pdjohn/bundestag-causal-attribution

Bundestag Causal Attribution Dataset

This dataset contains annotated sentences from the German Bundestag (parliamentary proceedings) with fine-grained causal relation annotations. It is designed for training multi-task models that perform both span detection (identifying causal indicators and entities) and relation classification (determining the type of causal relationship).

Dataset Structure

Data Splits

Split	Token Samples	Relation Samples	Total
Train	141	464	605
Test	36	117	153

Features

Each sample contains:

• sentence (string): The text of the sentence
• task (string): Either "token" or "relation"
• spans (list): For token tasks - character-level spans with:
  • start (int): Start character position
  • end (int): End character position
  • type (string): Either "INDICATOR" or "ENTITY"
• indicator (string): For relation tasks - the causal indicator phrase
• entity (string): For relation tasks - the entity involved in the causal relation
• relation (int): For relation tasks - the relation type ID (0-13)

Multi-Task Structure

The dataset combines two tasks in a single structure:

• Token-level samples (task="token"): Contains all causal spans (indicators and entities) for a sentence
• Relation-level samples (task="relation"): Contains individual indicator-entity-relation triples

This design enables efficient multi-task learning where both tasks share the same sentence encoder but have separate task-specific heads. The connection between token and relation samples is implicit through the sentence text, following standard multi-task learning practices.

Relation Types and Distribution

The dataset includes 14 fine-grained causal relation types. The table below shows the formal definition of each relation type alongside empirical counts across data splits:

ID	Relation Type	Description	Train	Test	Total
0	NO_RELATION	No causal relationship	8	1	9
1	MONO_POS_CAUSE	Monotonic positive cause	137	34	171
2	DIST_POS_CAUSE	Distributed positive cause	26	10	36
3	PRIO_POS_CAUSE	Priority positive cause	18	2	20
4	MONO_NEG_CAUSE	Monotonic negative cause	7	3	10
5	DIST_NEG_CAUSE	Distributed negative cause	3	0	3
6	PRIO_NEG_CAUSE	Priority negative cause	0	0	0
7	MONO_POS_EFFECT	Monotonic positive effect	147	40	187
8	DIST_POS_EFFECT	Distributed positive effect	10	4	14
9	PRIO_POS_EFFECT	Priority positive effect	2	1	3
10	MONO_NEG_EFFECT	Monotonic negative effect	92	18	110
11	DIST_NEG_EFFECT	Distributed negative effect	1	0	1
12	PRIO_NEG_EFFECT	Priority negative effect	1	0	1
13	INTERDEPENDENCY	Bidirectional causal relationship	12	4	16
	TOTAL		464	117	581

Note on Class Distribution: The dataset exhibits natural class imbalance reflecting real-world causal discourse patterns. The three most common relations (MONO_POS_EFFECT, MONO_POS_CAUSE, MONO_NEG_EFFECT) account for approximately 80% of all annotations. Three rare relation types (DIST_NEG_CAUSE, DIST_NEG_EFFECT, PRIO_NEG_EFFECT) appear only in the training set with 1-3 samples each. When using this dataset with CausalBERT, class weights are automatically computed using inverse-frequency weighting during training (see train.py) to mitigate the effects of class imbalance. For other frameworks, consider implementing similar class weighting or data augmentation strategies for rare classes.

Example Usage

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("pdjohn/bundestag-causal-attribution")

Example Entries

Token Task Example:

{
  "sentence": "Das Artensterben bringt das Erreichen der SDGs der Agenda 2030 in Gefahr.",
  "task": "token",
  "spans": [
    {"start": 4, "end": 16, "type": "ENTITY"},      # "Artensterben"
    {"start": 17, "end": 23, "type": "INDICATOR"},  # "bringt"
    {"start": 42, "end": 46, "type": "ENTITY"}      # "SDGs"
  ],
  "indicator": null,
  "entity": null,
  "relation": null
}

Relation Task Example:

{
  "sentence": "Das Artensterben bringt das Erreichen der SDGs der Agenda 2030 in Gefahr.",
  "task": "relation",
  "spans": null,
  "indicator": "bringt",
  "entity": "SDGs",
  "relation": 10  # MONO_NEG_EFFECT
}

Filtering by Task

# Get only token classification samples
token_samples = dataset['train'].filter(lambda x: x['task'] == 'token')

# Get only relation classification samples
relation_samples = dataset['train'].filter(lambda x: x['task'] == 'relation')

Use Cases

This dataset is suitable for:

• Multi-task learning for causal relation extraction
• Named entity recognition for causal indicators and entities
• Relation classification with fine-grained causal semantics
• German NLP research in the political/environmental domain
• Causal language understanding in formal discourse
• Environmental discourse analysis in parliamentary proceedings

Dataset Creation

Source Data

The data originates from official transcripts of German Bundestag proceedings, filtered for sentences containing specific environmental keywords related to various forms of "dying" or "death" in ecological contexts.

Annotation Process

Annotations were created with focus on:

• Identifying causal indicators (verbs, conjunctions, nouns indicating causation)
• Marking entities involved in causal relations (causes and effects)
• Classifying relations with three dimensions:
  • Polarity: Positive (POS) vs. Negative (NEG)
  • Salience: Monotonic (MONO), Distributed (DIST), or Priority (PRIO)
  • Direction: Cause vs. Effect

Data Processing

The dataset is structured as a multi-task dataset where:

• Token-level tasks contain all spans (indicators and entities) from a sentence
• Relation-level tasks contain individual indicator-entity pairs with their relation types
• Each sentence may generate multiple relation samples if it contains multiple causal relations

Model Compatibility

This dataset was created with CausalBERT (C-BERT) in mind but is model-agnostic and can be used with any multi-task transformer architecture that supports:

• Token classification for span detection
• Text classification for relation classification
• Multi-task learning frameworks

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.

Attribution Requirements:

• Source Data: German Bundestag stenographic reports (publicly available under open data policy)
• Annotations: Original causal relation annotations by Patrick John
• Citation: Please cite this dataset using the citation information provided below

You are free to:

• Share and redistribute the dataset
• Adapt, transform, and build upon the dataset
• Use for commercial purposes

Under the condition that you provide appropriate attribution.

Citation

If you use this dataset, please cite:

@dataset{bundestag_causal_attribution,
  author = {Johnson, Patrick},
  title = {Bundestag Causal Attribution Dataset},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/pdjohn/bundestag-causal-attribution}
}

Contact

For questions or issues, please open an issue on the CausalBERT GitHub repository.

Acknowledgments

This dataset is based on publicly available German Bundestag proceedings. All data has been filtered and annotated specifically for causal relation extraction research.

The research was funded by the Deutsche Forschungsgemeinschaft (DFG) as part of the project FOR 5182 / "Kontroverse Diskurse. Sprachgeschichte als Zeitgeschichte seit 1990". The author is deeply grateful for the financial support that made this work possible. I also thank the members of the research group for their invaluable feedback and stimulating discussions throughout the development of the Causal Semantics framework.