TVS-2WikiMultiHopQA Dataset

This repository contains the tvs-2wikimultihopqa dataset, which is a key component of the "Think, Verbalize, then Speak" (TVS) framework presented in the paper Think, Verbalize, then Speak: Bridging Complex Thoughts and Comprehensible Speech.

• Project Page: https://yhytoto12.github.io/TVS-ReVerT
• Paper: https://huggingface.co/papers/2509.16028
• Code: https://github.com/yhytoto12/TVS-ReVerT

Introduction

The "Think, Verbalize, then Speak" framework addresses the challenges of directly applying Large Language Models (LLMs) in spoken communication, where optimal textual and verbal delivery often mismatches. This dataset is specifically designed for the "verbalizing" step, an intermediate process that translates complex thoughts into natural, speech-ready text, thereby preserving the full reasoning capacity of LLMs.

The tvs-2wikimultihopqa dataset contains thought-verbalization pairs, which are crucial for training verbalizers (such as ReVerT) to transform intricate reasoning outputs into concise and natural spoken language. This enhances speech naturalness and conciseness with minimal impact on reasoning performance.

Citation

If you find our project useful for your research and applications, please kindly cite using this BibTeX:

@inproceedings{tvs2025@woolee,
  title={Think, Verbalize, then Speak: Bridging Complex Thoughts and Comprehensible Speech},
  author={Sang Hoon Woo, Sehun Lee, Kang-wook Kim, Gunhee Kim},
  booktitle={Proceedings of the EMNLP 2025},
  year={2025}
}