Can a language model perform it without a separate module?

Yes, a system can combine selection and wording. The conceptual distinction still helps identify whether an error concerns missing information or expression.

Is there always one correct selection?

No. Several selections may be reasonable for the same audience and goal. Evaluation should allow legitimate variation while identifying material omissions and unsupported additions.

AI Glossary / Natural language and speech / Intermediate

Content determination

Definition

Definition: Content determination is the natural language generation task of deciding what information a generated text should communicate. Often called content selection, it chooses relevant material before or alongside decisions about organization and wording.

What it means

What it means: A system may have access to far more information than a reader needs. Content determination decides which facts, comparisons, explanations or caveats belong in the output for a particular purpose and audience. It concerns the message, not just the choice of attractive sentences.
For example, a delivery summary might prioritize overdue orders, while a planning report might prioritize upcoming demand. Both can use the same underlying records but select different information. Selection also creates responsibility for omissions: a report can contain only true sentences yet mislead if it leaves out a qualification that changes their interpretation.

How it works

How it works: The system identifies candidate messages from its permitted evidence and evaluates them against the communication goal. Rules, learned models or combinations can assess relevance, importance and relationships among candidates. A length or attention budget may restrict how much can be included.
Candidates should not always be scored independently. One figure may need its denominator, comparison period or uncertainty statement to be understood correctly. The selected content then supports document planning and linguistic realization. These functions can be intertwined in a system, but it remains useful to evaluate whether the right information was selected separately from whether it was expressed fluently.

Three practical examples

Three practical examples:
1. An operations report generator selects unusual delays and unresolved incidents from a large event log. It retains the dates and scope needed to interpret each selected issue.
2. A customer-facing status message chooses the order's current state and next expected step. Internal diagnostic details may be irrelevant to that reader even though they remain useful to support staff.
3. A sports-summary experiment selects significant events and statistics from structured records. It checks whether the selected details form a coherent account rather than an unrelated list of individually prominent numbers.

Think of it this way

Think of it: Content determination resembles preparing the agenda for a briefing. The agenda decides which points deserve the audience's limited attention before someone crafts the final wording. The analogy is limited because automated selection needs explicit evidence and evaluation; a confident-looking shortlist does not establish that important context has been retained.

Worked example

Worked example: Imagine a toy report with space for three messages. Candidate A states that 90 of 100 orders arrived on time. Candidate B reports the same result as a 90% on-time rate. Candidate C says the previous comparable period's rate was 95%. Candidate D states that ten current orders were late. Candidate E explains that all ten late orders concern the same affected region.
Selecting A, B and D fills the allowance but largely repeats one numerical fact. Selecting A, C and E adds comparison and distributional context: the current rate is five percentage points below the previous rate, and the late orders are concentrated in one region. The figures alone do not establish why the change occurred.
Which set is best still depends on the audience and report purpose. If a customer asks only about one specific order, neither aggregate selection answers the question. This original example illustrates relevance, redundancy and context; it does not claim that a simple three-slot optimization captures every requirement of real reporting.

Illustrative case study

Illustrative case study: Fictional Elm Services generates weekly summaries from maintenance records. Its first prototype selects the most frequent event types and omits an infrequent but unresolved equipment failure. The resulting text reads smoothly but fails the managers' stated need to identify outstanding issues.
The team revises its selection criteria to consider unresolved status and operational relevance, then reviews both included and omitted events. It keeps wording evaluation separate so a stylistic improvement cannot hide a selection mistake. This invented case demonstrates task alignment, not a universal rule that rare events always outrank common ones or that one scoring function works for every department.

Common misconceptions and mistakes

Common misconceptions and mistakes: Content determination is not synonymous with summarizing an existing article; it can select from structured records or other evidence. It is also different from sentence ordering, although the choices interact. High frequency does not necessarily mean high importance, and a large numerical change does not automatically justify a causal explanation. Finally, the ability to generate a plausible statement does not make that statement an eligible content candidate. Its evidential basis must be established.

Business use

Business use: Specify the audience, purpose, mandatory information and acceptable omissions before evaluating a report generator. Keep test cases where a caveat, denominator or unresolved exception changes the reader's interpretation. Compare the selected content with the underlying evidence, including material the system omitted. Where editors provide feedback, distinguish requests for different facts from requests for clearer wording. Version the selection policy when reporting priorities change so readers can understand why successive reports emphasize different things.

Check your understanding

Check your understanding: A report includes a percentage but omits that it was calculated from only two observations. Is that solely a wording problem?

Answer: No. The omitted denominator or sample-size context is a content-selection issue that can change how readers interpret the percentage.

Related terms

Natural language generation
content selection
document planning
discourse planning
data-to-text generation
sentence aggregation
linguistic realisation
factual consistency

Learn next

choose ten facts from a small fictional dataset and draft three-message briefings for two different audiences. Explain each inclusion and omission. Then rewrite one briefing without changing its selected facts to see how content decisions differ from wording decisions.

Sources

Sources: Barzilay and Lapata, Collective Content Selection for Concept-To-Text Generation, 2005: https://aclanthology.org/H05-1042.pdf . Reiter and Dale, Building applied natural language generation systems, author publication listing: https://web.science.mq.edu.au/~rdale/publications/index.html . The latter paper's PDF endpoints timed out; its indexed description supplies terminology only. The order-count example and Elm Services case are original illustrations, not results from either paper.

Previous: Content-based image retrieval Next: Content format