A Study on Prompt Engineering Strategies and Context-Enhanced Translation for Video Game Localization

Abstract

Video game localization often depends on contextual understanding rather than sentence-level translation alone. This problem is especially evident in fragmented narrative games such as Elden Ring, where information about characters, places, factions, terminology, and world-building is scattered across item descriptions, equipment texts, skill explanations, and other isolated entries. When these entries are translated without sufficient background information, large language models may produce inconsistent terms, weakened narrative references, inaccurate interpretation of polysemous expressions, or a style that does not match the original atmosphere. To address this problem, this study proposes a context-enhanced prompt-based translation framework for game localization. The framework builds a local semantic knowledge base, retrieves relevant background information through contextual embeddings and cosine similarity, and inserts the retrieved context into structured translation prompts. Using Elden Ring item descriptions as the test corpus, this study compares Doubao, Gemini, and GPT-5.5 under four settings: baseline translation, prompt engineering, context enhancement, and the full method. The evaluation uses BERTScore F1, character-level BLEU-4, and the Global Terminology Enrichment Score (GTES). The results show that context retrieval is more effective than prompt engineering alone in improving domain terminology recall, and the full method achieves the strongest overall performance. This suggests that external knowledge retrieval can help reduce contextual loss in fragmented game text translation.