dictionary	×40, scaled 0.5×–2× by how much you mix that language in. A downloaded list with real word counts is capped lower, so its commonest word sits at half the primary’s commonest and never buries a word you typed in full in your main language
Your personal dictionary	×500
App names	×400
Contact names	×3,000
Contact emails (2+ character prefix)	flat 4,000

Tap a candidate to insert it (see Options for the space that follows). It’s cased to match what you typed: a leading capital gives you a capitalized candidate (“Hte” → “The”), and typing in all caps gives you the candidate in all caps (“HTE” → “THE”). Anything else is inserted as stored.

When the chip is replacing a word that’s already sitting in the field, it takes that word’s capitals instead. This is the strip after a swipe or after handwriting, where the chips are other readings of the word that just landed, and it’s also a word you tapped back into mid-sentence. So if auto-capitalization gave you “Van” at the start of a sentence and you swap it for “can” off the strip, you get “Can”, and swapping a shouted “VAN” gets you “CAN”. Only capitals that Shift, caps lock or auto-capitalization put there carry over. A name’s or an acronym’s own capitals are its spelling, not yours, so swapping “Joe” for “javelin” mid-sentence gives “javelin”, and “QWERTY” for “qwery” gives “qwery”. At the start of a sentence the new word still gets its capital. The same goes for the recent words row and for a word picked from Synonyms.

The context model

Everything above reads at most three words back. Context model (WM Keyboard › Typing › Automatic corrections, on) adds a small neural language model that reads the whole sentence so far. It is a file you download per language, from the language’s own screen under WM Keyboard › Languages & layouts (the Context model row, about 10 MB; the Download button also offers a Large English model of about 17 MB, which predicts the next word about a point more often and takes longer on each word on a slow phone, and the row’s menu switches between the two later), and it runs on the phone: nothing you type leaves it. English has one, German has one, and so does romanized Bangla, whose model reads English and Banglish in the same sentence. A language without a model keeps everything else on this page exactly as it was.

The model answers one question once per word you commit: given the sentence so far, how likely is each word of its vocabulary to come next. That one answer serves the strip three ways.

• Predictions. Its likeliest next words join the next-word strip, after your own learned pairs and before the downloaded word pairs, so “I was wondering if you” offers “could” and “would” where the pairs alone could only see “you”. Measured on sentences the model had never seen, top-one next-word prediction rose from 14% to 20% (top three from 23% to 34%), and the letters you could leave untyped by taking the first suggestion from 38% to 44% on careful typing.
• Ranking. Its opinion of each candidate for the word you are typing joins the score the strip is ranked by, before the word pairs have their say: a candidate the model expects more than the strip’s current leader moves up, one it expects less moves down, by a bounded amount. The keys you hit still outweigh it, so a far-fetched word the sentence happens to predict cannot climb over a word your taps spell. Measured on sentences the model never saw, the strip’s first word is right about half a point more often on careful typing (87.2% to 87.5%), 3 points more on fast typing (79.5% to 82.7%) and 4 points more on sloppy typing (72.1% to 76.0%), where the keys say least.
• Autocorrect. Its score for a candidate is one of the things the learned gate weighs before a correction is applied.

Each model has a vocabulary of about twenty thousand words plus four thousand word pieces. A word outside it, a name, a term you taught the keyboard, a word from another language, is handled one of two ways. A model that carries word pieces spells it out of short letter groups and still gives it an opinion, so a name the sentence calls for is not held back for being new. While you type, this is done for the few words close enough to the leader for the model to change their order, four at most on each letter. A model without pieces gives no opinion: the word is neither lifted nor sunk, and your dictionary ranks it as before. The sentence it reads begins at the last full stop, question mark or exclamation mark, and it also knows where it is being read: a chat, an email, a search box or anywhere else, and whether the app is one you write casually in (the same lists Adapt to app tone uses), so a chat leans conversational and an email formal.

Phrase suggestions (WM Keyboard › Typing › Automatic corrections, off) lets the model offer the next two or three words as one chip after a space, when it is sure of each of them in turn: “going to be” rather than “going”. Tap the chip and all of them go in. Nothing is offered when the model is not sure, which is most of the time.

Re-check the previous word (WM Keyboard › Typing › Automatic corrections, on) is the model reading backwards. The word you have just typed often shows what the word before it should have been: “I want to by” is fine until “a car” lands. Once the next word is in, the keyboard asks the model about the previous word’s other readings, the words the strip had for it at the time, with both sides of the sentence in sight, and it keeps asking as the words after it land, so a reading that two or three later words make obvious is still found. When one fits twenty times better than what stands there, a chip offers it, the same chip the confusable rule uses for “their” and “they’re”. It never rewrites anything on its own, and it only ever offers a word the strip had already found for the taps you made, so it cannot invent a correction from the sentence alone. When you stop typing for a second with the strip empty, it also reads the last few words together once and offers the one swap among them that fits best, if any clears the same bar; the chip goes away as soon as you type on.

Before you type

Words before you type, under Typing / Suggestions, puts your own words on the strip while nothing is typed. At the start of a sentence they replace the usual openers. After a word, predictions keep their slots, your words fill any left empty, and the rest head the grid behind the strip.

In a right-to-left script the best word starts at the right edge. Right-to-left words start on the right, in the same place, turned off keeps the strip and its buttons where they are in every other language.

With the toolbox button on the strip, a swipe up on it opens the pinned tools, the same as the arrow, and a swipe down on it or on the toolbar’s own toolbox button closes them. Where the tools go set to On a swipe makes that the only way: the strip keeps the suggestions until you swipe.

When the cursor lands in a word

Put the cursor inside a word you have already written, by tapping it or by walking the caret back with the spacebar, and the strip stops predicting the next word and answers about that one instead. You get its completions and corrections, the same as if you were typing it, plus the word itself as the top candidate, which with Best suggestion in the middle on is the middle chip.

A word you highlight gets the same answer, abbreviations with a full stop inside them included: select i.e. or e.g. and the strip offers that, not the words it showed before. A word that only ends in a full stop is left alone, because a chip picked over it would take the sentence’s full stop with it.

Phonetic layouts such as Avro answer the same way for a word written in their script: put the cursor in or after আমা and the strip offers আমার, আমাদের and the other words that begin with it, from the layout’s own word list. With the cursor inside the word, typing there writes a new word at the cursor, and picking a chip replaces the word.

Right after a word counts too, as in next word| to. There the word is normally picked back up as if you were still typing it, so a letter you type goes on the end of it. Some apps won’t let the keyboard do that, and in those the strip still answers about the word: picking a chip replaces it, and typing adds to the text.

A phonetic layout such as Avro does the same with words in its own script. The keyboard spells the word back into the Latin letters that type it, so with the cursor after আমি the word is underlined again as ami, and typing r makes it আমির. It only picks a word up when those letters give back exactly that word; any other word gets the strip, and its chips, but stays as it is while you type. See Going back to a word.

That chip is there for the hold, not the tap. Tapping it would replace the word with itself, so it does nothing but put the strip away. Holding it opens the word menu, which is how you add the word to your dictionary, open its card to respell or re-rank it, delete it, or block it, all without leaving the sentence. Holding any other chip reaches the same menu, and its Add and Search all words items are about the word the cursor is in rather than the chip you held.

Only offer new words (WM Keyboard › Typing › Suggestions, off) takes that chip away, the same way it takes away the slot the word you are typing normally holds. With it on, every slot offers something different from what is already there, and the word menu is then only a hold on some other chip away.

A word a swipe wrote gets one thing more: the strip leads with that stroke’s other readings instead of respellings of the letters standing in the field. See Going back to a misread swipe.

Retyping the last word

Retype word is a text editing action for the moment a word went in wrong and you would rather type it again than fix it letter by letter. It is in the menu a held suggestion opens, as Retype “word” naming the word it would take, and you can also put it somewhere of its own: on a key or a flick in the layout editor, on a long-press alternate, a key of the text editing panel, a multi-finger swipe or a ?123 chord.

It takes the word right behind the cursor back out of the field, with the one space after it if there is one, and leaves the cursor where the word began. The strip then shows that word first, so one tap puts it back exactly as it was. If you swiped the word, the other words that swipe could have been follow it, and taking one of them teaches the keyboard the same thing picking it right after the swipe would. A word you tapped out has no such list, so the strip offers only the word itself. Type anything and the strip goes back to ordinary suggestions.

A space the keyboard added after a picked or glided word goes with it; a space you typed yourself stays. It only ever takes the word at the cursor, never another copy of it further back. Behind the cursor there has to be a word, or a word and one space; after a full stop, a line break or two spaces it does nothing. It also does nothing while text is selected, in a field that won’t say what is behind its cursor or exactly where the cursor is, or when the field has changed since the keyboard last looked. A word still being typed is taken back whole, and the strip keeps what it was offering for it.

Fixing a word you’ve typed past

Turn on Recent words (WM Keyboard › Typing › Suggestions, off) and the three words you wrote last sit in a row of their own above the suggestion strip, the newest on the right. Each word is followed by what else it could have been, so fixing it is one tap. It is for the word you notice is wrong only after you have typed on. The strip keeps all its slots, and the keyboard grows by the row only while it has words to show; when they don’t fit, the row scrolls, starting at the newest. Under WM Keyboard › Rows & bars › Recent words you can move the row in the row order, put the words at the start of the strip instead so the keyboard never grows for them, or have the row keep its height while it’s empty so the keyboard doesn’t jump between messages.

What follows a word is what the keyboard had for it as you wrote it: the suggestions on the strip at that moment, and the other readings of its swipe if you swiped it. A word typed where nothing is suggested, such as a browser’s address bar, kept nothing, so it gets the dictionary’s corrections instead. Tap one and only that word changes, where it stands. The word you are still typing stays as it is, the cursor stays where it was, and the text after the word moves by the change in length. Your old spelling goes to the front of that word’s choices, so the same row changes it back.

A word on the row is a place in the text as well as a spelling, so if “form” appears twice, a tap changes the copy it was drawn for and not the other. If the text has changed since the row was drawn, such as an app rewriting it or you editing it by hand, the tap does nothing and the row is read again. The row holds only words you wrote in this field: it starts empty in every new field, and text that was already there never joins it, wherever you put the cursor. It also shows in a browser’s address bar, where searches are typed, and never in password fields or while incognito is on. It is kept only while you type, never saved.

Punctuation chips

Turn on Punctuation suggestions (off by default) and a short row of marks (. , ? ! ’) rides after the word candidates, set off by a thin divider. Tap one and it commits exactly as typing that key would, so a full stop or comma is one tap away without switching to the symbols layout. Those five marks are only the default. A Punctuation marks text field appears under the toggle and takes up to twelve characters, one chip per character, so a Bengali danda or the Spanish inverted marks can sit here instead. Clearing the field puts the five back.

Punctuation chips riding after the word candidates.

The row only appears while word candidates are already showing, and only when no emoji prediction has claimed the tail instead.

Shift re-cases the strip live

Arm Shift and every chip in the strip re-cases in real time to match what would be typed if you tapped it right then. Shift is a latch rather than a hold, so one tap gives you Title Case, a double tap (caps lock) gives you UPPERCASE, and another tap goes back to lowercase. The chips follow the shift state, so they revert the moment it clears, which includes the letter or suggestion that spends it. The whole row follows, not only the top candidate. One thing is left alone: a contact-email suggestion, meaning anything containing @, since an address isn’t meant to change case.

Suggestion chips re-cased to Title Case while Shift is armed.

This is a display rule only. It doesn’t reject or change which candidate is offered, and it’s unrelated to reverting an autocorrected word with backspace.

Emoji in the strip

Matching emoji append to the end of the word candidates once you finish typing a whole word. The same mechanism can offer an emoji on an empty buffer once it has learned that a word or two-word phrase you often type is usually followed by one, and after an emoji it offers the words you tend to type next. Emoji suggestions have their own settings and their own page: see Emoji suggestions while typing for the toggle, the default, and how tapping or holding one behaves. What matters here is the tail of the strip, which fits up to four emoji. The emoji tail and the punctuation chips both want that space, and emoji always wins.

Learning from what you type

Every word you commit, and every pair of words you type back to back, feeds a private on-device vocabulary. Both the completion ranking above and next-word prediction draw from it. How strongly a word is taught depends on how deliberately you typed it. Tapping a suggestion from the strip teaches harder than a word that was merely typed and committed, and a word autocorrect chose for you earns no credit at all, only the word pair around it. The Personal dictionary page covers that grading in full, along with adding, removing, and clearing what’s been learned. What belongs here is what learning does to ranking: a learned word carries a ×500 weight (see the table above), so it competes with common dictionary words almost immediately instead of needing to be typed dozens of times first.

Note

If you type in more than one language, the keyboard separately tracks which secondary language you reach for and how often. Older commits decay geometrically, so the weighting follows your recent habit rather than your whole history. That’s the 0.5×–2× scaling factor on secondary dictionaries in the table above. It’s a deeper multi-language topic than this page covers, and it isn’t tied to word-level personal-dictionary learning. A second, much faster signal reads the language of the field you’re in right now and leans ranking toward it. That one has its own page: Mixed-language typing.

Removing an unwanted suggestion

Press and hold the word on the strip. The menu that opens has Never suggest “word”, which puts it on the suggestion blacklist, and Delete “word” for a word the keyboard can forget: one you learned, one still waiting to be learned, one in Android’s dictionary, one you adjusted the rank of, or one from a word list you imported. Forget what you learned about “word” sits beside it for a list word you have taught something: it drops the learning, its capitals and rank included, and keeps the word. Synonyms lists words with the same meaning, and tapping one puts it in place of the word. Edit opens the word card, which carries Never suggest and Delete and the rank controls besides. With Selection actions on, selecting a word in the text offers Edit word, which opens the same card for it. The Personal dictionary page describes the menu and the card in full, and Word menu under WM Keyboard › Typing › Suggestions picks which of the optional items it shows.

The same two lists are editable from settings, if you would rather not chase the word on the strip: the Personal dictionary screen for learned words, and the suggestion blacklist screen for words that should never be offered or corrected to, whether or not they were ever learned.

Options

WM Keyboard › Typing › Suggestions

The Suggestions group on the Typing settings screen.

Setting	Default
Suggestions	On
Punctuation suggestions	Off
Number of suggestions	3 (range 2–6)
Swipe down for more suggestions	Off
Swipe up for more suggestions	Off
More suggestions in a grid	Off
Emoji takes a slot	Off
Suggested emoji	4 (range 1–4)
Keep suggestions in place	Off
Coloured suggestions	Off
Scroll the suggestions	Off
Suggestions in every field	On
Suggestion strip always visible	Off
Best suggestion in the middle	On
Shorten long words	In the middle
Numbers from the letter keys	Off
Remember numbers you type	Off
Space after a suggestion	On
Space after a suggestion (per language, under Languages & layouts)	On (Off for Bangla)
Block offensive words	On
Context-aware suggestions	On
Context model	On
Re-check the previous word	On
Phrase suggestions	Off
Adapt to app tone	On
Suggestion blacklist	Empty

Suggestions is the master switch for the whole strip described on this page. Turn it off and completions, correction candidates and next-word prediction all stop. Autocorrect keeps working: while its own toggle is on, the word you’re typing is still underlined as it’s composed, and space or Enter still corrects it, fixes a missing apostrophe and learns it. Turn both off and letters go straight into the field with no underline. Number of suggestions is how many word slots the strip splits itself into. Three suits a phone; a tablet or a landscape screen has the room for more, at the cost of each candidate getting narrower. Scroll the suggestions drops the fixed slots for a row that scrolls sideways, so more candidates cost nothing in legibility (see What’s on the strip). Suggestions in every field overrides the handful of apps that ask the keyboard to hide suggestions on an ordinary text field. It also covers a browser’s address bar: with it on, the words you type there are completed on the strip, though never autocorrected. Password fields and structured keypads (number, phone, and similar) never show suggestions whatever this is set to. Space after a suggestion adds a trailing space when you tap a candidate, the same as finishing a word yourself would. When the word you picked already has a space after it, as it does when you go back and fix a word mid-sentence, no second space is added. The caret moves past the space that’s there instead, so the next word you type starts after it rather than running into the one you just picked. The same goes for a chip that replaces a word in the middle of a sentence, with the cursor inside the word or the word highlighted, and for one picked in a browser’s address bar. In front of a mark that closes up onto the word, such as ?, ,, . or a closing bracket or quote, no space is put in: replace question in A question? and you get A questionable?. The space waits instead, and if you go on typing a word at that spot it goes in first, so A questionable spa…? comes out without a press of the spacebar. A mark typed there, or moving the cursor, drops it. A word typed into the middle of another one, with letters on both sides of the cursor, gets no space at all, since one would split the word in two. Each language also has its own Space after a suggestion switch, under Languages & layouts / your language / Spacing, on by default for every language but Bangla. Turn it off for one language and picks made while typing it go in bare, while every other language keeps the space. That suits Bangla, where a picked word often takes a case marker, postposition or suffix straight after it, and a kar picked in the middle of a word must not split it, so Bangla starts with it off. The main switch still rules every language, and the per-language rows are greyed out while it’s off.

None of these reach a conversion keyboard. Chinese, Cantonese and Japanese put the composer’s own character candidates on the strip, and that list is how the language is typed rather than a guess about it, so it survives Suggestions being off, a field that asks for a silent strip, and a password box alike. See Chinese, Japanese & Korean.

Numbers from the letter keys (off) is for boards with the number row hidden, where each top-row letter carries a small digit in its corner. When every letter of the word you’re typing sits on one of those keys, the strip’s last slot also offers the number those digits spell: qwe on QWERTY offers 123 beside the words, and one tap types it. It reads the keys you pressed rather than the letters, so AZERTY’s azerty spells 123456 and a layout with no digit hints offers nothing. A key’s digit counts wherever its popup lists it, so a is still 1 with à ahead of it. The number is never the first slot, never what a space commits and never an autocorrect target. Two letters at least, so a lone I stays a word rather than an 8. Five digits or more are grouped the way the number chip groups them, under its Grouping style, as long as Group long numbers is on. With the number row shown the digits have keys of their own, nothing is offered, and the row is hidden.

Remember numbers you type (off) completes the numbers you type often, the way Samsung’s keyboard does. A number of three digits or more that stands on its own is kept when you press space or enter after it, or leave the field. The next time you type its first digits, the strip shows the numbers you’ve typed that start with them, the most used first, and one tap replaces your digits with the whole number. No space follows it, since a number is as often followed by a full stop as by a word. A number is only one of its own after a space, at the start of the text, or after +, ( or #, so the 880 of +880 counts. Digits stuck to letters (win10) and the pieces of a number split by a comma, a colon or a dash (1,234, 12:30, 0171-234) are left alone. Nothing is kept from a password field, in incognito, or while Learn from typing is off. Number and phone fields have no strip at all, so this works in ordinary text fields. The list stays on the phone, holds up to 100 numbers, and is deleted with the other learned data under Storage.

Seven of the rows above aren’t in the Suggestions group at all. They live under WM Keyboard › Typing › Automatic corrections, and they’re listed here because what they do is change the strip:

• Space after a suggestion (on) is described above, and sits with the other spacing rules rather than with the strip.
• Block offensive words (on, and only visible while autocorrect is on) uses the same filtering mechanism as the suggestion blacklist to keep a bundled list of profanity and slurs out of the strip, not only out of autocorrect.
• Context-aware suggestions (on, also only visible while autocorrect is on) reorders the top eight candidates using the words you typed before them: your own learned word pairs and triples, a downloaded n-gram pack if you have one, common English pairs, and words already used in this conversation. It only reorders when there’s real evidence to reorder on, so an unfamiliar context keeps the usual order. It never changes which word autocorrect would apply.
• Context model, Re-check the previous word and Phrase suggestions (the first two on, the third off, all only visible while autocorrect is on) are the neural model described in The context model: the sentence-aware next words, the chip that re-reads the word before, and the few-words chip.
• Adapt to app tone (on, and visible whatever autocorrect is set to) tilts ranking by where you’re typing. Chat-speak like “lol” and “gonna” ranks a little higher in messaging apps and a little lower in email apps and email fields. It’s ranking only: no word is ever hidden or blocked, and autocorrect doesn’t change. The context model reads the same lists to know where it is.

A physical keyboard can also pick a strip candidate by number without touching the screen (see Hardware keyboard for how that’s bound).

Details & edge cases

• A known word gets only its two nearest slips. Anything in the active dictionary, a custom or secondary-language dictionary, or your personal dictionary is offered as-is, with its completions and at most two words one edit away, each scored as at most half as likely as the word you typed before context moves it. Wider corrections, two edits away, are only walked for a word nothing knows. Autocorrect never replaces a known word either way.
• A dictionary word and a contact name can collide on capitalization. A contact’s name keeps its own capitalization (“Wasi”) independent of how a dictionary word would be cased. If the same word would otherwise appear twice, once lowercase and once capitalized, only the better-scored one is shown.
• Missing-space splits need a real word on both sides. “ofthe” only becomes “of the” if both halves are independently known words. An arbitrary split of an unknown word never appears as a candidate. Turn off Add missing spaces under autocorrect and no split appears at all, so “history” half-typed is never offered as “his to”.
• The word as typed keeps a slot. Unless Only offer new words is on, the strip always holds what you have typed so far, exactly as typed, in the slot left of the middle, even when the engine has no idea what the word is, and in a browser’s address bar as much as a text box. A long word the dictionary lacks no longer empties the strip.
• The word as typed can go first. Word as typed first (WM Keyboard › Typing › Suggestions, off) draws it in the leftmost slot, wherever the engine ranked it, the way Multiling O Keyboard does. Only the drawing moves: the word a space writes is still the one in bold, and a hardware keyboard’s digit badges follow the slots as drawn.
• A long word borrows room from short ones. When one word would not fit its equal share of the strip and the words beside it are short, it takes the room they leave, so “international” next to “a” and “it” is drawn whole instead of shrunk. The short words keep exactly the width they need. While every word fits an equal share, the slots stay equal, so each word sits where your thumb expects it. A scrolling strip and Keep suggestions in place keep their own widths.
• Shrinking a word has a floor, and past it the ellipsis is back. The two stages together fit roughly twice the slot width. A word longer than that is cut, and so is one in a slot squeezed narrow by a long smart reply or a full row of predicted emoji riding the tail. The cut goes in the middle of the word by default, as most keyboards do, so its start and its ending both stay readable: “Punct…tion”. Set Shorten long words to At the end for “Punctuat…” instead.
• Only the word candidates scale. Conversion candidates (Pinyin, kana, jyutping) keep their fixed larger size, since Han and kana need it, and they scroll instead of splitting the bar evenly. Predicted emoji, punctuation chips and the clipboard pill are all fixed-size too.
• The strip fades rather than snapping. The composing buffer or field can empty out completely: you delete back to nothing, clear the field, or back out of a : emoji search. The last set of candidates stays visible and fades out over a short animation instead of disappearing instantly.
• A contact-email suggestion is never added to your personal dictionary, since an address isn’t a word to learn. Tap it inside a field the keyboard recognizes as an email input and it commits with no trailing space, because the address is usually the whole field. Tap it from an ordinary text field and it gets a trailing space like any other suggestion. Contact-aware typing has that behavior in full.

For the word-typo confidence math, and for how a correction is applied automatically instead of only offered, see Autocorrect. Personal dictionary has the full story on what gets learned, how strongly, and how to review or clear it. To keep a specific word out of the strip entirely, see the suggestion blacklist.