Is WYSIWYM the same as WYSIWYG?

No. WYSIWYG emphasizes the appearance of the final output. This WYSIWYM technique emphasizes feedback about the meaning represented in an underlying model.

Can the user change any wording freely?

Not necessarily. In the classic architecture, edits select semantic options through the feedback text. Other systems may add facilities, but those should be described explicitly.

Does a constrained interface guarantee correct knowledge?

No. It can restrict malformed or unsupported structures, yet a user can still select the wrong fact or misunderstand a relationship. Faithful feedback and review remain important.

AI Glossary / Data engineering and architecture / Intermediate

WYSIWYM (interaction technique)

Definition

WYSIWYM, meaning What You See Is What You Meant in this interaction technique, lets people edit structured knowledge through natural-language feedback generated by the system. Users select meaningful options associated with parts of the displayed text. Those choices update the underlying knowledge representation, and the system regenerates the text to show the revised meaning. The main object being edited is the knowledge, not the wording alone.

What it means

A domain expert may understand a business procedure without knowing the formal language used to represent it in a knowledge base. WYSIWYM makes that representation accessible through readable feedback and constrained editing choices.

For example, a sentence can contain a selectable phrase marking an unspecified participant or action. Choosing an option fills that part of the model. The next generated sentence reflects what the system now represents. This entry concerns the knowledge-editing technique associated with natural language generation, rather than every document editor that uses WYSIWYM to describe semantic formatting.

How it works

The system maintains a structured domain model, including incomplete information during editing. A language generator turns the current model into feedback text. Selectable spans, often called anchors, expose places where the user can add or change information.

An anchor offers choices compatible with the model's constraints. Selecting one changes the underlying structure, which may introduce further details to specify. The generator then produces updated feedback so that the user can inspect the represented meaning.

This loop differs from interpreting unrestricted typed prose after the fact. The original approach controls the available semantic edits and generates the language from those edits. Its usefulness depends on the domain model, available choices, and faithful generation of understandable feedback.

Three practical examples

1. Procedure authoring: A prototype lets an expert choose an action, its object, and an operating condition while generated instructions show the resulting procedure.

2. Knowledge-base maintenance: A subject specialist adds a relationship between two defined concepts through readable selections instead of editing a formal representation directly.

3. Multilingual documentation: A system generates supported language versions from the same structured content, allowing authors to review the shared meaning through their chosen feedback language.

Think of it this way

Think of building a sentence with a set of meaningful, connected pieces while a guide continually reads back what those pieces express. Changing a piece changes the underlying arrangement, and the read-back changes accordingly.

The analogy highlights feedback and constrained choices. Its limit is that a real system must represent relationships, references, and incomplete structures formally. A pleasing sentence is not sufficient: the text must accurately reflect the stored model, including what the author has not yet specified.

Worked example

Suppose a fictional authoring interface starts with the feedback: 'A person sends a document to someone.' The phrases for the sender, document, and recipient are editable anchors.

The author selects 'service coordinator' as sender, 'maintenance report' as document, and 'site manager' as recipient. The model now contains those three roles and their relationship to a sending action. The regenerated feedback becomes: 'The service coordinator sends the maintenance report to the site manager.'

If a timing condition remains unspecified, the interface should make that incompleteness visible rather than silently inventing 'every Friday.' Adding an approved weekly condition is a further semantic edit. The example illustrates the feedback loop, not a claim that one fixed set of words or interface controls is required by all WYSIWYM implementations.

Illustrative case study

Fictional case: Meaning Orchard maintains internal equipment instructions in a structured knowledge base. Experts previously submitted free-text changes that engineers manually converted into formal relationships, creating delays and misunderstandings.

The prototype interface exposes supported actions and objects through generated feedback. During a review, an author notices that a selected recipient applies to the entire procedure rather than one step. The team improves the feedback and editing choices so that the relationship's scope is visible before the update is accepted.

This case emphasizes representation and usability. It does not report measured productivity gains or suggest that constrained authoring can express every instruction an expert might need.

Common misconceptions and mistakes

A common mistake is assuming the technique means that the system understands any sentence the user types. The classic approach instead generates feedback from structured knowledge and offers controlled ways to change that knowledge.

Another mistake is confusing generated feedback with final document layout. An editing view can reveal incomplete content or provide anchors that would not appear in the finished publication.

Multilingual support also requires suitable generation resources for each supported language. A language-neutral model does not automatically supply flawless wording in every language. Authors still need ways to inspect meaning and identify gaps in the available domain concepts.

Business use

WYSIWYM is relevant where structured, reusable knowledge matters and domain experts need to edit it without learning its formal notation. Potential benefits include clearer feedback about represented meaning and multiple outputs from shared content.

The development effort includes designing the model, language generation, editing options, and usability testing. Evaluate whether users can express intended distinctions, detect incorrect relationships, and recognize incomplete information. For AI projects, the approach offers a useful lesson about making machine-readable meaning inspectable, even when a modern interface uses additional techniques beyond the historical architecture.

Check your understanding

An author changes a recipient through an anchor, and the sentence is regenerated. What is the primary edited object?

Answer: The underlying structured knowledge representation. The sentence is generated feedback showing the revised recipient relationship, rather than merely a manually rewritten string.

Related terms

Natural language generation
Knowledge representation
Ontology
symbolic authoring
controlled language
semantic authoring
user interface
WYSIWYG

Learn next

Design a tiny procedure model with an actor, action, object, and recipient. Write feedback for both complete and incomplete versions. Ask a reader to identify what is missing and what each edit would change before adding more complex relationships.

Sources

Sources: Power and Scott, WYSIWYM: knowledge editing with natural language feedback: https://aclanthology.org/W98-1437.pdf. Original 1998 paper read for the architecture and symbolic-authoring context.

Evans and Power, WYSIWYM—building user interfaces with natural language feedback: https://aclanthology.org/E03-2005.pdf. Original 2003 paper read for the feedback-and-anchor mechanism. This entry discusses the interaction technique and uses original non-medical examples.

Previous: Write - write conflict Next: X-11 decomposition