Two banks, one word

I sat on the grassy bank of the river. I walked to the bank for a loan. Same five letters, twice, and not the faintest relationship between the two things they name. Every automatic system that pairs words with pictures has to decide which one you meant, before the word has finished arriving, without asking you. That decision is the hardest problem in this product by a wide margin.

The temptation is to treat a word as a key. Look up bank, retrieve the picture filed under it, display the picture. It takes about four minutes of building that to discover it cannot work, and about four hundred words of English to discover how badly. Spring is a season, a coil, a source of water and a verb of movement. Crane is a bird and a machine and something you do with your neck. Light is illumination, and weight, and a match. Charge is electricity, an accusation, a fee, and a cavalry manoeuvre. Mine is a possessive, a hole in the ground and a buried explosive.

These are not exotic words. They are the ordinary furniture of English prose, and the ambiguous ones tend to be the frequent ones, because frequency is what wears a word smooth enough to take on extra jobs. A dictionary that maps one word to one picture will therefore be wrong most often on exactly the words that appear most often. That is a poor place for a system to be wrong.

The eight-word window

What we do instead is modest and mostly mechanical. Every sense in the curated lexicon — 528 hand-drawn icons carrying roughly 3,100 sense entries between them — arrives with a set of context cue words attached to it. When a target word comes up, the resolver takes the eight words either side of it and scores each candidate sense against what it finds there. River, water, grassy and sat pull hard towards the sloping edge. Loan, account, manager and walked to the pull towards the institution. Where nothing in the window decides it, frequency rank breaks the tie, which is a polite way of saying the system guesses the common sense and moves on.

Eight words is not an arbitrary number so much as a compromise you can defend from both sides. Too narrow and the cue simply is not in the window: the sentence that establishes you are reading about geology may be two sentences back. Too wide and the scoring drowns in irrelevant vocabulary from a different clause about a different subject, and precision falls without recall rising to pay for it. Everything here has to complete before the next frame, in a browser tab, on whatever machine you happen to be using, with no server to ask.

The literature is at least encouraging about the shape of this. Ambiguous words briefly activate more than one meaning before context selects a winner, and where two senses are roughly balanced in frequency, resolution is measurably slower — a real cost, visible in fixation times. Where a prior context strongly favours one sense, most of that cost disappears. So a correct image supplied alongside the word is not decoration sitting next to the sentence. On the good days it is doing a piece of the disambiguation work the sentence would otherwise have charged the reader for.

The cost of being wrong

Now the other side of the ledger, which is steeper than it looks. A wrong image is not a null result. If the resolver shows a bird when the sentence meant a lifting machine, the reader does not experience a blank or a shrug. They experience a bird, delivered through a channel that is very good at making concepts available quickly, at the precise moment they are assembling a clause. The wrong meaning does not sit off to one side. It goes in.

And it goes in quietly. Nothing about the reading experience flags it. At 400 words per minute you may carry a mistaken sense forward for three sentences before the text contradicts it firmly enough to register, and by then you have to unwind not just the word but everything you built on top of it. The repair costs far more than the image ever saved. This is why we treat accuracy on ambiguous words as a safety property rather than a quality metric: the downside is not symmetrical with the upside, and any honest account of the second channel has to start there.

It is worth conceding the obvious counter-argument here, because it is a good one. Every other tool in this category — Spritz, Reedy, Outread, the whole word-flashing family — shows no pictures at all, and therefore cannot possibly show you the wrong one. They have no exposure to this failure. We took the exposure on deliberately, on the view that a correct image is worth more than a wrong image costs, but that is a judgement rather than a measurement, and a reader who thinks the risk is not worth the return is reasoning from the same facts we are.

A wrong picture is not a missing picture. It is a meaning inserted into a sentence that never contained it.

Why disambiguation is a safety property

Where we lose

There is a specific list of cases where this approach does badly, and we would rather publish it than have readers discover it one bird at a time.

• Irony. The cue words in the window support the literal sense, because the literal sense is what the sentence says. The whole point of the construction is that the writer means the opposite, and nothing in a scoring function notices that.
• Metaphor. “The argument collapsed” gets a building. Live metaphor is systematically mishandled, and English prose is far more metaphorical than writers notice while writing it.
• Technical jargon reusing a common word. A bus in a circuit diagram, a tree in a data structure, a charge on a company’s assets. Frequency rank actively works against us here, because it votes for the everyday sense.
• Proper nouns. A person called Baker is not a baker, and a firm called Crane makes cranes about half the time. Capitalisation helps a little and is unreliable at the start of a sentence.
• Negation scope. We can mark a negation with an overlay. Knowing how far it reaches — which of the following clauses it governs — is a syntactic question, and we do not parse the sentence.

The unifying feature of that list is that each case requires understanding the sentence rather than scoring words near it. A perfect disambiguator would have to know what the writer was doing, and nothing that runs in 250 milliseconds inside a browser tab, with no network and no model of the discourse, is going to know what the writer was doing. We could make it better. We cannot make it right, and pretending otherwise would put us in the company of everyone else who has promised the language system away.

Confidence, and a pin

Two things follow from admitting all of that. The first is that the resolver has to carry a notion of how sure it is, rather than presenting every choice with the same flat certainty. A sense selected by four strong cues in the window is not the same object as a sense selected because it was the commonest option and nothing argued otherwise. Every image in the stream is already labelled with the tier that produced it — the mechanism is described in the article on coverage — and the same discipline applies to sense selection: the reader is entitled to know when the machine was guessing.

The second is that the reader has to be able to overrule us. If you read patent filings, claim means something specific and our lexicon will be wrong about it several times a page, forever, because the general frequency of English is against you. So a sense can be pinned in your own lexicon and it stays pinned. That is a paid feature, on the Practice tier, and we will not dress that up: the free app gives you the complete resolver and the complete icon set, and the thing you pay for is the ability to correct it and have the correction persist across sessions.

None of this closes the gap. It narrows it, and it makes the remaining gap visible, which is the most a system without a parser can honestly offer. If you want to see the failure mode for yourself, load a page of your own field’s writing into the app and watch the words your discipline has borrowed from ordinary English. The mistakes will be obvious to you and invisible to us, which is exactly why the pin exists.