# Reviewing German-Language Corpora: Why Keyword Search Fails Where Prompts Do Not

_2026-11-17_

A term list assumes that the documents you need contain the words you thought of, in the form you typed them. German tests that assumption harder than English does. Words fuse into compounds, nouns change shape with case and number, umlauts reach a collection in more than one spelling, and a business email can move into English halfway through a thread.

For litigation support teams handling German-language collections, which come from internal investigations, arbitrations, access requests and other matters we described in our story on [where the document work in Germany comes from](/stories/ediscovery-in-germany-without-discovery-where-the-document-work-comes-from), those gaps decide which documents a reviewer ever sees. This story takes the four problems in turn, then looks at how a review in Claira inside Nuix Discover can describe the issue in a prompt, and where term lists still earn their place.

## Compounds hide the word you searched for

German writes compound nouns as one word. Duden lists [Kaufvertrag](https://www.duden.de/rechtschreibung/Kaufvertrag), a contract concluded between a seller and a buyer, as a single noun formed from Kauf and Vertrag. In an index that stores whole words, Kaufvertrag is its own entry, separate from Vertrag, and a search for the base word reaches it only if the query is built to look inside longer words.

The usual remedy is a wildcard. Claira's help centre, describing Nuix Discover search syntax, explains that wildcards collapse grammatical variants, so that waterproof* reaches waterproofing and waterproofed ([Generating Search Term Families](https://support.claira.to/docs/en/workflows/generating-search-term-families)). That works when the stem comes first. In Kaufvertrag the word that matters comes last, after the kind of contract it is, and a contracts dispute can involve many such compounds. Each wildcard placed to catch them also widens the net, which is why the same article recommends sample-testing imported terms against the collection.

## Declension and the umlaut plural

German nouns change form with case and number. Duden gives the genitive of [Vertrag](https://www.duden.de/rechtschreibung/Vertrag) as Vertrags or Vertrages and the plural as Verträge. The plural adds an ending and also changes the vowel, which Duden defines as the [umlaut](https://www.duden.de/rechtschreibung/Umlaut), the shift of a, o, u and au to ä, ö, ü and äu.

For a term list, the consequence is concrete. Verträge does not begin with the letters Vertrag, so a stem with a trailing wildcard reaches it only if the index treats ä and a as the same letter. The person drafting the terms has to know how the index was built, or add the plural as a term of its own, and then do the same for every compound, since Kaufvertrag has the plural Kaufverträge.

## One word, several spellings

The same word can arrive in different spellings. Where ä, ö and ü are not available, German writes ae, oe and ue, and as the hosting provider IONOS notes in its help pages, [domain names that use those replacements are still common](https://www.ionos.com/help/email/glossary-email-terms-explained/umlaut-addresses-at-a-glance/). An email address or a file name in a collection may therefore spell a word with ae where the message body uses ä.

The letter ß adds another variant. Guidance published by the [Council for German Orthography](https://www.rechtschreibrat.com/?p=6130) on the reformed spelling sets out that ss stands for ß at the end of a word stem after a short vowel, as in Fluss, muss and dass, where the earlier standard wrote daß. A collection that includes older documents can hold both forms of the same word. Each variant is another row in the term list, multiplied by every compound and plural above.

## German and English in the same thread

German business writing borrows from English. Duden lists [Deadline](https://www.duden.de/rechtschreibung/Deadline) as a German noun, die Deadline, plural Deadlines, taken from the English. As an illustration, a thread about a late delivery might open in German, move to English when a colleague abroad joins, and close in German again. A term list drafted in one language catches only part of that conversation, and one drafted in both doubles the variants to test.

Craig Ball, writing for EDRM in February 2024, called lexical search "a crude tool that misses much more than it finds" and argued that "every lexical search is informed and improved by the iterative testing of queries" ([Lessons from Lousy Lexical Search](https://edrm.net/?p=194400)). In German each concept carries more forms, so each round of that testing has more to check.

## Describing the issue in a prompt

A prompt approaches the same collection from the other end. The review team writes down what it is looking for, and the model reads each document's text against that description, so the team describes the issue once without listing every form a word can take. Claira's [prompting guidance](https://support.claira.to/docs/en/ai-review/prompting/overview) asks for clear criteria, definitions of case-specific terms, a fixed output format and an instruction for documents that lack the information. For a German collection, a relevance prompt can describe the supply arrangement at issue, name the parties and products, note that documents may be in German, English or both, and ask for a short description of each document in English alongside the call.

Case Context keeps that background in one place. It is a structured briefing on the matter, shared by the people on the case, and the prompt generator draws on it to write prompts specific to the matter ([Case Context](https://support.claira.to/docs/en/case-context/overview)). The team then tests the prompt in [Single Review](https://support.claira.to/docs/en/ai-review/single-review) on 10 to 25 representative documents, as the help centre recommends, choosing ones with compounds, ae spellings and English passages among them. Once it behaves as counsel expects, the team runs it across the set as a bulk scan, with each result written to a field in Nuix Discover.

Where the languages in a collection are not known in advance, the help centre's [translation workflow](https://support.claira.to/docs/en/workflows/translation) identifies each document's primary language and summarises it in English in up to two sentences, which helps a team decide what goes to a professional translator. The help centre is clear that these summaries are not certified translations and that quality varies by language.

## Where term lists still help

Term lists still have uses, for an agreed cut of the data or a first look at a collection, and Claira can draft them. Ask the Agent for search terms, or click Generate search terms on a Case Context, and Claira writes a search term family for Nuix Discover: one term per concept, each with a short label and a query in Nuix Discover search syntax, checked before it is shown. Download spreadsheet saves an Excel file that you import with Add search terms, then Add terms from a file. The help centre notes that the queries follow the language of the documents, and its advice to sample-test wildcards applies here too.

Running both on a German set gives a useful comparison. The terms show what an agreed list would have returned, and sampling the documents the prompt flagged that the terms did not reach shows whether compounds, plurals, alternative spellings or English passages account for the difference.

## Where to start

Start with a German-language set your team has already coded. Run the existing term list, write a prompt that describes the same issue, test it in Single Review and compare both against the coding on a sample. Where the two disagree, read the documents themselves: the pattern tells you how much your term lists depend on anticipating every form of a word. This is general information about review methods, and the right approach for a given matter is for the lawyers on it to decide. If you would like to try the comparison on your own documents, you can [book a session with our team](/meet).