AI Glossary / Natural language and speech / Advanced
Reformulation
Definition
Reformulation is expressing an earlier message again in a different form. In dialogue, it often means proposing a paraphrase or summary of another speaker's contribution to display or check understanding. The wording changes, but a faithful reformulation preserves the relevant meaning. This entry focuses on conversational reformulation rather than mathematical reformulation or search-query rewriting.
What it means
If a colleague says, “The file is ready, but I still need to check the figures,” a listener might respond, “So the document is prepared, with the figures still awaiting review.” The response shows which interpretation the listener has taken from the original statement.
That makes reformulation more informative than a bare acknowledgment such as “okay.” It exposes an interpretation that the other speaker can accept, reject, or refine. However, fluent restatement is not proof of understanding. A response that changes “still need to check” into “have checked” reverses a consequential part of the message.
A reformulation can also be selective. Summarizing a long contribution requires choosing which information to preserve. Those choices should reflect the conversational purpose without adding unsupported commitments or hiding qualifications that matter.
How it works
A participant identifies the earlier contribution, interprets its content in context, and produces a differently worded version. The next response can provide evidence about whether that interpretation was acceptable. A correction should update the conversation's working interpretation rather than remain an ignored side comment.
Dialogue annotation schemes can define reformulation more narrowly than everyday usage. The Switchboard-DAMSL manual, for example, has a summarize-reformulate category for a speaker's proposed paraphrase or summary of another speaker's talk. Its coding distinctions separate this function from simple repetition and other responses. Such labels belong to that annotation scheme; they are not universal names for every possible act of rewording.
Function depends on context, not sentence shape alone. A statement beginning “So you mean...” may invite correction even without ordinary interrogative word order. Conversely, starting with so does not automatically make an utterance a reformulation. The relationship to prior talk must be examined.
Three practical examples
- A support assistant restates a customer's description of a recurring fault, preserving when it occurs and what has already been tried.
- A meeting participant summarizes a proposal before the group discusses it, making the interpretation visible without treating the summary itself as agreement.
- A writing tutor restates a learner's intended point in clearer language, distinguishing improved expression from an added claim the learner did not make.
Think of it this way
Think of reformulation as redrawing a sketch to show what you understood. A clearer drawing can reveal a misunderstanding that a nod would conceal. But if the redraw adds a door that was never in the original, its neatness does not make it faithful.
Worked example
Consider this original teaching instruction: “Prepare a draft for six reviewers by Friday, but do not send it.” Assume the document is already identified. The instruction contains four relevant constraints: prepare a draft; use six as the reviewer count; meet Friday as the deadline; and do not send the draft.
Candidate A says, “You need a draft for six reviewers ready by Friday, and it should remain unsent.” Under the stated interpretation, all four constraints are preserved.
Candidate B says, “Send the final version to six reviewers on Friday.” It retains the number six, changes draft to final version, changes a deadline into a particular sending date, and reverses the prohibition on sending. Sharing several topic words does not make it an adequate reformulation.
Candidate C says, “Prepare a draft for six reviewers by Friday.” It preserves the first three constraints but omits the instruction not to send. This is less explicitly contradictory than Candidate B, yet the omission matters if the reformulation becomes the sole instruction passed to another person or system.
A simple teaching checklist gives A four preserved constraints out of four, B one out of four, and C three out of four: 100%, 25%, and 75%. This local rubric is not a universal metric. The errors are not equally consequential, and a 75% score must not be interpreted as permission to proceed despite a missing prohibition.
Now suppose the user responds to Candidate A, “Six reviewers is right, but make that Thursday.” The count remains six and the deadline changes to Thursday. The correction does not cancel the unsent condition. Maintaining the updated interpretation requires applying the specific change while retaining the constraints that were not revised.
The example also separates understanding from execution. Restating a request accurately does not mean that the draft has been prepared, reviewed, or sent.
Illustrative case study
Fictional example: BriefRelay turns long project discussions into short handoff messages. Its initial summaries sound clear, but reviewers discover that phrases such as “tentatively,” “after review,” and “do not distribute” sometimes disappear.
The team evaluates reformulations against the original conversation, looking separately for preserved constraints, unsupported additions, and lost qualifications. It also tests whether subsequent corrections update only the intended details. The scenario illustrates a review method; it does not report a measured accuracy gain or claim that a checklist replaces contextual judgment.
Common misconceptions and mistakes
Reformulation is not necessarily repetition with synonyms. It may condense several statements or make an interpretation explicit. That flexibility also creates opportunities for distortion: negation, uncertainty, conditions, and participant roles can change even when many words remain similar.
Another mistake is treating a proposed summary as accepted before the other person responds. Silence does not establish that every detail was endorsed. Agreement that a restatement captures an idea also need not mean agreement with the idea itself, approval of a separate action, or confirmation that the action has been completed.
Business use
Reformulation can help with support handoffs, meeting summaries, and requirements discussions. Evaluate fidelity to constraints as well as readability. Preserve access to the original contribution so disputed interpretations can be checked. When an assistant uses a reformulation to update its task state, test corrections and partial acceptance rather than only uncomplicated confirmations.
People also ask
Must reformulation be shorter?
No. It can expand or condense the earlier message.
Is it always a question?
No. A declarative restatement can still invite a check of understanding.
Does a faithful paraphrase prove task completion?
No. Understanding and execution are different.
Check your understanding
Which candidate preserves all four constraints?
Answer: Candidate A.
What crucial detail does Candidate C omit?
Answer: The instruction not to send the draft.
After the Thursday correction, is sending permitted?
Answer: No. That condition was not changed.
Related terms
- Grounding
- Dialogue act
- Clarification
- Paraphrase
- summarization
- acknowledgement
- conversational repair
- semantic fidelity
Learn next
Compare several restatements of one message. Mark preserved claims, omissions, additions, and changes in certainty or negation, then test how each responds to a targeted correction.
More in Natural language and speech
- Referent - A referent is the person, object, event, or other entity that someone refers to through an expression in context.
- Referential density - Referential density is the degree to which participants in discourse are expressed overtly rather than left unexpressed and recovered from context.
- Referring expression generation - Referring expression generation is the task of producing language that helps a listener or reader identify an intended entity.
- Regular expression - A regular expression is a pattern notation used to describe and match sets of strings.
- Relatedness - Relatedness is the degree to which words, concepts, or other items are meaningfully connected.
- Relation extraction - Relation extraction is the task of identifying semantic relationships between entities mentioned in text and representing those relationships in a...
Sources
Jurafsky, Shriberg, and Biasca, Switchboard-DAMSL Coders Manual, section 6.2.2.4, summarize-reformulate and understanding checks: https://web.stanford.edu/~jurafsky/ws97/manual.august1.html
Previous: Reflexive Rule Next: Refutation