Every query takes something for granted. Presupposition is the name for what it assumes.

A presupposition is a background assumption that an utterance takes to be true without stating it. When did you stop smoking presupposes that you smoked. A query carries presuppositions in the same way, and they are part of the user intent.

A presupposition trigger is a word or construction that introduces an assumption, such as stop, again, the, regret or also. The trigger is visible in the words. The assumption it brings is not.

Also discussed at https://plato.stanford.edu/entries/presupposition/

A presupposition survives negation. I did not stop smoking still assumes that I smoked. This test separates a presupposition from what a sentence merely implies.

Also discussed at https://plato.stanford.edu/entries/presupposition/

Common ground is the body of knowledge and beliefs that people in a conversation take as shared. Presuppositions are drawn from it. Between a searcher and a search system, the common ground is thinner than the searcher assumes.

Presupposition accommodation is the process by which a listener silently accepts an assumption that was not yet part of the common ground. David Lewis described it in 1979. Search systems accommodate every assumption in a query, true or false.

If a model handled a false assumption as well as an ordinary question, both columns would match. For every model the right column is lower.

A definite description is a noun phrase with a definite article, such as the best treatment. It presupposes that such a thing exists and is unique. Bertrand Russell and Peter Strawson disagreed about it for decades.

Also discussed at https://plato.stanford.edu/entries/presupposition/

Entailment is a relation in which the truth of one sentence guarantees the truth of another. When the sentence is negated, an entailment falls away and a presupposition stays. The difference shows what a query really takes for granted.

Implicature is meaning a listener infers from context, as Paul Grice described it. It can be cancelled easily, while a presupposition is built into the words. Both carry more than the query says.

A false presupposition is an assumption in an utterance or a query that is not true. Why does vitamin C cure colds presupposes that it does. Answering such a question as asked confirms the error.

Also discussed at https://arxiv.org/abs/2212.10003

26 % of these questions rest on a false presupposition. If spotting it were easy, the model and the searching person would score close to the person holding the best reply.

Holes, in Lauri Karttunen's terms, are expressions that let the presuppositions of an embedded sentence pass up to the whole sentence, such as know or be surprised. The assumption travels outward. The speaker is still committed to it.

Plugs are expressions that block presuppositions of an embedded sentence, such as say or tell. The assumption stays inside the reported speech. The speaker is no longer committed to it.

Also discussed at https://plato.stanford.edu/entries/presupposition/

Filters are constructions such as if and or that let a presupposition pass under some conditions and block it under others. If John has children, his children are asleep does not assume that John has children. Context decides.

An anaphoric presupposition trigger, such as again or too, needs something in the prior context to refer to. It cannot simply be accommodated. In a single query without context, it points to something the system never saw.

Described in a Google patent: https://patents.google.com/patent/US9547690B2/en

When a presupposition fails, the sentence can lack a truth value, as Peter Strawson argued. The king of France is bald is neither clearly true nor clearly false while there is no king of France. A query built on a failed assumption has no correct answer in its own terms.

Also discussed at https://plato.stanford.edu/entries/presupposition/

User intent is the goal and motivation behind a query. Presuppositions are the part of that intent the person no longer questions. They mark where the search intent read from the words and the user intent begin to part.

What happens when the assumption is wrong. The usual treatment names triggers and accommodation. It leaves out how assumptions project through longer queries, what happens when they fail, how they can be cancelled, and how a system could ask back. For search, that missing part is the whole problem: the headache tablet already assumed a pill.

The projection problem asks when the presuppositions of a part of a sentence become presuppositions of the whole. Karttunen's holes, plugs and filters are one answer. A query with a conjunction or a condition can keep an assumption or drop it, and a system has to know which.

Also discussed at https://plato.stanford.edu/entries/presupposition/

Presupposition failure occurs when an assumption of an utterance is not met. The communication breaks down, whatever answer follows. A system that treats a failed assumption as an ordinary query answers a question that should not be answered as asked.

If unanswerable queries failed mainly because the search went wrong, the first row would be small. It is the largest cause by far.

A factive verb, such as realise, regret or know, presupposes that its complement is true. Paul and Carol Kiparsky described the class in 1970. A query like why do people regret buying electric cars assumes the regret is real.

A lexical presupposition trigger is a single word that brings an assumption, such as stop, again or manage. It differs from structural triggers built into sentence form. Separating the two shows where in a query the assumption sits.

An informative presupposition introduces new information in the form of an assumption, as in I have to pick up my sister from the airport said to someone who did not know about the sister. The listener learns it without being told. Queries do this constantly, and the new information shapes what the person needs.

A cleft sentence has the form it was X that Y, and it presupposes that someone or something did Y. It isolates the new part from the assumed part. Queries phrased this way say which part the person already takes as settled.

Presupposition cancellation is the removal of an assumption by context, as in he did not stop smoking, he never smoked. Some presuppositions can be cancelled more easily than others. A person can take back an assumption in the next turn, and a system should notice.

A pragmatic presupposition is an assumption that belongs to the speaker, the belief they take to be shared, as Robert Stalnaker described it. It depends on the context of use, beyond the sentence. It is the form of presupposition closest to user intent.

A semantic presupposition is a condition a sentence itself places on its truth. It is a property of the sentence, independent of who utters it. The two kinds together separate what the words assume from what the person assumes.

Stalnakerian common ground is Robert Stalnaker's formal model of common ground as the set of propositions the participants accept for the purposes of the conversation. Each assertion updates the set. It gives shared assumptions a structure a system could track.

Also discussed at https://plato.stanford.edu/entries/assertion/

A counterfactual conditional considers what would be true if something false were true, as in if I had started earlier, would I have ranked. It presupposes that the antecedent is false. A query of this kind asks about a possibility and should be answered as one.

Implicit query intent is the purpose a query does not state, including the assumptions behind it. It lives beyond the surface words. Reading only the words misses the part the person took for granted.

Conversational search lets a person search in several turns, as a dialogue. Assumptions can be stated, questioned and corrected along the way. A single query cannot do that, and a conversation can.

A clarification question is a question a system asks back to resolve what a person means. For a false or unclear presupposition, it is the only honest reply. It is also what a publisher can ask on a page, which is the idea of Supervised Search.

Also discussed at https://arxiv.org/abs/1907.06554

The question under discussion is a model by Craige Roberts in which every utterance addresses a question the conversation is trying to answer. Questions form a tree of larger and smaller goals. A query is a node in such a tree, and its place decides what counts as an answer.

Epistemic stance is the way a speaker signals how certain they are, through words such as surely, maybe or I think. A query can be firm or tentative about its assumption. The same words with a different stance need a different answer.

Pragmatic failure is a misunderstanding that comes from mismatched assumptions, beyond wrong words. Jenny Thomas described it in 1983. A result that fits the keywords and misses the assumption produces exactly this failure.

A task-oriented dialogue system helps a person complete a goal, such as booking or troubleshooting, through conversation. It has to track what the person assumes and what is still open. Presupposition handling is part of getting the task done.

Described in a Google patent: https://patents.google.com/patent/US8606568B1/en

Doxastic commitment is the commitment a person takes on to believing something. Accepting a presupposition as true and merely going along with it for the moment are different commitments. A searcher may entertain an assumption without holding it.

An assertion is the main claim an utterance puts forward, as distinct from what it presupposes. Answers address the assertion by default. For a query, the presupposed background can matter more than the asserted question.

A contextual defeater is information in a situation that undermines a belief that would otherwise be justified. It comes from epistemology. A new fact can defeat the assumption a query rested on, and the answer has to change with it.

An epistemic modal, such as might, must or probably, expresses how possible or necessary something is given what a speaker knows. It marks a query as tentative. Tentative queries and certain ones should not receive the same confident answer.

A query with a false or narrow assumption leads a system to answer inside that assumption. The person who typed headache tablet gets pharmacy results and never hears of acupressure. Every business whose offer lies outside the assumption of a query is invisible to that query.

Name the assumption and answer it as well as the question. A page can say what a query takes for granted, correct it where it is false, and ask back what the person actually wants. That is the clarification question moved onto the page, the idea of Supervised Search.