Multilingual Document Review: How E-Discovery Breaks in Cross-Border Disputes
A dispute between a US plaintiff and a Japanese supplier does not politely restrict itself to English. It arrives as three years of Slack, a shared drive nobody has cleaned since 2019, and email threads that switch language mid-sentence because the sales team writes to head office in one tongue and to customers in another. Somebody now has to read all of it, and the bill for doing that badly is the reason multilingual document review has become its own discipline.
Volume is the enemy, language is the multiplier
Domestic ediscovery is already an exercise in reduction. You collect a million documents, cull by date and custodian, deduplicate, run search terms, and hand a survivable pile to reviewers. Every one of those steps assumes the software can read the text.
Introduce a second language and each step degrades quietly rather than loudly. Deduplication misses near-identical documents because one is a translated copy. Search terms hit in English and sail past the same concept in Korean. Date filters work fine, which lulls everyone into thinking the rest did too. The failure is invisible until a document nobody produced turns up in the other side's exhibit list.
Language identification comes before everything
The first job on a cross-border matter is not review. It is finding out what you are holding. Automated language identification runs over the whole collected population and produces a count by language, by custodian and by document type. Two things usually surface. First, a language nobody mentioned in the kick-off call. Second, a large slice of mixed-language documents, where a single email chain contains Mandarin, English and a spreadsheet attachment with Japanese column headers.
Mixed-language material is the genuinely hard case. Classifiers assign one label per document, so a chain that is 70 percent English gets tagged English and routed to a reviewer who cannot read the 30 percent that matters. Serious teams classify at the segment level, or at minimum flag any document containing more than one script.
Building search terms that survive translation
Search term lists are still how most matters get narrowed, and they are where multilingual review most often breaks. A term list written in English and translated word for word is close to useless. Employees do not write like contracts. They use nicknames for projects, abbreviations invented in-house, romanised spellings of native words, and slang that a dictionary has never heard of.
Term lists that hold up share a few properties:
- They are drafted by a native speaker who has actually looked at sample custodian documents, not by translating the English list
- They include morphological variants, because agglutinative and heavily inflected languages bury the root inside a longer word
- They account for character sets, including traditional and simplified forms, and for text that has been transliterated into Latin script
- They are tested and iterated against the real corpus, with hit counts reviewed before anyone commits to them
Tokenisation deserves its own warning. Chinese, Japanese and Thai do not separate words with spaces. A platform that indexes by whitespace will simply not find your term inside a sentence, and it will report zero hits with total confidence.
Machines sort, humans decide
Technology assisted review and predictive coding both work in multilingual matters, but only if the model is trained on the language it will be judging. A classifier trained on English responsiveness decisions has learned nothing transferable about German documents. The usual answer is one model per language, each seeded by reviewers who read that language, which is more expensive than the sales deck implies.
Machine translation still has a place, and it is the same place it has in any serious litigation workflow: it tells you which documents deserve a human. Nothing produced by an engine goes into a brief, a deposition binder or a production without a qualified linguist having produced the version that will be relied on. That is where proper document translation is bought, and it is a small fraction of the population if the triage upstream was done properly.
The law gets in the way, deliberately
None of this happens in a vacuum. Data collected in Europe is personal data, and moving it to a review platform in the United States engages the General Data Protection Regulation before a single document has been read. France, Switzerland and several other jurisdictions maintain blocking statutes that make it an offence to hand certain material to a foreign court at all, and US litigants are routinely told to go through the Hague Evidence Convention instead, a route everyone agrees is slow.
The practical consequences reshape the technical plan. Review often has to happen in-region, on a platform hosted in the relevant country, with reviewers who are physically there. Personal data belonging to uninvolved employees gets redacted before anything crosses a border. Frameworks published by bodies such as EDRM give teams a shared vocabulary for these stages, and the working-level arguments about what actually functions play out in places like r/ediscovery.
Costing the review before you buy it
Foreign language review costs more per document than English review, and the multiplier is not small. Reviewers who are qualified lawyers and fluent in Japanese, Korean or Arabic are scarce, they bill accordingly, and there are not enough of them to staff a rush. A population that would take 30 contract attorneys three weeks in English can take three qualified bilingual reviewers three months, which is a schedule problem long before it is a budget problem.
That arithmetic is why the culling work matters so much. Every document removed by defensible language-aware filtering is a document nobody pays a premium rate to read. It is also why the language breakdown belongs in the first budget submitted to the client, not in a revised one issued after the platform contract is signed and the reviewers have been hired.
The pattern across failed cross-border reviews is the same. The team scoped the matter as if it were domestic, discovered the language problem after the platform was chosen and the budget was approved, and spent the rest of the case paying for that order of operations.