Multilingual Support

Status: Current behavior reference
Last verified: 2026-03-19

Current multilingual boundary

Batchalign supports multilingual corpora primarily through:

• file-level language metadata
• per-utterance language directives
• conservative handling of per-word code-switch markers

This is the current public contract. It is more explicit than older BA2-era single-language assumptions, but it is not the same as full per-word bilingual analysis.

Current morphosyntax behavior

Current morphosyntax handling is language-aware at the utterance level.

Practical consequences:

• utterances can be processed with language-aware routing rather than assuming the entire file is one language
• code-switched words marked at word level are handled conservatively rather than being forced through a possibly wrong language model
• cache and payload handling keep language as part of the processing boundary

Current output consequences

Users working with multilingual corpora should expect:

• better behavior than a file-wide single-language assumption
• clearer distinction between utterance-level language handling and word-level foreign-language marking
• some foreign/code-switched words to remain conservatively represented instead of receiving overconfident morphology

Current limits

Batchalign does not currently promise:

• full per-word routing into multiple language-specific NLP pipelines
• perfect code-switch analysis inside one utterance
• complete elimination of manual review for difficult bilingual material

Related references

• Language-Specific Processing, how each pipeline stage diverges per language
• Language Code Resolution, ISO mapping, model resolution
• Language Routing, per-utterance routing into Stanza, per-word routing limits, auto-detection
• Language Data Model
• L2 & Language Switching

This page last changed: 2026-06-19 (commit c82a6d03). The whole book last changed: 2026-09-16 (commit 34d249d8).