E2MoCase: A Dataset for Emotional, Event and Moral Observations in News Articles on High-impact Legal Cases

The way media reports on legal cases can significantly shape public opinion, often embedding subtle biases that influence societal views on justice and morality. Analyzing these biases requires a holistic approach that captures the emotional tone, moral framing, and specific events within the narratives. In this work we introduce E2MoCase, a novel dataset designed to facilitate the integrated analysis of emotions, moral values, and events within legal narratives and media coverage. By leveraging advanced models for emotion detection, moral value identification, and event extraction, E2MoCase offers a multi-dimensional perspective on how legal cases are portrayed in news articles.

In this repository, we provide the E2MoCase dataset and the full code to obtain the raw data from the SwissDox platform. The dataset is also available on HuggingFace, where sentence embedding of the source paragraphs, generated with various pretrained language models (e.g., bert-base-uncased, Qwen3-0.6B), along with their annotations can be downloaded.

A preprint paper is available here.

We are continuously refining and expanding the E2moCase dataset. Stay tuned for upcoming updates!

Data Description

E2MoCase contains 97,251 paragraphs extracted from a total of 19,250 news articles. These news articles were obtained from about 100 candidate real-world cases related to legal matters that had significant media impact due to evidence of cultural biases, such as religious, political, gender, racial, and media biases. For each case, we manually verified its accuracy in terms of reported news, we ensured it had significant media impact and it was covered by reputable newspaper agencies.

All paragraphs are labeled with emotions and moralities. Of these, 50,975 paragraphs are also labeled with events, whereas the remaining ones do not contain events. For additional information, please refer to our paper. The statistics of E2MoCase and its variants are shown as follows.

	E2MoCase	E2MoCase_noEvents	E2MoCase_full
# paragraphs	50,975	46,276	97,251
avg # tokens	275.106 ± 245.303	139.402 ± 220.950	210.532 ± 243.647
avg # emotions	1.164 ± 0.757	1.634 ± 0.680	1.678 ± 0.657
avg # morals	3.517 ± 3.870	1.773 ± 1.644	2.795 ± 2.424
avg # events	3.597 ± 2.940	0.0 ± 0.0	1.885 ± 2.785

E2MoCase_noEvents, is the dataset obtained by removing paragraphs that do not contain events, while E2MoCase_full, is the version that also includes paragraphs that do not contain events.

The dataset contains the following columns:

• content_id: Identification code of the news item within SwissDox.
• P : Paragraph identification code. It takes the form $P_i$, where $i$ is the $i$-th paragraph within the news item.
• subject : Main subject of the news item (e.g., Amanda Knox case).
• event : List of events in JSON format
• care, harm, fairness, cheating, loyalty, betrayal, authority, subversion, purity, degradation: Real-valued scores (within 0 and 1) associated with moral values
• anticipation, trust, disgust, joy, optimism, surprise, love, anger, sadness, pessimism, fear: Real-valued scores (within 0 and 1) associated with emotion values

After obtaining a valid SwissDox API key and running the instructions described in the reconstruct_data.ipynb notebook (see the Usage section below), you will also obtain the text column containing the original textual content retrieved from SwissDox.

Example data

Given the following paragraph:

"Mystery without an answer: Where is Sarah's murderer? 
Julia Rossi was acquitted of murdering Sarah Bianchi. 
But if it wasn't her, then who killed the Italian woman with 25 stab wounds?"

An annotated data instance associated with the paragraph is as follows:

event:

[
  {"mention": "murder", "entities": {"Julia Rossi": "murderer", "Sarah Bianchi": "victim"}},
  {"mention": "kill", "entities": {"Julia Rossi": "murderer", "Sarah Bianchi": "victim"}}
]

Moral columns:

care	harm	fairness	cheating	loyalty	betrayal	authority	subversion	purity	degradation
0.0	0.985	0.0	0.901	0.0	0.910	0.0	0.0	0.0	0.221

Emotion columns:

anticipation	trust	disgust	joy	optimism	surprise	love	anger	sadness	pessimism	fear
0.0	0.0	0.521	0.0	0.0	0.0	0.0	0.5	0.0	0.0	0.0

Note: in the above example, all references to real persons have been replaced with fictitious names..

Data access and reproducibility

For this work, we used media data made available via Swissdox@LiRI(see https://www.liri.uzh.ch/en/services/swissdox.html for more information).

The raw news paragraphs required to reproduce the dataset cannot be openly shared due to commercial restrictions. However, the original query (in YAML format) used for retrieving data from Swissdox@LiRI can be found in the folder queries. Additionally, aggregated/derived data (i.e., labels for emotions, morality, and events) for each paragraph are available in the e2mocase.csv file. In the notebook reconstruct_data.ipynb we describe how to retrieve the original text from which the labels for emotions, morality, and events were derived.

You can also download the dataset with the sentence embedding of the source paragraphs, generated with various pretrained language models (e.g., bert-base-uncased, Qwen3-0.6B) here.

Guidelines for human annotation

The file human_annotations/Guidelines.pdf contains the instruction provided to the human annotatore to label the dataset, as described in our paper.

Prompt for event extraction

The prompt descriptions for event extraction are available in the prompts folder.

Hyper-parameters selection

All experiments fine-tuning PLMs were run for five epochs with a batch size of 8, using the Adam optimizer and a learning rate of 5e-5. The MLP with a TF-IDF configuration was run for 50 epochs, using a single hidden layer with 100 units and, where possible, the same hyper-parameters as the experiments with PLMs.

References

This work results from a collaboration between the University of Calabria (UNICAL) and the University of Lausanne (UNIL). The resource paper describing the dataset is currently under review at CIKM 2025.

If you use this resource, please cite:

@misc{candida_maria_greco_2025,
	author       = { Candida Maria Greco and Lorenzo Zangari and Davide Picca and Andrea Tagarelli },
	title        = { E2MoCase (Revision 745e678) },
	year         = 2025,
	url          = { https://huggingface.co/datasets/lorenzozan/E2MoCase },
	doi          = { 10.57967/hf/5819 },
	publisher    = { Hugging Face }
}

You can also refer to the following preprint (dated 2024):

@article{greco2024e2mocase,
  title={E2MoCase: A Dataset for Emotional, Event and Moral Observations in News Articles on High-impact Legal Cases},
  author={Greco, Candida M and Zangari, Lorenzo and Picca, Davide and Tagarelli, Andrea},
  journal={arXiv preprint arXiv:2409.09001},
  year={2024}
}

Also you might refer to the following paper on the topic:

@inproceedings{zangari2025me2,
  title={ME2-BERT: Are Events and Emotions what you need for Moral Foundation Prediction?},
  author={Zangari, Lorenzo and Greco, Candida M and Picca, Davide and Tagarelli, Andrea},
  booktitle={Proceedings of the 31st International Conference on Computational Linguistics},
  pages={9516--9532},
  year={2025}
}