Skip to content

Extend to trilingual code-switching (JA-EN-X) #9

Description

@Awshesh12

Everything so far is bilingual JA–EN. Real multilingual workplaces in Japan routinely involve a third language, and the script-accuracy framing generalises naturally — it is defined per-language, not per-pair.

Why it is interesting

  • Script accuracy scales to N languages without redefinition: the metric asks whether each language's words land in that language's script. A trilingual extension is a real test of whether the metric is principled or was tuned to a two-language case.
  • Script-distinct triples are the strongest test: JA–EN–KO (kanji/kana, Latin, Hangul) or JA–EN–ZH (the hard case — kanji and hanzi share glyphs, so script alone cannot disambiguate).
  • CS-FLEURS covers 113 code-switched pairs across 52 languages, so evaluation data for several triples partly exists.

Open questions worth a paper

  • Does encoder LoRA still carry the gains at three languages, or does capacity become the binding constraint?
  • Does adding a third language degrade the first two, and does the balanced-mix trick that fixed Japanese forgetting still work?
  • Does JA–ZH break script accuracy, given shared characters? If so, what replaces it?

Suggested scope

Start with JA–EN–KO: three visually distinct scripts, good FLEURS coverage, and a clean test of whether the method generalises before tackling the JA–ZH ambiguity.

Metadata

Metadata

Assignees

No one assigned

    Labels

    dataCorpora, augmentation, dataset constructionresearchExploratory direction, new scope

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions