Can true information be irrelevant?

Yes, if it does not bear on the current task.

Can relevance change?

Yes, when the question, context, or constraints change.

Does an indirect reply prove intent?

No. Its interpretation can depend on revisable assumptions.

AI Glossary / Natural language and speech / Advanced

Relevance

Definition

Relevance is the property of bearing on a particular question, task, or choice of action. Information is relevant when it helps address what matters in that context. In conversation, relevance also helps people interpret why a speaker offers a particular response, including meanings that are suggested rather than explicitly stated.

What it means

Relevance is a relationship between information and a purpose. A delivery time may matter when choosing a service that can meet a deadline. The same detail may contribute little to a question about packaging color. Changing the purpose can change which facts deserve attention.
In AI, relevance influences what an agent observes, which evidence a retrieval system presents, and how a dialogue system interprets a reply. These applications share a concern with usefulness to the current task, but their operational definitions and evaluation methods can differ.
Relevant information is not automatically true, sufficient, or authoritative. A false statement can directly concern the question being asked. A correct statement can be irrelevant to it. Good assistance therefore requires both selecting pertinent information and assessing its reliability and coverage.

How it works

Start by identifying the question or objective. For an action-selection problem, specify constraints and the measure used to compare acceptable outcomes. Then determine which observations could affect feasibility, preference, or uncertainty about the choice. A feature's usefulness depends on that formulation.
A retrieval system may estimate relevance from lexical matches, learned representations, or judgments about whether a passage answers a query. A high matching score is evidence produced by a method, not a guarantee that the passage resolves the user's actual need.
In dialogue, people often assume that a contribution bears on the ongoing exchange. This expectation helps them infer an indirect answer. If someone asks whether a room is available at two and receives “It is booked until three,” the reply suggests an availability problem. Interpreting it still depends on assumptions about the same room, date, and booking arrangements.
Such conversational inferences are defeasible: additional context can change them. A system should distinguish the explicit content from its interpretation, and seek clarification when the uncertainty affects an important next step. Relevance alone does not establish the speaker's intent with certainty.

Three practical examples

A scheduling assistant selects opening hours and existing appointments as pertinent to finding a feasible meeting time.
A search tool distinguishes a document that mentions a product from one that answers the user's specific question about that product.
A dialogue assistant recognizes that a statement about a room booking may indirectly answer an availability question, while retaining uncertainty about unstated details.

Think of it this way

Think of packing for a particular activity. An item earns space because it serves the activity's needs. Changing the activity changes the useful selection. The analogy also has a limit: leaving something out does not make it worthless in every setting, and including it does not prove it is dependable.

Worked example

Consider a delivery-choice exercise. Option A costs 12 units and arrives at 14:40. Option B costs eight units and arrives at 15:20. Assume these times and prices are certain, both options otherwise satisfy the requirements, and the objective is to minimize cost subject to arrival by a stated deadline.
With a 15:00 deadline, A arrives 20 minutes before the limit and B arrives 20 minutes after it. Only A is feasible, so the rule chooses A at a cost of 12 units. The arrival times and deadline are relevant because together they determine which option qualifies.
Now change only the deadline to 16:00. A arrives 80 minutes before the limit and B arrives 40 minutes before it. Both are feasible. The cost comparison now selects B, saving four units relative to A. B is one-third cheaper than A because 4/12 is 33.3%.
Suppose A uses blue packaging and B uses green packaging. Under the stated objective and assumptions, packaging color does not affect the choice. This conclusion is conditional on the task definition; it is not a universal claim that color never matters.
Add a new requirement that the package must be blue. With the 16:00 deadline, B now fails that requirement, so A becomes the selected option again. A previously unused attribute has become relevant because the constraints changed.
Finally, consider a system that retrieves the true statement “Option B costs eight units” for the original 15:00 request and recommends B immediately. The price statement is pertinent to cost, but insufficient for the decision. Omitting the arrival constraint produces the wrong recommendation despite using a relevant and accurate fact.
These invented numbers illustrate task-dependent relevance. Real planning also requires uncertainty assessment.

Illustrative case study

Fictional example: RoomReply answers “Can we use Room Cedar at two?” with a paragraph describing the room's equipment. The paragraph contains correct facts, but misses the availability question.
Reviewers change the evaluation so that responses must address the requested room and time. When the booking record says “occupied until three,” the assistant explains the apparent conflict and checks the date if it is missing. Equipment details remain available for a later suitability question. This scenario illustrates an evaluation change, not a measured performance improvement.

Common misconceptions and mistakes

Keyword overlap does not by itself establish relevance. A passage can repeat the query's words while discussing a different issue. Conversely, an answer can be relevant without repeating those words.
Another mistake is treating relevance as equivalent to truth or completeness. Selecting a pertinent fragment does not justify ignoring constraints elsewhere. In conversation, a plausible inference about why someone said something should not be presented as their explicit statement. Relevance judgments also need a defined task rather than an unexplained universal score.

Business use

Evaluate whether an assistant addresses the user's actual question and includes the constraints needed for the decision. Separate relevance judgments from checks of factual accuracy and evidence quality. Include examples where apparently useful information becomes misleading because a deadline, exception, or qualification is omitted.

Check your understanding

Which option meets the 15:00 deadline?

Answer: A.

Why does B win with a 16:00 deadline?

Answer: Both qualify, and B costs less.

When does packaging color matter?

Answer: When a requirement makes it affect eligibility or preference.

Related terms

Information retrieval
Performance measure
Evidence
query
context
conversational implicature
dialogue state
task environment

Learn next

Compare relevance, factual accuracy, and completeness for several answers to the same question. Then change one constraint and reconsider which evidence matters.

Evidence

Sources

Russell/Norvig, AIMA, Intelligent Agents, section2.3: relevance depends on action choice and performance measure (consulted): https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap02.pdf Jurafsky/Martin, Speech and Language Processing, chapter 26, conversational relevance (consulted): https://web.stanford.edu/~jurafsky/slp3/26.pdf

Previous: Relaxed sequential Next: Reliability