mdeberta-v3-base-subjectivity-english

This model is a fine-tuned version of microsoft/mdeberta-v3-base on the CheckThat! Lab Task 1 Subjectivity Detection at CLEF 2025. It achieves the following results on the evaluation set:

• Loss: 0.5845
• Macro F1: 0.7921
• Macro P: 0.7952
• Macro R: 0.7941
• Subj F1: 0.7885
• Subj P: 0.8364
• Subj R: 0.7458
• Accuracy: 0.7922

Model description

This model is part of AI Wizards' participation in the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles. Its primary goal is to classify sentences as subjective (opinion-laden) or objective.

The model is based on the mDeBERTaV3-base architecture. Its core innovation lies in enhancing transformer-based classifiers by integrating sentiment scores, derived from an auxiliary model, with sentence representations. This approach aims to significantly improve upon standard fine-tuning for subjectivity detection, resulting in consistent performance gains, particularly in subjective F1 score. To further address class imbalance, prevalent across languages, decision threshold calibration optimized on the development set was employed.

Intended uses & limitations

Intended uses:

• Classifying sentences in news articles as subjective or objective.
• Supporting fact-checking pipelines by identifying opinionated content.
• Assisting journalists in distinguishing between facts and opinions.

The model was evaluated across various settings, including monolingual (Arabic, German, English, Italian, and Bulgarian), multilingual, and zero-shot transfer (Greek, Polish, Romanian, and Ukrainian).

Limitations:

• Performance may vary on domains significantly different from news articles, as the model was fine-tuned specifically on news data.
• While the sentiment augmentation significantly boosts performance, the overall effectiveness can depend on the quality and nature of the input text.
• As noted by the authors in their GitHub repository, due to a mistake in their submission process for the multilingual track, the official Macro F1 score was initially lower than its actual performance. The corrected score would have placed them 9th overall in the challenge.
• The model primarily processes text inputs and is not designed for other modalities.

Training and evaluation data

The model was trained and evaluated on datasets provided for the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles.

Training and development datasets were provided for multiple languages including Arabic, German, English, Italian, and Bulgarian. For the final evaluation, additional unseen languages such as Greek, Romanian, Polish, and Ukrainian were included to assess the model's generalization capabilities. The training process involved augmenting transformer embeddings with sentiment signals and applying decision threshold calibration to effectively address class imbalances observed across languages.

How to use

You can use this model for text classification with the transformers library:

from transformers import pipeline

# Load the text classification pipeline
classifier = pipeline(
    "text-classification",
    model="MatteoFasulo/mdeberta-v3-base-subjectivity-english",
    tokenizer="microsoft/mdeberta-v3-base",
)

# Example usage for an objective sentence
text1 = "The company reported a 10% increase in profits in the last quarter."
result1 = classifier(text1)
print(f"Text: '{text1}' Classification: {result1}")
# Expected output: [{'label': 'OBJ', 'score': ...}]

# Example usage for a subjective sentence
text2 = "This product is absolutely amazing and everyone should try it!"
result2 = classifier(text2)
print(f"Text: '{text2}' Classification: {result2}")
# Expected output: [{'label': 'SUBJ', 'score': ...}]

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use 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: 6

Training results

Training Loss	Epoch	Step	Validation Loss	Macro F1	Macro P	Macro R	Subj F1	Subj P	Subj R	Accuracy
No log	1.0	52	0.6894	0.5697	0.7150	0.6233	0.4375	0.875	0.2917	0.6104
No log	2.0	104	0.4827	0.7785	0.7793	0.7782	0.7910	0.7782	0.8042	0.7792
No log	3.0	156	0.4996	0.7922	0.7931	0.7932	0.7931	0.8214	0.7667	0.7922
No log	4.0	208	0.5405	0.7986	0.8027	0.8008	0.7938	0.8483	0.7458	0.7987
No log	5.0	260	0.5838	0.7920	0.7978	0.7948	0.7848	0.8495	0.7292	0.7922
No log	6.0	312	0.5845	0.7921	0.7952	0.7941	0.7885	0.8364	0.7458	0.7922

Framework versions

• Transformers 4.49.0
• Pytorch 2.5.1+cu121
• Datasets 3.3.1
• Tokenizers 0.21.0

Code

The official code and materials for this submission are available on GitHub: https://github.com/MatteoFasulo/clef2025-checkthat

Citation

If you find our work helpful or inspiring, please feel free to cite it:

@misc{fasulo2025aiwizardscheckthat2025,
      title={AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles}, 
      author={Matteo Fasulo and Luca Babboni and Luca Tedeschini},
      year={2025},
      eprint={2507.11764},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.11764}, 
}