RooseBERT-cont-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 — cont-cased — was trained via continued pre-training (CONT) of bert-base-cased, initialising from its original weights and vocabulary and training for additional steps on the political debate corpus. This allows the model to leverage BERT's general language understanding while adapting its representations to the political domain. The cased variant preserves capitalisation, making it sensitive to distinctions between proper nouns, acronyms, and common words — particularly relevant in political text where named entities and institutional references are abundant.

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

Model Details

The CONT approach initialises from bert-base-cased's pre-trained weights and continues training on the political debate corpus, retaining BERT's standard vocabulary. This means the model benefits from BERT's broad linguistic knowledge while adapting its contextual representations to the political domain — without the overhead of training a new tokenizer or initialising weights from random. CONT models require fewer training steps (150K vs. 250K for SCR) and can be trained in approximately 24 hours on 8× A100 GPUs.

Training Data

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

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

The CONT-cased variant is a strong general-purpose choice for political NLP tasks. It achieves the lowest perplexity of all RooseBERT variants (cased or uncased), reflecting the strongest overall adaptation to the political debate domain. It is particularly recommended for tasks where capitalisation carries semantic weight, such as NER, or where compatibility with the standard BERT vocabulary is required.

How to Use

from transformers import AutoTokenizer, AutoModelForMaskedLM

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

For fine-tuning on a downstream classification task:

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("ddore14/RooseBERT-cont-cased")
model = AutoModelForSequenceClassification.from_pretrained(
    "ddore14/RooseBERT-cont-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-cont-cased (Macro F1 unless noted).

RooseBERT-cont-cased outperforms BERT-base-cased on 9 out of 10 tasks, with the strongest gains on stance detection (+9% on AusHansard), motion policy classification (+8%), and sentiment analysis (+7% on HanDeSeT). NER performance is comparable to BERT, as the NEREx dataset uses general rather than politically specific entity categories. Results are averaged over 5 runs with different random seeds.

Perplexity on held-out political debate data:

RooseBERT-cont-cased achieves the lowest perplexity of all models evaluated, including all cased and uncased variants, indicating the strongest adaptation to the political debate domain.

Available Variants

CONT (continued pre-training) models inherit BERT's standard vocabulary and pre-trained weights, requiring fewer training steps. SCR (from scratch) models use a custom political vocabulary that encodes domain-specific terms as single tokens (e.g., deterrent, bureaucrat, statutorily). Cased models preserve capitalisation; uncased models lowercase all input.

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.
• Because CONT models retain BERT's standard vocabulary, domain-specific political terms may still be split into sub-tokens (e.g., deterrent → ['de', '##ter', '##rent']). For richer domain vocabulary encoding, consider the SCR variants.
• Performance on NER tasks does not benefit meaningfully 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.