Samo-preverjanje prevodov: naj stroj ujame napake

Pri pisanju petjezičnih prevodov sem vedno delal enako napako: nemški glagol v indoneščini, angleško besedilo v kitajskem slotu, en slot manjka. Za oko skoraj nevidno, na produkcijski strani pa takoj.

First, understand the slot layout

Five languages are not always a 5-tuple. Real projects mix at least four layouts; a script that assumes "everything from index 1 is foreign" will flag perfectly legal content.

# 5  = (zh, en, id, sl, it)
# 10 = 问答按语交替 (zh_q, zh_a, en_q, en_a, ...)
# 6  = (name, zh, en, id, sl, it)   前两位是名称与中文
# 4  = (zh_q, zh_a, en_q, en_a)
# 3  = (id, sl, it)               根本没有中文槽位

The 3-tuple line is the trap: asserting "index 0 must be Chinese" flags perfectly fine Indonesian-only data as broken.

Three classes of error to catch

• 外语槽位含 CJK：印尼语里混进「訳」这样的日文汉字，或中文漏进英文槽Foreign slot containing CJK: a Japanese kanji like 訳 slipping into Indonesian, or Chinese leaking into an English slot
• 中英混排：一句里两种文字拼在一起，比如「如此Jdahit」「特征CORE功能」Mixed scripts mid-sentence: two writing systems fused, like 「如此Jdahit」 or 「特征CORE功能」
• Misaligned or empty slots: a tuple short by one element, or a language left blank

The third is the sneakiest: an empty slot silently falls back to a neighbouring language, so the page looks filled while the language is wrong.

Infer from structure, not position

def latin_indices(n):
    """给定槽位总数，返回「应该是拉丁语」的下标集合。"""
    return {1: {1, 2, 3, 4},
            2: {1},
            3: {1, 2},
            4: {2, 3},
            5: {1, 2, 3, 4},
            6: {2, 3, 4, 5},
            9: {2, 3, 4, 5, 6, 7, 8},
           10: {2, 3, 4, 5, 6, 7, 8, 9}}.get(n, set())

With that table the check becomes project-agnostic, and reusable for the next set of pages without rewriting it.

Don't hand-edit data files

When appending large translation blocks via a shell script I produced a duplicated robots.txt group; editing the JSON by hand created nested lists and duplicated columns.

Conclusion: data files are only ever written by scripts, and one-off scripts must validate before writing — abort when validation fails.

if bad:
    print("发现 %d 处问题，未写入" % bad)
    return 1          # 中止，不碰文件
write_file(path, data)

Sorodna orodja