Improved Dialogue Localization and Translation with Dialogue Act Scripting

ACL ARR 2025 February Submission6570 Authors

Abstract: Non-English dialogue datasets are scarce, and models are often trained or evaluated on translations of English-language dialogues, which can introduce artifacts that reduce their naturalness and cultural relevance. This work proposes Dialogue Act Script (DAS) as a structured framework for encoding, localizing, and generating multilingual dialogues. Rather than directly translating, DAS generates new dialogues in the target language by adapting a language-independent representation, ensuring greater cultural relevance and naturalness. By using structured dialogue act representations, DAS improves multilingual dialogue localization by enhancing cultural adaptability, reducing translationese, and providing an interpretable framework for structured adaptation. The results show that DAS-generated dialogues consistently outperform machine and human translations across Italian, German, and Chinese in all evaluation criteria, particularly in cultural relevance, coherence, and situational appropriateness, suggesting that functional abstraction allows explicit adaptation to conversational norms that straightforward machine translation may not capture.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: evaluation and metrics, multilingual / low resource, conversational modeling

Contribution Types: NLP engineering experiment

Languages Studied: Italian, German, Chinese, English

Submission Number: 6570