Source-linked AI summary
A Survey and Classification of Controlled Natural Languages
Tobias Kuhn
TL;DR
CNL lacks agreed characteristics because diverse languages emerged under different names and vary widely in ambiguity, precision, expressiveness, and scope. This paper proposes a common definition and PENS-based classification, surveys 100 English-based CNLs, and finds them distributed between natural and formal languages.
Problem
Controlled natural language remains a fuzzy term because approaches arose across different environments and disciplines, used different names, and exhibit substantial variation in properties.
Method
The paper establishes a common definition and models CNLs using the PENS scheme, which classifies inherent language properties by precision, expressiveness, naturalness, and simplicity.
Results
100 English-based CNLs are surveyed, and the languages form a scattered but connected cloud between natural languages and formal languages.
Takeaways & Limitations
The survey provides common terminology and a model for describing CNLs, while showing that existing languages pursue different goals including comprehensibility, translatability, and formal representation.
Takeaways & Limitations
PENS simplifies continuous, potentially multidimensional properties into discrete dimensions and therefore provides only a rough measure of concepts such as expressiveness.
Abstract
from arXiv · showhide
What is here called controlled natural language (CNL) has traditionally been given many different names. Especially during the last four decades, a wide variety of such languages have been designed. They are applied to improve communication among humans, to improve translation, or to provide natural and intuitive representations for formal notations. Despite the apparent differences, it seems sensible to put all these languages under the same umbrella. To bring order to the variety of languages, a general classification scheme is presented here. A comprehensive survey of existing English-based CNLs is given, listing and describing 100 languages from 1930 until today. Classification of these languages reveals that they form a single scattered cloud filling the conceptual space between natural languages such as English on the one end and formal languages such as propositional logic on the other. The goal of this article is to provide a common terminology and a common model for CNL, to contribute to the understanding of their general nature, to provide a starting point for researchers interested in the area, and to help developers to make design decisions.
1. Introduction
Controlled natural languages are engineered, restricted forms of natural language whose varied names, origins, and properties have made the category difficult to define. This article addresses that problem by surveying existing CNLs and introducing a shared classification framework.
- CNLs have been described with many attributes and names, including controlled, simplified, technical, structured, and basic.
- No general agreement exists on CNL’s characteristic properties because these approaches emerged across different environments, disciplines, and decades.
- CNLs vary widely in ambiguity, expressiveness, formal precision, and grammatical completeness, making their fundamental properties difficult to characterize.
- The article aims to clarify CNL’s nature, establish common terminology and a common model, support researchers, and help developers make design decisions.
- CNL research is relevant to computational linguistics through lexical analysis, grammar checking, ambiguity detection, machine translation, and computational semantics.
- The survey focuses on English-based CNLs, while its proposed classification scheme is intended to apply beyond English.
2. Background
The paper defines CNL as an engineered, restricted language based on one natural language while preserving much of its intuitive naturalness. It distinguishes CNL from related concepts and classifies approaches by their intended problems and environments.
- Definition: A CNL is a constructed language based on one natural language that restricts its lexicon, syntax, and/or semantics while preserving most natural properties.
- Definition: The definition excludes natural languages, Esperanto, and common formal languages because they fail one or more requirements concerning construction, linguistic basis, or intuitive understandability.
- Related terms: CNL differs from a sublanguage because CNLs are consciously defined, whereas sublanguages arise naturally within expert communities.
- Related terms: Controlled vocabularies standardize terms but do not specify how to combine them into complete grammatical sentences.
- Types and properties: The survey collects nine clear-cut properties because the many properties proposed in prior work mainly describe application environments and remain too fuzzy for strict categorization.
- Types and properties: CNLs are roughly divided into communication, translation, and formal-representation types, denoted C, T, and F.
3. PENS Classification Scheme
The PENS scheme classifies controlled natural languages along four dimensions—precision, expressiveness, naturalness, and simplicity—positioning them between natural and formal languages. Its five-class scales make vague properties comparable, but simplify continuous and potentially multidimensional characteristics.
- Dimensions: PENS condenses language properties into four dimensions: precision, expressiveness, naturalness, and simplicity.Ambiguity, predictability, and formality of definition map to precision; grammar modifications, understandability, and natural look-and-feel map to naturalness.
- Conceptual space: PENS places languages in a conceptual space between natural languages, which are expressive but complex and imprecise, and formal languages, which are simple and precise but unnatural and inexpressive.Each dimension is scaled from 1 to 5 using English and propositional logic as conceptual endpoints.
- Scope and limitations: The four dimensions are continuous or fine-grained, so representing them as single dimensions and five classes is an explicit simplification.The choice of five classes is described as somewhat arbitrary, balancing classification detail against strict, objective criteria.
- Precision: Precision measures how directly textual form determines meaning, ranging from context-dependent natural language to formally defined logic.The scheme distinguishes imprecise, less imprecise, reliably interpretable, and deterministically interpretable languages along this dimension.
- Expressiveness: Expressiveness describes the range of propositions a language can express, but unequal languages cannot always be objectively ranked in a total order.PENS therefore uses selected general expressiveness features rather than every possible feature.
- Interpretation: PENS describes a language’s nature rather than ranking its quality or usefulness, and optimal dimension levels depend on application, environment, and goal.Higher values are not necessarily better in practice because additional performance beyond an application’s needs may bring no benefit.
4. Languages
The survey catalogs English-based controlled languages using a common classification approach, while illustrating the range from highly restricted systems to practical technical languages. Its examples show recurring trade-offs among precision, expressiveness, naturalness, and enforceability.
- Survey scope: The survey restricts its main inventory to English-based controlled languages and reports exactly 100 languages, with classifications and descriptions also available as an online CSV table.It excludes CNLs based on other natural languages and adds natural English and propositional logic for comparison.
- Survey scope: The selected sample contains twelve influential, well-documented, or sufficiently distinct CNLs arranged roughly chronologically.The appendix provides the complete list with short descriptions.
- Sowa’s syllogisms: Sowa’s syllogisms use a small set of English sentence patterns that map directly to first-order logic, yielding an exact and comprehensive description.The language remains perfectly natural despite its simple structure, but supports only very simple sentences and one-place relations.
- Basic English: Basic English restricts grammar and vocabulary to 850 root words, including only 18 supported verbs, while retaining natural text flow across topics.Its lexical and grammatical restrictions increase precision, but informal grammar rules do not significantly reduce PENS complexity.
- E-Prime: E-Prime forbids every inflectional form of the verb “to be,” requiring speakers to rephrase otherwise permitted statements.The survey finds no considerable gain over full English in PENS precision or complexity, although rephrasing remains possible and natural, often at greater length.
- Industrial and technical languages: Caterpillar Fundamental English combined practical communication goals with rules for positive, short, uniform, and consistently named technical sentences.CFE was discontinued in 1982 because its basic guidelines were not enforceable; Caterpillar Technical English instead emphasized enforceable restrictions and reduced translation costs.
5. Analysis
The analysis places CNLs across a broad, continuous PENS space while identifying recurring relationships among their properties, goals, domains, uses, and historical development. It also reports evidence of benefits in comprehension, translation, and usability, while emphasizing that outcomes depend on context and design quality.
- 5.1 PENS Classes: 25 distinct PENS classes are observed, widely scattered across the conceptual space despite identifiable hotspots.Theoretically, 625 classes are possible, but the surveyed languages occupy a broad variety of the observed space.
- 5.1 PENS Classes: CNL classes form one connected cloud between natural English and propositional logic rather than separate clusters.This supports using CNL as a broad umbrella term and makes clean subdivision difficult.
- 5.1 PENS Classes: Precision and simplicity correlate positively (ρ = 0.90), whereas expressiveness and simplicity correlate negatively (ρ = −0.82).Naturalness and expressiveness also show a strong positive correlation (ρ = 0.77).
- 5.2 Properties: CNL goals divide roughly into comprehensibility or translatability versus formal representation, with type F languages favoring precision and simplicity over expressiveness and naturalness.A bit less than half target comprehensibility, about 22% target translatability, and another roughly half target formal representation.
- 5.2 Properties: More than 90% of surveyed languages are intended to be written, while only seven are intended to be spoken.Six of the seven spoken languages originated in government, and written languages have higher average PENS values in all four dimensions.
- 5.5 Evaluations: The survey reports improved comprehension, faster translation or post-editing, and positive usability results for several CNLs and tools.Reported examples include a five-to-one translation-time gain for MCE, three- or four-times-faster post-editing for PACE, and a 20% average post-editing reduction for CLCM.
6. Conclusions
The survey concludes that diverse controlled natural languages can be covered by one definition and positioned between natural and formal languages. Its classification model organizes their language properties and environments, while the survey provides resources for researchers and developers.
- Theoretical conclusions: Despite their diversity, controlled natural languages form a scattered but connected cloud between natural and formal languages.The paper presents this as support for viewing CNLs as more formal than natural languages but more natural than formal ones.
- Common terminology and model: The proposed model distinguishes application environments from inherent language properties and describes the latter using the four-dimensional PENS scheme.PENS classifies precision, expressiveness, naturalness, and simplicity on a discrete scale.
- Research resources: The survey presents a diverse sample of twelve important languages and a longer list of collected CNLs as a starting point for researchers.The model is also intended to help identify research focuses and relevant prior work.
- Design support: Survey data can guide developers toward existing CNL approaches and indicate whether a proposed usage is common, rare, or previously unaddressed.The data also expose typical trade-offs among precision, expressiveness, naturalness, and simplicity for design decisions.
- Field outlook: The field is dynamic and interdisciplinary, spanning small academic, industrial, and governmental niches that together form a substantial body of work.The conclusion connects this diversity to ongoing and future study of controlled languages.
Appendix A: Full List of English-based Controlled Natural Languages
The appendix lists English-based controlled natural languages spanning technical documentation, translation, queries, scientific publishing, reminders, and biblical texts. The entries illustrate varied restrictions, purposes, domains, and computational connections.
- Technical and industrial languages: Several entries target technical communication, including Airbus Warning Language, CTE, Avaya Controlled English, and Global English.Their documented aims or restrictions include readability, consistency, comprehension, translation, lexicon, grammar, semantics, and style.
- Application domains: Other entries support specialized applications, including biomedical querying, scientific assertions, automatic reminders, and biblical-text translation.BioQuery-CNL interfaces with answer set programming, AIDA supports RDF-based nanopublications, Atomate defines context-sensitive tasks, and EasyEnglish supports biblical texts and translation.
- Rule-based restrictions: The appendix includes languages defined through explicit rules, such as Bull Global English’s ten rules covering sentence form, voice, abbreviations, punctuation, and nomenclature.These examples show controls operating at lexical, syntactic, stylistic, and discourse levels.
- Computational representations: Some listed languages connect controlled English with formal or computational representations, including Lite Natural Language’s deterministic mapping to DL-Lite and Atomate’s mapping to RDF.These entries illustrate CNL use as an interface to logical formalisms or executable structures.