A Unicode-Aware Term Inventory for Manuscript Review, Without Automatic Replacements

#javascript

Disclosure: This tutorial was prepared and locally checked by an AI agent for the PaperHumanizer product team. Examples are synthetic. It is not a report of customer behavior or measured rewriting quality.

A writing application can change an instrument name into a plausible synonym. A literal counter cannot decide whether that change is scientifically valid, but it can help a reviewer find the forms they asked to inspect.

This small Node.js utility takes an explicit JSON dictionary and UTF-8 text from stdin. It prints the dictionary labels and per-variant counts, without manuscript excerpts or input paths. It has no dependencies beyond Node's built-in modules and makes no network request. Keep the dictionary and counts private as well: they may disclose unpublished terminology.

Start with author decisions

Our fictional manuscript measures “revision quality” with an invented “Revision Quality Index (RQI).” If a candidate says “writing ability scale,” that is a review item, not an instruction to replace it everywhere. That phrase might be correct in a quotation or another study.

The free terminology style sheet separates preferred names, permitted variants, review-only forms, and sources to consult. The literal inventory supports that manual process; it does not supply the decisions.

For example:

[
  { "label": "instrument", "variants": ["Revision Quality Index", "RQI"], "caseSensitive": true },
  { "label": "constructs", "variants": ["revision quality", "writing ability", "confidence"] }
]

Escape input instead of treating it as a regex

Here is the core of the runnable module:

function escapeLiteral(value) {
  return value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
}

export function inventory(text, entries) {
  if (typeof text !== 'string' || !Array.isArray(entries) || entries.length === 0) throw new Error('Invalid input');
  return entries.map(entry => {
    if (!entry || typeof entry.label !== 'string' || !entry.label.trim() || !Array.isArray(entry.variants) || entry.variants.length === 0 || (entry.caseSensitive !== undefined && typeof entry.caseSensitive !== 'boolean')) throw new Error('Invalid entry');
    const counts = entry.variants.map(variant => {
      if (typeof variant !== 'string' || !variant.trim()) throw new Error('Invalid literal');
      const expression = new RegExp(`(?<![\\p{L}\\p{N}_])${escapeLiteral(variant)}(?![\\p{L}\\p{N}_])`, entry.caseSensitive ? 'gu' : 'giu');
      return { variant, count: [...text.matchAll(expression)].length };
    });
    return { label: entry.label, caseSensitive: !!entry.caseSensitive, counts };
  });
}

The user supplies literal strings. Escaping metacharacters keeps C++ from turning into regex syntax. Validation rejects blank labels, empty dictionaries, empty variants, and a non-boolean case-sensitivity option instead of producing reassuring but meaningless output.

Unicode property boundaries stop RQI from matching inside aRQI, RQI_2, or 前RQI. The definition is deliberately limited: letters, numbers, and underscore are word characters here; a hyphen is not. Thus RQI-based contains a match. This is not a tokenizer for every language or scientific notation.

Run it locally

The repository includes term-inventory.mjs, its test file, and a synthetic starter dictionary. Inspect the code, save the files together, and use Node.js:

node term-inventory.mjs term-sheet.example.json < private-draft.txt > private-counts.json
node --test term-inventory.test.mjs

The CLI catches file, JSON, or dictionary errors and returns a nonzero status with a generic message, without printing the dictionary path or input text. It writes no manuscript file. A source code inspection is still necessary before using any script with confidential data.

Test boundaries that change the interpretation

Five locally executed tests cover:

Literal punctuation and regex metacharacters, including C++ and (RQI).
Case-sensitive and insensitive matches, word embeddings, Unicode neighbors, and hyphen boundaries.
Independent overlapping phrase counts, exact whitespace, and a decomposed accent that is not normalized to a precomposed one.
Invalid dictionary entries that must fail.
The CLI's stdout/stderr contract, including malformed JSON, without manuscript or private-path leakage.

Overlapping counts are not unique mentions. If revision quality and quality are both in the dictionary, the same words can contribute to both counts. Likewise a two-space variant will not match a one-space phrase. These are choices to document, not evidence that the author made a mistake.

Keep the semantic review human

The script does not inspect Word archives, formulas, figure text, abbreviation definitions, references, or the first occurrence of a term. A zero count does not prove an omission from the full manuscript. It does not infer synonyms or judge which construct was measured. Use it to locate review work, then consult the sources and your co-authors.

We build PaperHumanizer for academic prose revision. Generation requires sign-in and credits; the worksheet and local counter are free without a product account. Treat generated prose as a candidate, verify facts and sources, and follow your institution's and journal's AI-use policies. No detector result or publication acceptance is guaranteed.

DEV Community