Multilingual KokoroChat: A Multi-LLM Ensemble Translation Method for Creating a Multilingual Counseling Dialogue Dataset

This dataset was created by translating KokoroChat, a large-scale manually authored Japanese counseling corpus, into both English and Chinese. We developed and employed a novel Multi-LLM Ensemble method. Our approach first generated diverse hypotheses from multiple distinct LLMs. A single LLM then produced a high-quality translation based on an analysis of the respective strengths and weaknesses of all presented hypotheses.

This work has been accepted to the main conference of LREC 2026.

🤖 LLMs Used for Translation

For the English translation, we selected three models representing the state-of-the-art at the time of manuscript preparation: GPT-5 (gpt-5-2025-0907), Gemini-2.5-Pro (gemini-2.5-pro), Grok-4 (grok-4-0709). For Chinese translation, we replaced Grok-4 with Qwen-Plus (qwen-plus-2025-07-28).

We selected Gemini-2.5-Pro as the refiner LLM.

📊 Dataset Statistics

Language	Dialogues	Avg. utterances/dialogue
Japanese (KokoroChat)	6,589	91.2
English	6,565	91.2
Chinese	6,582	91.2

Note that the slight reduction in the number of dialogues for the multilingual versions is due to the exclusion of content that triggered the LLM's safety filters.

📁 Directory Structure

Chinese/                 # Chinese translation dialogues by Multi-LLM Ensemble
└── zh*.json

English/                 # English translation dialogues by Multi-LLM Ensemble
└── en*.json

src/                     # Source code
├── TestHyp/             # Hypotheses generated by each LLM (used for refinement)
|      ├── Chinese/
|      |      └── {LLM}4.json
|      └── English/
|             └── {LLM}4.json
|
├── CancelBatchProcess.py  # Script to cancel running batch jobs
├── config.json            # API keys for each LLM
├── Hypothesis_{LLM}.py    # Prompts and code to generate hypotheses with each LLM
├── main.py                # Main entry point to run the experiments
├── Refine_Gemini.py       # Prompt and code to refine translations by integrating 3 hypotheses
└── utils.py               # Shared utility functions

The English translation dialogue data is stored in English/ as en*.json, while the Chinese translation dialogue data is stored in Chinese/ as zh*.json.

src/Hypothesis_{LLM}.py contains the prompts and code used to generate hypotheses with each LLM. src/Refine_Gemini.py contains the prompt and code used to refine translations by integrating 3 hypotheses.

src/TestHyp/ contains, for each language, one dialogue of hypotheses generated by each LLM, which are used to reproduce the integration performed in src/Refine_Gemini.py.

If you want to run each file, please execute src/main.py and specify the file ID to be translated, the LLM to use, and the target language.

💬 DataFormat

The table below describes the fields of each utterance object in the dialogue array.

Key	Type	Description
role	String	Speaker's role (counselor or client).
time	String	The timestamp of the utterance in ISO 8601 format.
origin	String	The original utterance in Japanese.
content	String	The translated text of origin in the target language.

📄 Citation

If you use this dataset, please cite the following paper:

@inproceedings{suzuki2026multilingualkokorochat,
  title     = {Multilingual KokoroChat: A Multi-LLM Ensemble Translation Method for Creating a Multilingual Counseling Dialogue Dataset},
  author    = {Ryoma Suzuki and Zhiyang Qi and Michimasa Inaba},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year      = {2026},
  url       = {https://github.com/UEC-InabaLab/MultilingualKokoroChat}
}

⚖️ License

Multilingual KokoroChat is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) license.

6