CHORUS Uyghur Dataset

# 📊 CHORUS: Uyghur Text Classification Dataset
**Topic and Sentiment Analysis with Multi-LLM Consensus Voting**

[](https://doi.org/10.5281/zenodo.18804285)
[](https://creativecommons.org/licenses/by-nc/4.0/)

## 📖 Overview
The **CHORUS Uyghur Dataset** is a high-quality, comprehensively annotated text classification dataset specifically designed for low-resource Natural Language Processing (NLP). The text corpus is widely collected from mainstream short-video social media platforms (Douyin, Kuaishou) and authoritative news portals (Tianshannet / Xinjiang Daily).

The core feature of this dataset is the pioneering **Multi-LLM Consensus Voting** mechanism. Each Uyghur text entry in the dataset has been independently evaluated for both Topic and Sentiment by four state-of-the-art Large Language Models:
* 🤖 **Claude 4.5 Sonnet**
* 🤖 **Gemini 3 Flash**
* 🤖 **ChatGPT 5**
* 🤖 **Deepseek V3.2**

The final ground truth labels are rigorously determined based on the voting status among the models (e.g., Majority Win, Tie) or subsequent human arbitration (Human Review). This dataset is not only ready-to-use for training and fine-tuning downstream Uyghur NLP models but also serves as an excellent benchmark for evaluating the alignment, multilingual comprehension, and potential domain biases of current mainstream LLMs in non-English, non-Western contexts.

---

## 🏛️ Creators & Affiliations
* **Authors:** Weize Sun, Xiao Du
* **Affiliations:** * School of Computer Science and Technology, Kashgar University
  * Xinjiang Key Laboratory of Multimodal Intelligent Computing and Large Models

---

## 🏷️ Label Mapping

**Topic Codes (0-4):**
* `0` - News/Society
* `1` - Emotion/Philosophy
* `2` - Family/Life
* `3` - Culture/Ent
* `4` - Economy/Biz

**Sentiment Codes (0-2):**
* `0` - Negative
* `1` - Neutral
* `2` - Positive

---

## 🗂️ Data Dictionary
The core data file is `CHORUS_Uyghur_Dataset.csv`, containing the following key fields:

| Column Name | Description |
| :--- | :--- |
| **`id`** | Unique identifier for each data entry. |
| **`text_raw`** | Raw text with privacy scrubbing (e.g., replaced with `[PHONE]`) and Unicode normalization, retaining original punctuation for human review. |
| **`text_cleaned`** | Cleaned text with all punctuation removed from `text_raw`, designed specifically as direct input for model training (e.g., SentencePiece, SVM). |
| **`label_topic_[LLM]`** | Original topic prediction (0-4) independently generated by Claude, Gemini, ChatGPT, or Deepseek. |
| **`label_sentiment_[LLM]`** | Original sentiment prediction (0-2) independently generated by Claude, Gemini, ChatGPT, or Deepseek. |
| **`source`** | Data collection platform (e.g., Douyin, Kuaishou, Tianshannet). |
| **`Parent Company/Organization`** | The entity behind the source platform, useful for domain bias analysis. |
| **`topic_status` / `sentiment_status`**| The voting status recording the level of consensus reached among the models (e.g., Majority_Win (4vs0), Tie). |
| **`final_topic` / `final_sentiment`** | The final ground truth label for Topic (0-4) and Sentiment (0-2). |

---

## 📜 License
This dataset is distributed under the **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)** license. You are free to share and adapt the material for non-commercial purposes, provided you give appropriate credit.

---

## 📝 Citation
If you use this dataset in your research or project, please cite it using the following BibTeX entry:

```bibtex
@dataset{chorus_uyghur_dataset_2026,
  author       = {Sun, Weize and Du, Xiao},
  title        = {CHORUS: Uyghur Text Classification Dataset},
  year         = {2026},
  publisher    = {Zenodo},
  note         = {School of Computer Science and Technology, Kashgar University \& Xinjiang Key Laboratory of Multimodal Intelligent Computing and Large Models},
  doi          = {10.5281/zenodo.18804285},
  url          = {[https://doi.org/10.5281/zenodo.18804285](https://doi.org/10.5281/zenodo.18804285)}
}