Controlled natural language

Overview

A controlled natural language is a natural language with extra rules that remove ambiguity or complexity. The survey names three aims for those rules: easier human comprehension, translation, and standing in for a formal notation 1507.01701 §2.3. It defines a CNL by four properties: it is based on exactly one natural language; it is more restrictive than that base, mainly in lexicon, syntax or semantics; it keeps most of the base language's properties so that its speakers still understand it; and it is built on purpose rather than arising by natural change 1507.01701 §2.1.

Many names are in use, such as "simplified", "structured" or "restricted" English, and they do not agree on what the languages are 1507.01701 §1. Search for those names as well as this slug.

The shared core of all variants is the same: the text reads as the base language, and the restrictions are fixed in advance. What differs is which restriction the designer prioritises. The survey covers only designs with English as the base language, and it gives no algorithm for building one. A designer still makes most decisions alone 1507.01701 §5.3.

Variants

The survey has no implementations. Its catalogue of 100 English-based CNLs 1507.01701 §4 is grouped below by the aim each language was built for. A language can fall into more than one group, because any two of the survey's nine environment properties may overlap 1507.01701 §2.3.

Comprehension-oriented (written, human readers)

Built so that people read the text quickly and correctly, often in one domain. The survey's averages over the 45 languages tagged for comprehension are high for expressiveness (4.3) and naturalness (4.7), and low for simplicity (1.2), on its 1 to 5 scale 1507.01701 tab. 3. Examples the survey describes in detail are ASD Simplified Technical English and FAA Air Traffic Control Phraseology 1507.01701 §4.1. Card: A Survey and Classification of Controlled Natural Languages.

Translation-oriented

Built so that a text can be translated, by machine or by a human post-editor. Across the 22 languages tagged for translation, the averages are E 4.8, N 5.0 and S 1.1 1507.01701 tab. 3. The survey reports a five-to-one gain in translation time for one such language, MCE, and a 20 percent reduction in post-editing time for another, CLCM 1507.01701 §5.5.

Formal-representation-oriented

Built so that a machine can process or execute the text, or so that the text maps onto a formal notation. The 54 languages tagged this way average P 4.4, E 2.3, N 3.8 and S 3.2 1507.01701 tab. 3. They are the most precise group and the most readable as a formal language. Attempto Controlled English is described in detail 1507.01701 §4.1. For the step from text to a formal representation, see semantic parsing.

Spoken

Rare in the catalogue. Over 90 percent of the CNLs are intended to be written, and only seven are intended to be spoken, six of them governmental 1507.01701 §5.2.

When to use which

The survey's scale is PENS: precision (P), expressiveness (E), naturalness (N) and simplicity (S), each with five classes 1507.01701 §3. The averages below describe the catalogue, not the quality of any single design. They do not weight by maturity or popularity 1507.01701 §5.2.

The survey gives evaluation evidence but no general rule for choosing. Its findings depend on domain, users and design quality 1507.01701 §5.5. Pilot-test the chosen design with its real readers before you commit.

Implementation sketch

The survey has no algorithm. The sketch below is our own structure for the PENS record and the CNL test, built from the definitions. It is not the paper's code.

# Store one language on the PENS scale  [arxiv:1507.01701 §3]
# Input:  name, four classes 1..5, environment property codes  [§2.3]
# Output: record; is_cnl() applies the naturalness threshold

def pens_record(name, precision, expressiveness, naturalness,
                simplicity, props):
    # Each dimension has five classes, 1 (low) to 5 (high)  [§3]
    for value in (precision, expressiveness, naturalness, simplicity):
        assert 1 <= value <= 5
    # Codes from the nine environment properties  [§2.3]
    return {"name": name,
            "pens": (precision, expressiveness, naturalness, simplicity),
            "props": set(props)}

def is_cnl(record):
    # Naturalness N3 or higher counts as a CNL  [§3.3]
    # pens order is P, E, N, S, so index 2 is N
    return record["pens"][2] >= 3

Contradictions

None across cards. The wiki has one card on this concept. Inside the survey, the evaluation studies disagree. For AECMA Simplified English, two studies report a significant comprehension gain and a third reports no significant result. The survey reports mixed usability and comprehension results for Rabbit 1507.01701 §5.5. The survey does not resolve these, and neither does this page.

Papers

• A Survey and Classification of Controlled Natural Languages: defines a CNL, proposes the PENS scale, and surveys 100 English-based CNLs. It has no method or experiments, and its literature runs to about 2013.

Only one card covers this concept. Other CNL papers are not yet compiled in the wiki, so their content is not used here.