inews / README.md
pitehu's picture
Update README.md
50b0dce verified
---
license: cc-by-nc-sa-4.0
language:
- en
task_categories:
- text-classification
- zero-shot-classification
- multiple-choice
- visual-question-answering
tags:
- emotion detection
- news analysis
- personalization
- psychology
- individual differences
- affective computing
size_categories:
- 10K<n<100K
extra_gated_prompt: >-
This dataset contains sensitive demographic and personality information
collected from human participants. Access is restricted to ensure participant
privacy and ethical use. By requesting access, you agree to use this data
responsibly.
extra_gated_fields:
Full Name: text
Email: text
Institution: text
Website / Google Scholar: text
Position:
type: select
options:
- PhD Student
- Postdoc
- Faculty Member
- Industry Researcher
- Other
Supervisor/PI (if applicable): text
I agree to not attempt re-identification of participants: checkbox
I agree to acknowledge this dataset in all publications: checkbox
I agree to implement appropriate data security measures: checkbox
I agree to not share this dataset with third parties and direct readers to this official page: checkbox
extra_gated_heading: Request Access to Full Dataset with Persona Information
extra_gated_description: >-
This dataset contains sensitive personal information. We will evaluate your
application and you will receive email notification once processed.
extra_gated_button_content: Submit Access Request
configs:
- config_name: default
data_files:
- split: train
path: train.csv
- split: dev
path: dev.csv
- split: test_personalization_public
path: test_personalization_public.csv
- split: test_generalization_public
path: test_generalization_public.csv
- split: test_cold_start_public
path: test_cold_start_public.csv
---
# iNews: A Multimodal Dataset for Personalized Affective News Responses
This is the full, gated dataset for the paper: **iNews: A Multimodal Dataset for Modeling Personalized Affective Responses to News** (ACL 2025).
**iNews** is a large-scale, multimodal dataset designed to model *personalized* emotional responses to news content. Unlike datasets with aggregated labels, iNews captures the rich individual variability in affect, providing the granular data needed for developing human-centered AI. The comprehensive persona profiles in this dataset explain **15.2% of annotation variance**, and their inclusion improves zero-shot affective prediction accuracy by up to **7%**.
## ๐Ÿ”— Official Links
| Resource | Link | Description |
| :--- | :--- | :--- |
| ๐Ÿ“„ **Paper** | **[https://arxiv.org/abs/2503.03335](https://arxiv.org/abs/2503.03335)** | The full research paper on arXiv. |
| ๐Ÿ’ป **Code** | **[https://github.com/pitehu/inews](https://github.com/pitehu/inews)** | Official GitHub repo. |
| ๐Ÿ’พ **Public Dataset** | **[https://huggingface.co/datasets/pitehu/inews_public](https://huggingface.co/datasets/pitehu/inews_public)** | Non-gated version with emotion labels only (no persona information). |
## Citation
If you use this dataset in your research, please cite our paper:
```bibtex
@article{hu2025inews,
title={iNews: A multimodal dataset for modeling personalized affective responses to news},
author={Hu, Tiancheng and Collier, Nigel},
journal={arXiv preprint arXiv:2503.03335},
year={2025}
}
```
โš ๏ธ **GATED ACCESS REQUIRED** - Contains sensitive persona data
## ๐ŸŒ Public Version Available
**Just need emotion labels?** Use our **[public repository]([link-to-public-repo](https://huggingface.co/datasets/pitehu/inews_public))** - no application required.
This gated version adds **demographic and personality features** for personalization research.
## Dataset Overview
Complete dataset with individual differences data:
- **Same 12,276 annotations** as public version (*Note: A subset of annotations is withheld for a future workshop shared task.*)
| Feature Category | Description | Examples |
| :--- | :--- | :--- |
| **News Content** | Multimodal posts from major UK news outlets. | Post text, url, headline, outlet information. |
| **Affective Annotations** | Fine-grained emotional responses from each user. | Valence, Arousal, Dominance (1-7), Discrete Emotions, Relevance, Sharing Likelihood. |
| **๐Ÿ‘ค Persona Profiles** | **(This Gated Version Only)** Comprehensive annotator persona. | **Demographics:** Age, gender, income, education.<br>**Personality:** Big Five (BFI-10), PANAS, PERS.<br>**Cognitive:** Cognitive Reflection Test (CRT).<br>**Habits & Beliefs:** News trust, consumption patterns, political affiliation. |
### Data Fields and Codebook
For a detailed explanation of each column in the dataset, please refer to the `survey_codebook.json` file included in this repository. **For most use cases, just use the **System_Prompt** column as the system persona prompt and the **User_Prompt** as the user prompt, before diving into the specific columns.**
### A Note on Multimodal Data
Due to copyright restrictions, the news post images/screenshots are not directly included in the dataset. However, we provide:
1. The original `post_url` linking to the Facebook post.
2. The associated text of the posts
3. All metadata about the posts (engagement information as well as topics)
Researchers interested in multimodal analysis will need to retrieve the images themselves (see github repo linked above).
### Main Splits
| Split | Samples | Users | Posts | Purpose |
|-------|---------|-------|-------|---------|
| `train` | 7,350 | 202 | 2,028 | Model training |
| `dev` | 155 | 30 | 128 | Hyperparameter tuning |
| `test_personalization_public` | 1,641 | 202 | 580 | **Personalization:** Known users, new content. |
| `test_generalization_public` | 1,676 | 59 | 1,410 | **Generalization:** New users, seen content. |
| `test_cold_start_public` | 498 | 59 | 401 |**Cold-Start:** New users, new content. |
### Test Set Scenarios Explained
#### ๐ŸŽฏ **Personalization (`test_personalization`)**
- **Scenario**: A user from the training set rates a completely new news post.
- **Tests**: The model's ability to learn and adapt to a *specific individual's* preferences and response patterns.
- **Real-world Analog**: Recommending a new article to an existing user of your platform.
#### ๐ŸŒ **Generalization (`test_generalization`)**
- **Scenario**: A completely new user rates a news post that was already seen in the training set.
- **Tests**: The model's ability to generalize from *persona profiles* to predict reactions without any prior behavioral data from that user.
- **Real-world Analog**: Predicting how a new user will react to trending or popular content.
#### โ„๏ธ **Cold Start (`test_cold_start`)**
- **Scenario**: A completely new user rates a completely new news post.
- **Tests**: The model's most fundamental reasoning capability, relying solely on the relationship between a persona profile and novel content.
- **Real-world Analog**: The first-time interaction for a new user on a platform with new content. This is the hardest and most realistic "zero-shot" scenario.
*Note: For reproducing our paper's baseline results: please use (`paper_few_shot_dataset`, `paper_test_dataset`).*
## Usage Example
```python
from datasets import load_dataset
# Requires approved access token
dataset = load_dataset("pitehu/inews", use_auth_token=True)
# Access demographic features
sample = dataset['train'][0]
print(f"Age: {sample['Age']}")
print(f"Extraversion: {sample['Extraversion']}")
print(f"Political affiliation: {sample['Political.affiliation..uk.']}")
# Personalization example
demographic_features = ['Age', 'Sex', 'Extraversion', 'Agreeableness']
personality_data = {feat: sample[feat] for feat in demographic_features}
```
**For most use cases, just use the **System_Prompt** column as the system persona prompt and the **User_Prompt** as the user prompt, before diving into the specific columns.**
## Key Research Applications
- **Personalized Content Systems**: User-specific emotion prediction
- **Algorithmic Bias**: Demographic differences in AI systems
- **Political Psychology**: News perception across political groups
- **Cross-cultural Studies**: Emotion perception differences
### โŒ Prohibited Uses
- Re-identification attempts
- Commercial applications
- Political targeting or profiling
- Sharing data with third parties