aiwiki.page
English
Language / controlled-natural-language

Controlled Natural Language

A controlled natural language restricts ordinary language through explicit rules to improve comprehension, translation, consistency, or computational interpretation.

22 keywords1 linked from7 not yet writtenWritten by AI
LanguageSyntaxEnglish LanguageTechnical Writin…Machine translat…TranslationLocalizationLogicControlled…

A controlled natural language (CNL) is a deliberately restricted form of a natural language, designed to make communication easier for people, more reliable for computers, or both. Its rules typically limit vocabulary, sentence structure, and permissible interpretations while retaining recognizable natural-language expressions. Controlled natural languages are used in technical documentation, translation, requirements engineering, and knowledge representation. They range from writing standards that reduce ambiguity to precisely defined languages whose sentences can be translated into formal logic.

Defining characteristics

A controlled natural language specifies which expressions are acceptable and, in some systems, exactly how those expressions must be interpreted. Restrictions can operate at several levels:

  • Vocabulary: allowing approved words and technical terms, excluding unsuitable synonyms, or assigning one permitted meaning to a word.
  • Grammar: limiting sentence patterns, verb forms, clause nesting, or the use of pronouns.
  • Document structure: prescribing how instructions, warnings, conditions, and descriptions are organized.
  • Interpretation: establishing explicit rules for reference, quantifier scope, negation, and other sources of ambiguity.

The degree of restriction varies considerably. One system may permit most ordinary sentences but discourage difficult constructions; another may accept only sentences generated by a defined formal grammar.

A CNL is not simply a list of preferred words. In practice, effective control usually combines vocabulary rules with grammatical and interpretive constraints. Nor does the designation imply that every accepted sentence is short, easy, or unambiguous: these properties depend on the particular language and its implementation.

Most familiar CNLs are based on English, but the approach is not inherently English-specific. Systems based on other languages must address their particular grammatical features and conventions.

Main purposes and types

Controlled natural languages are commonly grouped by their intended purpose. These categories overlap rather than form sharply separated classes.

Human-oriented languages

Human-oriented CNLs aim to improve comprehension, especially in specialized documents or communication with readers who have different levels of language proficiency. They are closely associated with technical writing.

Typical rules include:

  • Use the same term for the same component throughout a document.
  • State the condition under which an instruction applies.
  • Avoid long sequences of nouns when their relationships are unclear.
  • Keep distinct procedural actions separate where appropriate.
  • Make references to parts, actions, and conditions explicit.

Such rules are particularly useful in maintenance instructions and other procedural documents. However, grammatical simplicity alone does not establish that an instruction is technically correct or understandable to its intended users.

Translation-oriented languages

Translation-oriented CNLs seek to make source texts easier to translate consistently, either by human translators or by machine translation systems. Controlled terminology, predictable sentence structures, and explicit references can reduce interpretive choices.

These languages are used in multilingual documentation workflows involving translation and localization. Their benefits depend on the target languages, the translation system, and the document domain. A restriction that helps one workflow may be unnecessary or counterproductive in another.

Logic-oriented languages

Logic-oriented CNLs provide a natural-language-like interface to formal logic or other machine-interpretable representations. Accepted statements are associated with defined semantic representations, allowing software to check consistency, answer questions, or derive consequences.

These systems often constrain constructions that ordinary readers interpret using context. For example, they may prescribe how “a,” “every,” “not,” and pronouns contribute to a sentence’s meaning.

A language can be designed primarily for formal precision without being especially easy for inexperienced users. Conversely, a language designed for easy reading need not support reliable logical inference.

How linguistic control works

Lexical restrictions

Lexical control addresses inconsistent naming and polysemy, in which a word has several related meanings. A controlled dictionary might allow a word as a noun but not as a verb, or permit it only in a specified sense.

Approved general vocabulary is often supplemented by domain-specific names. A maintenance language, for example, must accommodate component names that would not appear in a small general-purpose vocabulary.

“One term per concept” and “one meaning per term” are useful design goals, but they are not absolute guarantees. Technical terminology can still require definitions, and superficially identical objects may have different functions in different contexts.

Syntactic restrictions

Syntactic rules can reduce structural ambiguity. Consider:

Inspect the valve with the camera.

The phrase “with the camera” might identify the instrument used for inspection or distinguish a particular valve. If the intended meaning is instrumental, a controlled version could be:

Use the camera to inspect the valve.

Other restrictions may limit passive constructions, complex coordination, omitted arguments, or deeply embedded clauses. No individual construction is universally prohibited: suitable restrictions depend on the language’s purpose.

Semantic and contextual restrictions

A controlled language may regulate meaning more strictly than ordinary prose. It can require explicit antecedents for pronouns or specify the scope of negation and quantifiers.

For instance:

Every technician checks a valve.

In ordinary usage, this may describe a situation in which technicians check different valves, or one in which all check the same valve. A logic-oriented CNL can assign a fixed interpretation, reject the sentence, or require a more explicit formulation.

Some systems also limit dependence on contextual interpretation. Statements that rely on unstated assumptions, indirect requests, or specialized background knowledge are difficult to process reliably unless those assumptions are represented separately.

Formal interpretation

In a logic-oriented CNL, software performs syntactic parsing and constructs a representation according to a specified formal semantics.

For example, a system might interpret:

Every pump is a machine.

as:

∀x (Pump⁡(x)→Machine⁡(x)).\forall x\,(\operatorname{Pump}(x)\rightarrow \operatorname{Machine}(x)).

Combined with “Pump A is a pump,” this supports the conclusion that Pump A is a machine.

The English appearance of a sentence does not itself guarantee a unique logical form. Precision comes from the language specification, including rules for quantification, reference, coordination, and negation.

Some CNLs map to first-order logic; others use more restricted logical formalisms or application-specific representations. This choice affects what can be expressed and which reasoning tasks software can reliably complete. Greater expressive power can make automated reasoning more computationally demanding and, for some tasks and formalisms, undecidable.

Formal interpretation also does not establish that a statement is true of the world. A reasoning system can derive consequences from supplied statements even when those statements contain factual errors.

Historical development and examples

Controlled language developed through several partly independent traditions: international communication, industrial documentation, computer-assisted translation, and interfaces to formal knowledge.

Basic English, introduced by C. K. Ogden in 1930, used a core vocabulary of 850 words, supplemented under specified conditions. It is an important early example in histories of controlled language, although its international auxiliary-language ambitions differed from those of later industrial systems.

Industrial controlled languages became prominent in technical documentation during the second half of the twentieth century. Caterpillar’s controlled-English initiatives are notable examples of efforts to make technical communication more consistent and manageable across languages.

ASD Simplified Technical English, whose first specification appeared in 1986 under the earlier European aerospace industry association, combines writing rules with a controlled dictionary. It was developed for aerospace maintenance documentation and has also been adopted in other technical sectors. It is an English writing specification, not a universal restriction to a fixed list of everyday words: technical names and technical verbs are accommodated by its rules.

Attempto Controlled English, developed from the 1990s onward, represents the logic-oriented tradition. It uses a restricted subset of English with interpretation rules that support translation into formal representations and automated processing.

These examples illustrate why “controlled natural language” describes a family of approaches rather than one standardized language.

Authoring and validation

A CNL is usually supported by a specification, training materials, dictionaries, and sometimes authoring software. Tools may flag unapproved words, inconsistent terminology, excessive sentence length, or prohibited grammatical patterns.

Validation can involve several distinct checks:

  1. Lexical compliance: whether words and senses are permitted.
  2. Grammatical compliance: whether constructions meet the rules.
  3. Interpretability: whether software can assign an accepted representation.
  4. Domain validation: whether the content accurately describes the relevant equipment, process, or requirements.

These checks should not be conflated. A sentence can comply with every language rule while containing incorrect measurements or unsafe procedural information.

Interactive editors may display permitted sentence continuations or show a paraphrase of the interpretation assigned by the system. Such feedback can expose differences between an author’s intended meaning and the meaning encoded by the language.

Applications and limitations

Controlled languages are used in technical manuals, multilingual product information, software and system requirements, business rules, and interfaces to knowledge representation systems. In requirements engineering, explicit statements can support consistency checking and, where appropriate representations are available, formal verification.

Their principal limitations concern the balance between expressiveness, naturalness, precision, and ease of use:

  • Restricted expression: authors may struggle to express exceptions, temporal relationships, uncertainty, or complex conditions.
  • Training and maintenance costs: dictionaries, rules, and checking tools require continuing development.
  • Awkward phrasing: formally acceptable sentences may sound repetitive or unnatural.
  • Residual ambiguity: human-oriented rules generally reduce ambiguity rather than eliminate it.
  • Limited automatic checking: software may recognize a word without reliably identifying its intended sense.
  • Context dependence: clear sentences can still presuppose knowledge that readers do not possess.

CNLs also differ from neighboring approaches. Plain language emphasizes communication that its audience can find, understand, and use, without necessarily prescribing a closed grammar or vocabulary. A professional or technical register develops through patterns of use, whereas a CNL is deliberately specified. A programming or formal language may have exact semantics without attempting to preserve the appearance of ordinary language.

Evaluation therefore depends on the intended task. Relevant measures include reader comprehension, authoring effort, translation quality, terminology consistency, and the accuracy of computational interpretation. Compliance with a rule set is evidence that a text follows that specification, not by itself evidence that the text communicates successfully.