Skip to content

Glossary drafting: Wikipedia cross-check as an automatable first pass (revert to English if unsure, flag for native review) #300

Description

@mmcky

Proposed by @jstac on PR #69 (2026-08-17 and 08-19) after the Japanese glossary review: a first-pass, automatable rule for machine-drafted glossary terms — cross-check each candidate term against the target-language Wikipedia; if still unsure, revert to English; flag for native-speaker review.

Why

The ja review showed the shape of the problem. @xuanguang-li ran the check by hand with ChatGPT on the 357-term draft: 268 terms were found on Japanese Wikipedia, of which only 3 differed from the draft (産業連関表 vs 産業連関モデル — a false positive, the table vs the model; 限界収益 vs 限界収入 and 割引現在価値 vs 現在割引価値 — both adopted). 89 terms were not found at all, and that set overlapped heavily with the terms the native reviewers ended up sending back to English (Forcing variable, Forward-looking difference equation, Authority centrality, Cost-to-go function, Lake model). So "not on Wikipedia" is a useful signal for "no established term — keep English", and "on Wikipedia under a different name" is a short, reviewable diff.

The policy the check would serve is recorded in .dev/decisions/D-2026-09-01-ja-terminology-policy.md: Japanese where a well-known, standard counterpart exists, English otherwise, if in doubt English. The same shape should apply to any future glossary that sits between full translation (zh-cn, fa, fr) and keep-English-dominant (ml).

Sketch

A script (alongside the glossary-review skill, which already finds the terms two models disagree on) that takes glossary/<lang>.json, queries the language's Wikipedia search API per en term, and emits three buckets: match (article title equals the draft rendering), differs (article exists, title differs — show both for review), not found (candidate for reverting to English under the if-in-doubt rule). Output as a markdown table a native reviewer can work through, the way @xuanguang-li's table was. Name-context entries are skipped (names stay Latin script by policy).

Open questions: proper-noun entries with a Wikipedia article under a katakana title (e.g. economists) would pass the check but the policy still wants English — so the policy filter runs before the lookup, not after; and disambiguation pages / redirects need handling so a redirect from the draft term to the canonical title counts as a match, not a difference.

Refs

PR #69 thread (2026-08-17 → 08-19), .dev/decisions/D-2026-09-01-ja-terminology-policy.md, glossary/README.md "Which terms belong in a glossary".

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions