Reproduce “Multilingual Dialogue Generation and Localization with Dialogue Act Scripting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Justin Vasselli, Eunike Andriani Kardinata, Yusuke Sakai, Taro Watanabe
The Dialogue Act Script (DAS) framework, developed at Nara Institute of Science and Technology, introduces a three-step pipeline for generating culturally appropriate multilingual dialogues by abstracting communicative intent, localizing it, and then decoding into target languages. It consistently produced dialogues preferred over machine and human translations in fluency, coherence, and cultural relevance across multiple languages.
Key insights
- Abstracting dialogue into intent-based representations before localization effectively mitigates anglocentric biases and translation artifacts, leading to more culturally relevant content.
- A modular, three-step pipeline (encoding, localization, decoding) outperforms a single-step, end-to-end prompting approach for multilingual dialogue generation.
- Leveraging LLMs for generating and processing structured dialogue act scripts offers a scalable method for creating high-quality, culturally appropriate multilingual data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.