TrendFact

TrendFact is the Chinese fact-checking benchmark from our ACL 2026 paper TrendFact: A Benchmark Towards Hotspot Perception in Automatic Fact-Checking.

Unlike earlier datasets built mostly on Wikipedia, most claims in TrendFact are drawn from trending platforms, so every sample comes with real dissemination signals. We use it to study a mostly overlooked question: when claims differ in social impact, should — and can — a fact-checking system adapt how much reasoning effort it spends on them? We call this the Hotspot Perception Ability (HPA).

Data

• 7,643 samples across five domains: public health, science, society, politics, and culture
• an evidence library of 366,634 entries
• supports three tasks: evidence retrieval, fact verification, and explanation generation

TrendFact.json

The main data. Each sample has the following fields:

Field	Description
claim	the claim to be verified
label	0 = SUPPORT, 1 = REFUTE, 2 = NEI (not enough info)
evidence	gold evidence, with text / url / target_idx
explanation	human-annotated explanation
domain	domain the claim belongs to
views / Discussion / Engagemen / Post	four hotspot indicators
influence_score / final_influence_score	influence scores
ori_title / url / date	original title, source link, and date of the claim

evidence.parquet

The evidence library, 366,634 entries in total, with fields url / text / publish_date. Used for the evidence retrieval task.

Citation

@inproceedings{zhang2026trendfact,
  title={TrendFact: A Benchmark Towards Hotspot Perception in Automatic Fact-Checking},
  author={Zhang, Xiaocheng and Wang, Xi and Lu, Yifei and Wang, Jianing and Ye, Zhuangzhuang and Bao, Mengjiao and Yan, Peng and Su, Xiaohong},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={26494--26513},
  year={2026}
}

Code: https://github.com/zxc123cc/TrendFact