AI Translation Methods Compared for Rock Art Documents
A recent study on arXiv evaluates three different English machine translation systems applied to a Spanish academic text about rock art, emphasizing the precision of terminology. The systems analyzed included DeepL (a robust NMT baseline), Gemini-Simple (an LLM utilizing basic prompts), and Gemini-RAG (an LLM that enhances prompting with glossary-based term-pair retrieval). Through human assessments using multi-way Direct Assessment and focused terminology checks with a limited MQM taxonomy, Gemini-RAG achieved the highest exact-match terminology accuracy, reaching 81.4%. This research underscores the significance of precise specialized terminology in sharing cultural heritage to prevent misinterpretation by non-experts.
Key facts
- Study compares three MT setups for Spanish rock art text translation to English
- Setups: DeepL (NMT baseline), Gemini-Simple (LLM basic prompt), Gemini-RAG (LLM with glossary)
- Evaluation methods: multi-way Direct Assessment and terminology auditing with MQM taxonomy
- Gemini-RAG achieved 81.4% exact-match terminology accuracy
- Focus on simple, operationally feasible interventions rather than complex model-side modifications
- Cultural heritage institutions face constraints in multilingual dissemination due to limited budgets and staffing
- Small lexical errors in terminology-dense domains can mislead non-specialists and reduce reuse
- Published on arXiv with ID 2605.14679
Entities
Institutions
- arXiv