ste100-checker

Rule-based ASD-STE100 (Simplified Technical English) compliance checker for procedural and descriptive technical text. It analyzes files with spaCy and reports writing-rule violations together with approved-alternative suggestions.

STE is a controlled natural language: a restricted vocabulary, one meaning per word, short simple sentences. No public labeled STE corpus exists, so all compliance logic is rule-based on top of spaCy's statistical tokenizer, POS tagger, dependency parser, and lemmatizer.

Features

• Ten automatable STE rules: vocabulary, part of speech, spelling, noun clusters, verb tense, passive voice, sentence length, semicolons, phrasal verbs
• Approved-alternative suggestion with every violation
• Text output for humans, JSON output for pipelines
• Curated, extensible JSON dictionary (approved, unapproved, technical)

Installation

Requires Python 3.14.

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
python -m spacy download en_core_web_sm

Usage

ste100 <file.txt> [--format text|json]

Text output:

$ ste100 samples/non_compliant.txt
Words: 45
Sentences: 3
Violations: 20

1. [1.1] "Utilize" is not an approved word (57-64)
   suggestion: use
2. [1.1] "colour" is not in the approved or technical dictionary (69-75)
...

JSON output:

{
  "words": 45,
  "sentences": 3,
  "total": 20,
  "violations": [
    {
      "rule": "1.1",
      "start": 57,
      "end": 64,
      "text": "Utilize",
      "message": "\"Utilize\" is not an approved word",
      "suggestion": "use"
    }
  ]
}

Each violation carries the rule id, character offsets into the source text, the offending span, a message, and an approved alternative when one exists. Exit status is 0 whether or not violations are found; parse the JSON and branch on total if you need a nonzero signal.

Rule coverage

Non-automatable rules (approved meanings, topic sentences, comprehension) are documented as out of scope; checkers are aids, not proofs.

Extending the dictionary

data/ holds the vocabulary as JSON. technical.json stores the domain words you add under rules 1.5 / 1.12:

{
  "nouns": ["actuator", "alternator", "antenna", "duct", "nozzle", "piston"],
  "verbs": ["bond", "drill", "weld"]
}

Append your own terms and rules 1.1 and 1.2 accept them automatically. approved.json maps approved words to their one allowed part of speech; unapproved.json maps unapproved words to their approved alternatives.

Evaluation

samples/eval.json is a 16-sentence annotated suite. Run the harness:

python -m src.evaluate

Current results: precision 0.941, recall 1.000, F1 0.970 (one false positive: rule 1.1 flags generator, a legitimate technical noun absent from the small curated dictionary). For context, Boeing BSEC reports 79%/89% and SECC 87%/93% on their much larger suites.

Development

ruff check .
ruff format --check .
python -m pytest

Project layout

├─ data/          curated approved / unapproved / technical dictionaries
├─ src/           package root: dictionary, pipeline, engine, rules, report, cli, evaluate
├─ src/rules/     one module per rule family
├─ samples/       compliant / non-compliant examples + annotated eval suite
├─ tests/         per-rule unit tests and CLI/eval tests
├─ case-study/    college case-study report (Typst)
└─ .github/       CI and release workflows

License

MIT. See LICENSE.