Is clarification always a question?

It often takes question form, but conversational repair can also use a brief repetition or other signal of non-understanding.

Should every uncertain detail trigger clarification?

The detail's relevance and consequences matter, along with existing context and user preferences.

Can the user clarify the assistant's output?

Yes. Clarification is a conversational function available to either participant, not only a system behavior.

AI Glossary / Natural language and speech / Advanced

Clarification

Definition

Clarification is a conversational act that seeks to resolve uncertainty about what another participant said or meant. In dialogue systems, it can target a misheard word, an ambiguous reference, an unclear intention, or an insufficiently specified detail. Its purpose is to improve shared understanding so the conversation can proceed on an appropriate interpretation.

What it means

If someone asks an assistant to “open the latest report”, the assistant may find two reports that fit. A useful clarification identifies that specific uncertainty: “Do you mean the sales report or the delivery report?” The question preserves what is already understood instead of restarting the request.
Clarification differs from simply asking for any new information. It addresses a problem in interpreting or completing the current exchange. It can overlap with confirmation when a proposed interpretation is checked, but understanding a request and obtaining permission to perform an action are separate matters. Resolving ambiguity does not automatically supply authorization that was otherwise missing.

How it works

A dialogue system maintains an account of the user's task and the information supplied so far. When several interpretations remain plausible, it can identify which unresolved distinction affects the next step. It then chooses a question that the user can answer and updates its interpretation from the response.
The question can be open, such as asking what a phrase means, or targeted, such as offering two known candidate objects. A confirmation question instead proposes a specific interpretation for acceptance or correction.
The response must actually be incorporated. Asking a precise question but then ignoring the answer does not repair the dialogue. Nor should silence be treated as if the user selected one of the options. The system should preserve uncertainty when an answer remains ambiguous.

Three practical examples

A document assistant finds two files named “budget”. It asks which department's file is intended, using distinguishing information already available in the file list.
A speech interface is unsure whether a user said fifteen or fifty. It asks about that quantity rather than requesting the entire instruction again.
A writing assistant receives “make this lighter”. It uses context to determine whether the user may mean a less formal tone, a paler color, or a shorter document, and asks only if the relevant distinction remains unresolved.

Think of it this way

Think of clarification as adjusting one uncertain coordinate on a map. If the destination street is known but the building number is unclear, asking for the building number is more useful than discarding the whole address.

Worked example

Consider an original hypothetical policy exercise about opening one of two internal reports. Assume the action is already permitted and the only uncertainty is which report the user means. The system assigns probabilities 0.6 to report A and 0.4 to report B. These probabilities are stipulated for the exercise, not calibrated measurements from a deployed model.
Suppose opening the wrong report incurs an inconvenience cost of ten units, while opening the correct report costs zero. Choosing A immediately gives expected cost 0.4×10=4 units. Choosing B gives 0.6×10=6 units, so A is the better immediate guess under these assumptions.
Now suppose a clarification question costs one unit of user effort and its answer identifies the intended report perfectly. Asking has expected cost one, lower than four. If the answer instead leaves a 10% chance of an incorrect selection, the expected cost becomes 1+0.1×10=2 units. It is still lower than four, but the benefit depends on the assumed reliability of the answer.
Change the cost of a wrong report to one unit while leaving the question's effort cost at one. The best immediate guess now costs 0.4×1=0.4 units in expectation, whereas even perfect clarification costs one. Under this simplified model, clarification no longer minimizes expected cost.
The calculation demonstrates a tradeoff between uncertainty, consequences, and interruption. It is not a universal rule for acting without clarification, and it does not override authorization requirements. Real decisions may also involve unequal error costs, more than two interpretations, user preferences, and answers that introduce new information.

Illustrative case study

Fictional example: ReportLane's assistant handles repeated requests for project files. Its initial recovery message is always “Please repeat your request.” Users repeat the same wording even when the actual issue is two matching filenames.
The team changes that case to a targeted question showing each candidate's project and date. The assistant keeps the requested action and updates only the file identity after the answer. Reviewers also test responses such as “neither” and descriptions that identify a third file. The scenario illustrates a more informative repair strategy; it reports no measured reduction in turns or improvement in user satisfaction.

Common misconceptions and mistakes

More questions do not automatically produce better understanding. Repeatedly asking about information already supplied can increase effort without resolving anything. A question can also be misleading if its choices exclude the user's intended answer while presenting themselves as exhaustive.
Model confidence is not necessarily a calibrated probability, and a fluent interpretation may still be wrong. Another mistake is bundling several uncertainties into an unclear yes/no question. Finally, confirmation of meaning, acknowledgment of receipt, and approval of an action should not be collapsed into one undifferentiated signal.

Business use

Clarification design matters in support assistants, document tools, and conversational workflows. Track whether questions resolve the intended uncertainty, whether answers are retained, and whether users must repeat themselves. Evaluate task success and user effort together. Review unnecessary questions as well as missing ones, and test corrections, alternative answers, and situations where no available option is correct.

Check your understanding

What makes the report question targeted?

Answer: It distinguishes the two candidate files while preserving the already understood action.

Why is the initial best-guess cost four units?

Answer: Report A is wrong with probability 0.4 and the error costs ten units.

Does selecting a report automatically authorize a separate action on it?

Answer: No. Resolving identity and authorizing an action are distinct.

Related terms

Dialogue act
Grounding
Ambiguity
Common ground
confirmation
dialogue state tracking
conversational repair
intent recognition

Learn next

Identify the exact uncertainty in several short dialogues and write a question that resolves only that uncertainty. Then compare open clarification, candidate selection, and confirmation, and test whether each answer updates the system's state without losing information already established.

Sources

Jurafsky and Martin, Speech and Language Processing, Appendix K, dialogue state and dialogue policy: https://web.stanford.edu/~jurafsky/slp3/K.pdf Columbia Games Corpus, question-function annotation guidelines, clarification and confirmation categories: https://www.cs.columbia.edu/speech/games-corpus/guidelines-question-function.php

Previous: Citation recall Next: Class