pip1237/LLM-Chinese-Textual-Disambiguation

Chinese Textual Ambiguity Dataset

This dataset is the accompanying dataset for the paper:

Uncovering the Fragility of Trustworthy LLMs through Chinese Textual Ambiguity

• Paper (arXiv): https://arxiv.org/abs/2507.23121
• Project repository: https://github.com/ictup/LLM-Chinese-Textual-Disambiguation

Dataset Summary

This release contains 925 Chinese textual ambiguity records collected and annotated for research on ambiguity detection, ambiguity understanding, disambiguation, and trustworthy handling of ambiguous Chinese inputs.

The public release contains the human-facing dataset fields only. Internal annotation/review workflow fields and model prediction outputs have been removed.

There is currently no predefined train/validation/test partition in this release. data.csv contains the complete released corpus.

Data Fields

The dataset contains 13 fields:

• 歧义句: ambiguous sentence
• 歧义句及上下文: ambiguous sentence together with its context
• 歧义文本位置: span or expression responsible for the ambiguity
• 歧义原因（解读选项）: explanation of the ambiguity and plausible interpretations
• 歧义句消岐1: first disambiguated rewrite
• 歧义句消岐2: second disambiguated rewrite
• 歧义句消岐3: third disambiguated rewrite, where available
• 歧义句消岐4: fourth disambiguated rewrite, where available
• 绝对索引: source record index
• 歧义类型: human-annotated ambiguity category
• 歧义内容语体类型: register/style annotation
• 歧义内容所属领域: domain annotation
• 母语者消歧难度: native-speaker disambiguation difficulty annotation

Intended Uses

This dataset is intended for research on:

• Chinese textual ambiguity detection
• Ambiguity source localization
• Enumeration of plausible interpretations
• Explicit disambiguation and rewriting
• Evaluation of LLM uncertainty handling and trustworthy response behavior

Limitations

Chinese textual ambiguity is context-sensitive and can admit multiple defensible interpretations. Some annotation fields may be unavailable for individual records. Researchers should inspect the relevant fields before defining task-specific experimental splits or evaluation protocols.

License

The authors should select an explicit dataset license before public release. The license is intentionally not declared in the metadata of this draft Dataset Card.

Citation

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

@misc{wu2025uncoveringfragilitytrustworthyllms,
  title        = {Uncovering the Fragility of Trustworthy LLMs through Chinese Textual Ambiguity},
  author       = {Xinwei Wu and Haojie Li and Hongyu Liu and Xinyu Ji and Ruohan Li and Yule Chen and Yigeng Zhang},
  year         = {2025},
  eprint       = {2507.23121},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2507.23121}
}

Contact

For questions, please use the project repository: https://github.com/ictup/LLM-Chinese-Textual-Disambiguation