Benefits

Benefits

Constraining English to a single, fully specified dialect is more useful than surface consistency alone would suggest.

Most natural-language systems treat English as an open mix of regional variants, spelling systems, punctuation habits, and near-synonyms. That variation is noise for machines. Eliminating it — fixed spelling, fixed number formats (ISO 80000-1 + ISO 8601), fixed preferred terms, epicene they as the unmarked third-person pronoun, CGEL grammar constraints, consistent punctuation — yields something closer to a formal language that still reads as English.

Regional forks of English, with different date and number formats, spelling (colour/color), and vocabulary, present barriers to communication in the guise of reducing friction for local users. Large software companies ship dozens of locale-specific date and number formats, despite ISO having developed standards. CE adheres to ISO [5]. ISO 80000-1 permits either a point or a comma as the decimal marker; CE picks the point, so 3.14 is a number and 3,14 is not. The world benefits from a single English standard without regional variation. Not only do machine protocols benefit from canonical serialization, but humans do as well.

Sense ambiguity is another barrier to communication:

To improve the handling of ambiguity, we advocate for extending redundant, ambiguity-aware annotation protocols (with attention to disagreement) from single-label tasks (e.g. QA, NLI) to complex, sequential outputs like semantic parsing. Improving both zero-shot generalization and data collection would help models capture the full range of utterance interpretations. This could lead to robust, interactive systems in which agents ask for confirmation or clarification on ambiguous examples (Stengel-Eskin & Van Durme, 2023), ultimately improving safety for critical systems.

Useful Applications of Canonical English

• Reliable semantic parsing / knowledge extraction. A closed, well-defined grammar and lexicon means CE can be converted into knowledge-graph triples without further interpretation.
• High-consistency knowledge bases and wikis. Every article, every claim, every date, number, and unit follows the same rules. Machines (and humans) can trust the surface form more. This is valuable for large-scale biographical, historical, or technical wikis where regional drift normally creeps in over time.
• Training-data hygiene and synthetic data generation. Existing corpora can be normalized, or new text generated, so that the result is internally consistent. Downstream models trained on it learn fewer spurious regional correlations.
• Automatic treebanks of paragraphs. noah --parse --to_tree can generate CGEL-style treebanks of paragraphs on a best-effort basis. That is useful to researchers working in the CGEL annotation framework [11], and as a quality-control loop: parse failures and disagreements become candidates for new grammar rules, so later CE versions move closer to the ideal set described on the methodology page.
• A generation constraint that raises prose quality. Because every content lexeme has one sense and every sentence has one structure, grammar that ordinary English cannot check deterministically — parallel structure across bullet items, matching tense and function across coordinated clauses, consistent attachment — becomes decidable. The same grain of access reaches every expression, so the checker can refuse nonsensical or unnatural model prose as surely as a misspelling: a selectional clash such as Grammaticality sits in the check step (grammaticality does not sit), or a stock contrast frame (Not X, but Y; §10). Idioms and metaphors are a finite lexicon, so each one can be counted and switched on or off. Wired into grammar-constrained decoding, that checker masks illegal tokens on each beam: generation barely slows, and the finished string is CE by construction.
• Deterministic suppression of Claudish. Language models, Claude especially, fall into a recognizable register — unusual metaphors and similes, marked or inverted verb–object order, stock contrast frames — that a system prompt does not reliably prevent. Those constructions fail CE’s grammar and figurative-language rules, which are a finite check rather than a taste score. Wired into beam search, the same check refuses them on each hypothesis, so the finished string cannot contain them.
• Controlled interfaces for formal systems. Regulations, contracts, scientific claims, technical specifications, or API documentation written in this dialect become easier to validate, diff, version, and machine-check.
• Pivot language for multilingual systems. A standardized English is a cleaner intermediate representation when translating to or from other languages, or when building cross-lingual knowledge graphs.
• Claim verification and fact-checking pipelines. Surface variation is reduced, so matching claims against a knowledge base or against each other becomes more precise.
• A bounded, versioned list of English rules. CE gives linguists a place to refer to rules — including complex ones like count properties that do not have a stable semantic-web URI — and therefore to discuss them. See A bounded, versioned list of English rules.

A bounded, versioned list of English rules

CE gives linguists a bounded, versioned list of rules of English, including complex ones like count properties [15], which do not have a stable semantic-web URI. CE provides a place to refer to rules and therefore to discuss them. It gives a starting point for linguistics experts to propose new rules, to modify them, and therefore to asymptotically approach as perfect a set of rules for English as possible — rather than starting from scratch, or losing progress and fellow linguists’ attention because of a lack of a shared terminology.

CE also forces linguistics folks to propose their rules formally, because they must execute inside of noah-linter and they must pass a very large regression suite of rules that already exist.

This also means the CE linter gives linguists a powerful tool to experiment and develop novel rules or observations about English. Some of these might get added to a future version of CE. Such rule discovery could potentially also be automated by using agents looping over large corpora of English text; for instance it may discover rules for less-than-formal registers, enabling new types of checking.

Corpus operations on the CGEL tree

Every CE sentence is a CGEL constituency tree with grammatical attributes on each node: clause-level time (present, past, or future), person, voice, valency frame, and syntactic depth. Because the tree records them as fields, a corpus written in CE can be filtered and rewritten by structure. A rule can exclude every clause of one time, or rewrite one feature as another: present time to future time, third person to first person, passive voice to active voice. Further rules can flag unusual constructions, normalize sentence length, split nested subordinate clauses, or refuse the structures a house style names as forbidden.

Those operations walk the tree, as a compiler walks an abstract syntax tree. A rule that retargets one feature applies to every matching node. The rule does not change with corpus size.

• Zero semantic drift. When a language model rewrites a chapter to change time or point of view, the rewrite often changes vocabulary, drops a metaphor, or invents a detail. A CE rewrite edits only the grammatical features the rule names. Every untargeted node keeps its lemma, sense URI, and wording.
• Determinism at corpus scale. A language model’s context window is finite. The model may convert a protagonist from first person to third in chapter 1 and revert to first person by chapter 40. A tree rule applies to the whole corpus at once. On a 1,000-page text, it executes the same way on the last sentence as on the first.
• Structural querying and auditing. Keyword search finds strings. A CE tree uses CGEL trees, which can be queried by structure: every instance of a given valency frame (ditransitive give him the book; locative put it on the table), every passive clause whose agent is omitted (The letter was sent), every subject–verb agreement mismatch, every clause past a depth limit.
• Reversible edits. A tree rewrite leaves the input tree intact, so restoring that tree restores the original wording. A document can be shifted to future time to see how it reads, then restored exactly. A language-model rewrite is a new sample, and the original wording is not recoverable from the result.
• Algorithmic style enforcement. Organizations can write structural guidelines and check them on the tree: parallel structure in lists, a closed set of valency frames for technical documentation, a hard limit on syntactic depth. Each constraint is a finite verdict on the nodes.

The output still reads as natural English. The constraints remove the degrees of freedom that create ambiguity and inconsistency, rather than replacing English with an artificial-looking notation.

Resolve Ambiguity at the Write Step, Not the Read Step

Uncontrolled English requires the reader to have implicit context in order to recover the meaning of the text. CE instead passes that work to the author, who must commit to exactly one sense of each word, and one URI for each proper name, before the document is transmitted to any reader.

A document-level default is not enough. On SemCor with fine-grained WordNet senses, about 33 percent of ambiguous words take more than one sense within a single discourse, with no meaningful difference across parts of speech [9]. Each occurrence must be committed, not the lemma.

The author already knows which sense they meant at the time of writing, so recording it is cheap. Every later reader has to reconstruct that same fact from context, and pays for it in compute, in latency, and in the risk of reconstructing it wrongly. Done by the author, the work happens once. Left to the readers, it happens once per reader, per reading, indefinitely.

Agent systems make that cost asymmetry unjustifiable. A system prompt, a tool description, a policy, a specification, or a contract is now read thousands or millions of times, mostly by machines, at a measurable cost per read — and every one of those reads re-derives meaning the author already knew and did not record. A large part of what retrieval and embedding pipelines do is guess, probabilistically and repeatedly, at what the writer knew and did not write down.

Recording senses by hand used to cost more than letting readers guess, which justified transmitting ambiguous text. That cost argument no longer holds: noah-parser performs the upconversion mechanically and noah-linter confirms it, so the author’s commitment costs a build step. With that tooling, there is no remaining reason to transmit ambiguous text. Disambiguate once, before the document is sent, rather than requiring every recipient to do it again.

CE does not claim to have solved word sense disambiguation. It moves the problem from the read side to the write side and records the result. A sense chosen once, written into the document, and transmitted with the text can be checked, diffed, and disputed. A sense inferred silently on every read cannot.

Read-time inference is measurably unreliable on the cases that matter. On AmbiEnt — 1,645 linguist-annotated English examples spanning lexical, syntactic, and pragmatic ambiguity — GPT-4’s generated disambiguations were judged correct only 32 percent of the time by crowdworkers, against 90 percent for the dataset’s own disambiguations [2]. The same model reached 63.0 percent on true/false judgments about whether an interpretation is valid, but when required to answer all four templates correctly for a single disambiguation, performance collapsed to 2.5 percent. That figure is the claim that two readings need not agree, measured directly: the same model, reading the same sentence, does not hold one interpretation consistently. The output is fluent, the meaning is wrong, and nothing flags the error. That is not an acceptable reading mode for a control surface: a system prompt, a tool description, a policy, or a contract.

Binding Versus Hinting

CE does not disambiguate by writing a glossary of definitions into the prose. A definition offered as ordinary sentences is a hint: a human reader may skip it, and a language model may ignore it in favour of context and attention. When a familiar word is replaced by an unfamiliar one and the reading must be inferred from context and stipulation alone, accuracy falls sharply, to near chance for smaller models [3].

Instead, CE records each commitment as JSON-LD frontmatter: a URI bound to each word, shipped with the document. Ordinary reading can still ignore that block. Audit and checking cannot. Downstream processes — noah-linter, conformance checks, archival tools — can be required to consume the URIs rather than re-infer the senses from the prose.

Benefits of using CE

• Less room for error. Every re-derivation is an opportunity to derive something different. In ordinary prose a misread sense produces a confusing sentence. In a control surface it produces an agent that does something other than what it was told, and because the text read fluently, nothing raised an alarm — an alignment failure from ordinary misreading. Removing the inference step removes that class of failure, rather than making it rarer.
• Deterministic veto of unnatural model prose. Full grammar-level access to every English expression lets the linter refuse predicates a noun cannot bear, constructions the style rules ban, and idioms or metaphors that are off the closed list — not by scoring “naturalness”, but by a finite check. Coupled to beam search, the same verdict barely slows generation, and the output is CE by construction.
• No drift over time. An uncontrolled document grows more ambiguous as it ages, because the reader’s dictionary keeps changing after the writer has stopped. A CE document names each content word by a dereferenceable sense IRI — an Open English WordNet synset page, a Wikidata lexeme for a function word, or a Wiktionary English section for a mark or digit — so a later reader follows the same link the author chose, not a compact prefix whose host no longer exists.
• Native to the semantic web. The metadata is JSON-LD, so a CE document is already RDF: its entities are Wikidata URIs and its sentences are triples that join the wider graph without a mapping layer.
• Durable and citable. Because each assertion carries its own identity and provenance, a CE document decomposes into nanopublications — individually addressable, attributable claims designed for archiving and citation [4].
• Deterministic grammatical rewrite. Time, person, and voice are attributes on the CGEL tree, so a corpus can be retargeted by walking the tree rather than by sampling a language model. Untargeted nodes keep their lemmas and sense URIs.
• A bounded, versioned list of English rules. CE encodes rules, even complex or obscure ones, like count properties. Therefore, CE's definition itself, hosted on GitHub at a stable URL, provides linguistics specialists with a stable semantic-web URI — and therefore to discuss these rules. Linguistics folks can propose their rules formally, and because they must execute inside the linter, they can be checked rigorously, and verified against a very large regression suite of rules that already exist. The linter is also a powerful tool to experiment; some novel rules or observations might get added to a future version of CE, and agents looping over large corpora of English text may discover rules for less-than-formal registers, enabling new types of checking. More.