Evaluation of Dialogue Translation in a Self-made Game Using a Pre-trained BERT Model

Abstract

This paper investigates the effectiveness of using a pre-trained multilingual BERT model to translate in-game dialogue from Japanese to English in a custom-developed co-operative action game. In-game dialogues often contain informal expressions, character-specific tones, and emotional nuances that are critical for preserving narrative immersion. As a baseline, we use ChatGPT, which generally produces fluent and contextually appropriate translations. We qualitatively evaluate BERT’s translations against this baseline, focusing on style fidelity, speaker consistency, and conversational naturalness. While BERT provides structurally accurate and literal translations, it tends to lack contextual sensitivity and character-aware adaptation. These findings highlight the limitations of applying unadapted BERT models in narrative contexts. Based on the results, we propose future enhancements through character-specific fine-tuning and context-aware modeling. This study serves as a first step toward building an adaptive, character-sensitive translation system for dialogue-intensive game environments.

Ähnliche Arbeiten