nllg/arg_emo_fallacy

---
license: cc-by-4.0
task_categories:
- text-classification
- zero-shot-classification
language:
- en
size_categories:
- 1K<n<10K
---
This dataset accompanies the paper [Emotionally Charged, Logically Blurred: AI-driven Emotional Framing Impairs Human Fallacy Detection](https://arxiv.org/abs/2510.09695). It includes annotations for logical fallacy labels, emotion categories, and argument convincingness ratings. Please refer to the paper and its [repository](https://github.com/NL2G/EMCONA-UTN/tree/main/emotion_fallacy) for more details.

If you use this dataset, please include the following citation：
```
@misc{chen2026emotionallychargedlogicallyblurred,
title={Emotionally Charged, Logically Blurred: AI-driven Emotional Framing Impairs Human Fallacy Detection},
author={Yanran Chen and Lynn Greschner and Roman Klinger and Michael Klenk and Steffen Eger},
year={2026},
eprint={2510.09695},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2510.09695},
}
```

Relevant columns in the `.tsv` file:

| Column | Description | Notes |
|---|---|---|
| `id_ori` | ID of the original argument (source item). | All rows derived from the same original argument share the same `id_ori`. |
| `id_gen` | ID of the variant within an original argument group. | Values range from **-1 to 3** (**-1** = original argument; **0–3** = synthetic variants). |
| `batch` | Annotation batch ID. | Values: **1–20**. |
| `model_gen` | Model used to generate the synthetic argument. | For original arguments (`id_gen = -1`), this is `"N/A"`. |
| `strategy_gen` | Emotional framing strategy used for synthetic generation. | |
| `emotion_gen` | Target emotion specified for synthetic generation. | |
| `fallacy_gold` | Gold fallacy label from the original dataset. | All variants from the same original argument share this label. |
| `argument` | Argument text (original or synthetic). | |
| `claim` | Claim associated with the argument (generated by LLMs). | All variants from the same original argument share the same claim. |
| `emo_0`, `emo_1`, `emo_2` | Emotion labels from annotators 0/1/2. | |
| `fallacy_0`, `fallacy_1`, `fallacy_2` | Fallacy labels from annotators 0/1/2. | |
| `conv_0`, `conv_1`, `conv_2` | Convincingness ratings from annotators 0/1/2. | If an annotator judged the claim does **not** match the argument, they skip this rating and the dataset uses the placeholder value **`100`**. |
| `conv_zscore_0`, `conv_zscore_1`, `conv_zscore_2` | Z-score–normalized convincingness ratings per annotator. | Same missing/skip rule as above: **`100`** indicates “not annotated”. |
| `emo_best_annotator`, `fallacy_best_annotator`, `conv_best_annotator` | Annotator ID with the highest agreement within a batch. | Values: **0, 1, 2**. |
| `emo_final`, `fallacy_final`, `conv_final`| Final labels. | Majority vote + best-annotator rule / average (as defined in the paper). |