| --- |
| library_name: transformers |
| license: mit |
| base_model: FacebookAI/xlm-roberta-base |
| tags: |
| - generated_from_trainer |
| metrics: |
| - accuracy |
| - f1 |
| - precision |
| - recall |
| model-index: |
| - name: EUroBerta-xlm |
| results: [] |
| --- |
| |
| <!-- This model card has been generated automatically according to the information the Trainer had access to. You |
| should probably proofread and complete it, then remove this comment. --> |
|
|
| # EUroBerta-xlm |
|
|
| This model is a fine-tuned version of [FacebookAI/xlm-roberta-base](https://huggingface.co/FacebookAI/xlm-roberta-base) on national parliamentary speeches in Europe, using ParlSpeech dataset. |
| It achieves the following results on the evaluation set: |
| - Loss: 0.1751 |
| - Accuracy: 0.9442 |
| - F1: 0.9442 |
| - Precision: 0.9455 |
| - Recall: 0.9442 |
|
|
| ## Model description |
|
|
| EUroBerta-xlm is a fine-tuned multilingual text classification model designed to identify whether a sentence discusses the European Union (EU). |
| Given any sentence as input, the model returns a binary label: EU (the sentence references the European Union) or No EU (the sentence does not discuss the EU). |
|
|
| This model is the first of three steps in a full EU stance classification pipeline: |
|
|
| - EUroBerta-xlm → identifies EU-relevant sentences (EU / No EU) ← this model |
| - EuOppostion_Classifier → classifies EU sentences as Opposition / Non-Opposition |
| - EuSupport_Classifier → classifies Non-Opposition sentences as Neutral / Support |
|
|
| The final model achieves an F1 score of 0.94 on the held-out evaluation set, with high and balanced precision and recall. |
| The confusion matrix below reflects the model's strong ability to distinguish EU-relevant from non-EU sentences, including after correction for false positives introduced by geographic or sporting references to "Europe". |
|
|
| ## Intended uses & limitations |
|
|
| The model was developed to support large-scale computational text analysis of political discourse, enabling researchers to efficiently filter and categorise party manifestos, parliamentary speeches, and other political documents across multiple European languages. |
| Altought the model can be use for any EU-related task, it was specifically designed for assisting annotation pipelines that require downstream classification of EU attitudes (pro-EU, Eurosceptic, neutral). |
|
|
| ## Training and evaluation data |
|
|
| The training data was constructed through two rounds of annotation, resulting in a final corpus of 5,950 sentences drawn from plenary speeches from 7 European national parliaments (Austria, Denmark, Germany, |
| the Netherlands, Spain, Sweden, and the United Kingdom). These speeches are sourced from the ParlSpeech dataset, covering a 20-year period from 1999 to 2019 (Rauh & Schwalbach, 2020) |
|
|
| The source data originally used five categories of EU stance: |
| - No EU — sentence does not reference the EU |
| - Pro-EU — sentence expresses a favourable view of the EU |
| - Neutral — sentence mentions the EU without taking a stance |
| - Soft Eurosceptic — sentence expresses mild scepticism toward the EU |
| - Hard Eurosceptic — sentence expresses strong opposition to the EU |
|
|
| These five categories were consolidated into a binary format for this model: sentences labelled No EU were retained as one class, while Pro-EU, Neutral, Soft Eurosceptic, and Hard Eurosceptic were merged into a single EU class. |
| After an initial training run, the model showed a systematic tendency to overclassify sentences as EU-related, particularly sentences referencing individual countries (including non-EU states) or generic mentions of "Europe" (e.g., UEFA, European geography). To address this, a second round of targeted annotation was conducted: |
| - 250 additional sentences per country were selected for re-annotation. |
| - Radical right parties were intentionally oversampled (~50 sentences per country) to improve coverage of oppositional discourse, which differs structurally from mainstream party language. |
| - The model was then retrained on the full combined corpus of 5,950 sentences. |
|
|
| ## Training procedure |
|
|
| ### Training hyperparameters |
|
|
| The following hyperparameters were used during training: |
| - learning_rate: 2e-05 |
| - train_batch_size: 24 |
| - eval_batch_size: 24 |
| - seed: 42 |
| - optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments |
| - lr_scheduler_type: linear |
| - num_epochs: 3 |
|
|
| ### Training results |
|
|
| | Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 | Precision | Recall | |
| |:-------------:|:-----:|:----:|:---------------:|:--------:|:------:|:---------:|:------:| |
| | 0.2651 | 1.0 | 198 | 0.2666 | 0.9154 | 0.9154 | 0.9185 | 0.9154 | |
| | 0.2581 | 2.0 | 396 | 0.2291 | 0.9171 | 0.9171 | 0.9242 | 0.9171 | |
| | 0.128 | 3.0 | 594 | 0.1751 | 0.9442 | 0.9442 | 0.9455 | 0.9442 | |
|
|
| ### Confusion Matrix |
|
|
|  |
|
|
| ### Framework versions |
|
|
| - Transformers 4.52.4 |
| - Pytorch 2.6.0+cu124 |
| - Datasets 3.6.0 |
| - Tokenizers 0.21.1 |
|
|
