Can Controlled Languages Scale to the Web? A 2006 Study of Restricted Natural Language

Take a plain sentence that a person would read without trouble and ask a machine to reason from it. The machine has to decide who or what each phrase refers to, which word sense is meant, and what the speaker is committing to. “Can Controlled Languages Scale to the Web?” was presented at CLAW 2006, the 5th International Workshop on Controlled Language Applications. Its author, Jonathan Pool, asked whether restricted, precise versions of ordinary languages could solve this problem for content on the web.

The problem: precision without knowledge engineers

The Semantic Web is a vision in which web pages carry meaning that machines can reason with. In its most prominent form, ordinary authors write such pages with off-the-shelf software. They are not knowledge engineers. The paper considered one way to get there. Controlled natural languages look informal but are formally defined, and they might give authors semantic precision without special training.

The paper added a multilingual twist. If equivalent controlled varieties existed for every language, the Semantic Web could work in all of them, not only in English. The paper called this a panlingual web.

What a controlled language is

A controlled natural language is an engineered subset of a natural language. Its grammar and vocabulary are restricted in a systematic way, which reduces both ambiguity and complexity. Some are designed so that people read them without misunderstanding. Others are machine-oriented: they can be translated automatically, and often deterministically, into a formal language such as first-order logic.

The paper stressed a further demand. Unlike a natural language, a controlled language for this purpose must require precision from the author, not merely permit it. Its documentation would have to show authors how to express things such as illocutionary force (whether a sentence asserts, asks or commands), evidentiality, probability, time, aspect, mood and number.

The approach: survey, then test

The study first asked whether such languages exist. It found 41 attempts to define written controlled varieties. They covered English, Esperanto, French, German, Greek, Japanese, Mandarin, Spanish and Swedish. Only 4 of the 41 met two conditions: they were designed for multiple domains and genres, and they were documented well enough to evaluate.

Those candidates were then checked against a sample of statements taken from the web on health and human rights. Test sentences of this kind are the reason other sites later referred to the work when evaluating controlled languages.

What it found

• Each of the evaluated languages showed limited expressivity or limited precision on the sample.
• The most expressive one avoided structural ambiguity, where a sentence can be parsed in more than one way.
• That same language still allowed semantic ambiguity, where the meaning stays open. The paper noted this could frustrate both human and machine comprehension.

The conclusion was cautious. A practical, Web-scale controlled language remained undemonstrated but unrefuted. The study did not say the idea had failed. It said no one had yet shown it working across the range of content the web contains.

How the field looks afterward

The paper became a standard reference in later surveys. A 2010 survey of controlled languages for knowledge representation, presented at the COLING conference and available through the ACL Anthology, cites it. That survey covers machine-oriented languages that can express software specifications, formal ontologies, business rules, and legal and medical regulations. Its authors list requirements for an ideal language of this kind. It needs a well-defined syntax and a precise semantics, given by an unambiguous mapping into a logic-based representation. It should also look as natural as possible.

A collaborative report on the properties and prospects of controlled natural languages also cites the paper. That report poses questions about goals, design, linguistic aspects, evaluation and tools, in an attempt to structure the field and encourage systematic discussion. A later survey and classification of controlled languages found limited tool support and a lack of validation through case studies for many approaches. That echoes the concern behind the 2006 question: whether these languages have been tested at the scale and variety the web demands.