RooseBERT-scr-cased

RooseBERT is a domain-specific BERT-based language model pre-trained on English political debates and parliamentary speeches. It is designed to capture the distinctive features of political discourse, including domain-specific terminology, implicit argumentation, and strategic communication patterns.

This variant — scr-cased — was trained from scratch (SCR) with a custom cased WordPiece tokenizer built from the political debate corpus, giving it a vocabulary better suited to political language than standard BERT.

📄 Paper: RooseBERT: A New Deal For Political Language Modelling
💻 GitHub: https://github.com/deborahdore/RooseBERT

---

Model Details

Property	Value
Architecture	BERT-base (encoder-only)
Training approach	From scratch (SCR)
Vocabulary	Custom cased WordPiece (28,996 tokens)
Hidden size	768
Attention heads	12
Hidden layers	12
Max position embeddings	512
Training steps	250K
Batch size	2048
Learning rate	3e-4 (linear warmup + decay)
Training objective	Masked Language Modelling (MLM, 15% mask rate)
Hardware	8× NVIDIA A100 GPUs
Frameworks	HuggingFace Transformers, DeepSpeed ZeRO-2, FP16

The SCR approach trains BERT entirely from scratch on domain-specific data, using a custom tokenizer. This allows political terms like deterrent, endorse, bureaucrat, statutorily, and consequential to be represented as single tokens — whereas standard BERT would split them into multiple sub-tokens. The custom vocabulary shares only ~56% of its tokens with bert-base-cased.

---

Training Data

RooseBERT was pre-trained on 11GB of English political debate transcripts spanning 1919–2025, drawn from:

Source	Coverage	Size
African Parliamentary Debates (Ghana & South Africa)	1999–2024	573 MB
Australian Parliamentary Debates	1998–2025	1 GB
Canadian Parliamentary Debates	1994–2025	1.1 GB
European Parliamentary Debates (EUSpeech)	2007–2015	110 MB
Irish Parliamentary Debates	1919–2019	~3.4 GB
New Zealand Parliamentary Debates (ParlSpeech)	1987–2019	791 MB
Scottish Parliamentary Debates (ParlScot)	–2021	443 MB
UK House of Commons Debates	1979–2019	2.6 GB
UN General Debate Corpus (UNGDC)	1946–2023	186 MB
UN Security Council Debates (UNSC)	1992–2023	387 MB
US Presidential & Primary Debates	1960–2024	16 MB

All datasets were sourced from authoritative, official political settings. Pre-processing removed hyperlinks, markup tags, and collapsed whitespace.

---

Intended Use

RooseBERT is intended as a base model for fine-tuning on downstream NLP tasks related to political discourse analysis. It is especially well-suited for:

• Sentiment Analysis of parliamentary speeches and debates
• Stance Detection (support/oppose classification)
• Argument Component Detection and Classification (claims and premises)
• Argument Relation Prediction and Classification (support/attack/no-relation)
• Motion Policy Classification
• Named Entity Recognition in political texts

---

How to Use

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("ddore14/RooseBERT-scr-cased")
model = AutoModelForMaskedLM.from_pretrained("ddore14/RooseBERT-scr-cased")

For fine-tuning on a downstream classification task:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, TrainingArguments, Trainer

tokenizer = AutoTokenizer.from_pretrained("ddore14/RooseBERT-scr-cased")
model = AutoModelForSequenceClassification.from_pretrained(
    "ddore14/RooseBERT-scr-cased",
    num_labels=2
)

# Recommended fine-tuning hyperparameters (from paper):
# learning_rate ∈ {2e-5, 3e-5, 5e-5}
# batch_size ∈ {8, 16, 32}
# epochs ∈ {2, 3, 4}

---

Evaluation Results

RooseBERT was evaluated across 10 datasets covering 6 downstream tasks. Results below are for RooseBERT-scr-cased (Macro F1 unless noted).

Task	Dataset	Metric	RooseBERT-scr-cased	BERT-base-cased
Sentiment Analysis	ParlVote	Accuracy	0.79	0.69
Sentiment Analysis	HanDeSeT	Accuracy	0.71	0.67
Stance Detection	ConVote	Accuracy	0.75	0.72
Stance Detection	AusHansard	Accuracy	0.60	0.54
Arg. Component Det. & Class.	ElecDeb60to20	Macro F1	0.63	0.61
Arg. Component Det. & Class.	ArgUNSC	Macro F1	0.61	0.61
Arg. Relation Pred. & Class.	ElecDeb60to20	Macro F1	0.63	0.58
Arg. Relation Pred. & Class.	ArgUNSC	Macro F1	0.72	0.57
Motion Policy Classification	ParlVote+	Macro F1	0.62	0.54
NER	NEREx	Macro F1	0.88	0.92

RooseBERT outperforms BERT on 9 out of 10 tasks, with improvements ranging from 4% to 15% on classification tasks. NER performance is comparable, as the NEREx dataset uses general (non-political) entity categories. Results are averaged over 5 runs with different random seeds.

Perplexity on held-out political debate data:

Model	Perplexity (cased)
BERT-base-cased	22.11
ConfliBERT-cont-cased	4.37
RooseBERT-cont-cased	2.61
RooseBERT-scr-cased	2.80

---

Available Variants

Model	HuggingFace ID
RooseBERT-scr-cased (this model)	ddore14/RooseBERT-scr-cased
RooseBERT-scr-uncased	ddore14/RooseBERT-scr-uncased
RooseBERT-cont-cased	ddore14/RooseBERT-cont-cased
RooseBERT-cont-uncased	ddore14/RooseBERT-cont-uncased

SCR (from scratch) models use a custom political vocabulary; CONT (continued pre-training) models are initialised from the original BERT weights and vocabulary.

---

Limitations

• RooseBERT is trained exclusively on English political debates. Cross-lingual use is not supported.
• The model may reflect biases present in official political speech, including over-representation of certain geopolitical perspectives.
• Performance on NER tasks does not benefit from domain-specific pre-training when entity categories are general rather than politically specific.
• As with all encoder-only models, RooseBERT is best suited to classification and labelling tasks rather than generation.

---

Citation

If you use RooseBERT in your research, please cite:

@article{dore2025roosebert,
  title={RooseBERT: A New Deal For Political Language Modelling},
  author={Dore, Deborah and Cabrio, Elena and Villata, Serena},
  journal={arXiv preprint arXiv:2508.03250},
  year={2025}
}

---

Acknowledgements

This work was supported by the French government through the 3IA Côte d'Azur programme (ANR-23-IACL-0001). Computing resources were provided by GENCI at IDRIS (grant 2026-AD011016047R1) on the Jean Zay supercomputer.