Skip to main content

Dictionary-Augmented LLM

The idea: Force known, verified translations for specific terms from a bilingual dictionary, and let the LLM handle sentence structure and unknown vocabulary. The dictionary provides anchors of correctness; the LLM provides fluency.

:::info This is a cookbook, not a finished implementation This guide sketches the approach. The specific dictionary matching and injection strategy will depend on your language pair and available lexical resources. :::

When to Use This​

  • A bilingual dictionary exists for your language pair (even a small one)
  • The LLM consistently hallucinates key terms — inventing words that don't exist
  • You need terminological consistency across translations (same word translated the same way everywhere)
  • You're translating domain-specific content where standard LLM translations are wrong (legal, medical, educational)

How It Works​

  1. Load a bilingual dictionary — key→value pairs mapping source terms to verified target translations
  2. Match source text against dictionary — identify terms in the input that have known translations
  3. Inject matches into the prompt — tell the LLM "these terms MUST be translated as follows"
  4. LLM generates translation — with dictionary constraints as hard requirements
  5. Post-process — verify dictionary terms appear in the output; retry if they don't

Dictionary Format​

dictionaries/crk-terms.json
{
  "school": "kiskinwahamâtowikamik",
  "teacher": "okiskinwahamâkêw",
  "student": "kiskinwahamâkan",
  "book": "masinahikan",
  "home": "kīwēwin",
  "water": "nipiy"
}

Prompt Structure​

Translate the following English to Plains Cree (SRO).

REQUIRED TERMINOLOGY — use these exact translations:
- "school" → "kiskinwahamâtowikamik"
- "teacher" → "okiskinwahamâkêw"

Source: "The teacher went to the school"

Key Design Decisions​

Matching strategy: Exact match is simplest. Lemmatized matching ("teachers" matches "teacher") catches more but requires a source-language lemmatizer. Fuzzy matching risks false positives.

Inflection handling: In polysynthetic languages, the dictionary form may need inflection to fit the sentence. You can provide the root and let the LLM inflect, or provide multiple inflected forms. An FST can validate the result.

Conflict resolution: What if the LLM ignores a dictionary term? Options: (a) retry with stronger instruction, (b) post-process by string replacement, (c) accept and flag for review.

Pros and Cons​

✅ Eliminates hallucination for known terms❌ Dictionary coverage is always incomplete
✅ Guarantees consistency for key vocabulary❌ Inflection/conjugation may not match sentence context
✅ Easy to audit and update❌ Over-constraining can produce unnatural output
✅ Dictionary is a reusable asset❌ Requires a dictionary to exist in the first place

Where to Find Dictionaries​

  • itwêwina — Plains Cree–English (FST-powered, open source)
  • Wolvengrey Dictionary — comprehensive Plains Cree reference
  • Apertium — bilingual dictionaries for dozens of language pairs
  • Giellatekno — dictionaries for Sámi, Uralic, and other minority languages
  • Community-created glossaries, educational materials, term lists

Combines Well With​

See Also​