This model is a fine-tuned version of microsoft/mdeberta-v3-base for Subjectivity Detection in News Articles. It was developed by AI Wizards as part of their participation in the CLEF 2025 CheckThat! Lab Task 1.

The model aims to classify sentences as subjective (opinion-laden) or objective. Its primary strategy involves enhancing transformer-based classifiers by integrating sentiment scores, derived from an auxiliary model, with sentence representations. This approach has been shown to significantly boost performance, especially the subjective F1 score.

For more details, refer to the paper: AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles. The official code repository is available at: https://github.com/MatteoFasulo/clef2025-checkthat.

It achieves the following results on the evaluation set:

Loss: 0.7922
Macro F1: 0.7490
Macro P: 0.7409
Macro R: 0.7602
Subj F1: 0.6402
Subj P: 0.6020
Subj R: 0.6836
Accuracy: 0.7961

Model description

This mdeberta-v3-base-subjectivity-italian model is a transformer-based classifier specifically designed for subjectivity detection in news articles. It distinguishes between subjective (opinion-laden) and objective sentences. The model's innovation lies in augmenting transformer embeddings with sentiment signals from an auxiliary model, leading to consistent performance gains, particularly in the subjective F1 score. It also incorporates robust decision threshold calibration to counter class imbalances prevalent across different languages.

This model was evaluated across monolingual settings (Arabic, German, English, Italian, Bulgarian), zero-shot transfer (Greek, Polish, Romanian, Ukrainian), and multilingual training, demonstrating strong generalization capabilities.

Intended uses & limitations

Intended Uses:

Subjectivity Detection: Classifying sentences in news articles as subjective or objective.
Fact-Checking Pipelines: Serving as a component to identify opinionated content that might require further scrutiny.
Journalism Support: Aiding journalists in analyzing content for bias or sentiment.
Misinformation Combatting: Contributing to systems designed to detect and combat misinformation by flagging subjective claims.

Limitations:

Class Imbalance Sensitivity: While decision threshold calibration was applied, the model's performance can be sensitive to the class distribution of the evaluation data. An initial submission error during the CLEF 2025 challenge illustrated this sensitivity.
Domain Specificity: Optimized for news articles; performance might vary on text from significantly different domains.
Sentiment Model Dependency: The effectiveness of sentiment augmentation depends on the quality and relevance of the auxiliary sentiment model used.

Training and evaluation data

This model was fine-tuned on data from the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles. The training and development datasets were provided for 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 specifically addressed class imbalance, which was a notable characteristic across these languages, by employing decision threshold calibration optimized on the development set.

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	101	0.6818	0.5159	0.6822	0.5414	0.1827	0.6129	0.1073	0.7451
No log	2.0	202	0.5257	0.6735	0.6922	0.7465	0.5977	0.4567	0.8644	0.6912
No log	3.0	303	0.6269	0.7393	0.7355	0.7437	0.6209	0.6043	0.6384	0.7931
No log	4.0	404	0.6516	0.7448	0.7353	0.7597	0.6373	0.5885	0.6949	0.7901
0.4326	5.0	505	0.7883	0.7463	0.7413	0.7522	0.6322	0.6105	0.6554	0.7976
0.4326	6.0	606	0.7922	0.7490	0.7409	0.7602	0.6402	0.6020	0.6836	0.7961

Framework versions

Transformers 4.49.0
Pytorch 2.5.1+cu121
Datasets 3.3.1
Tokenizers 0.21.0

How to use

You can use this model for text classification (subjectivity detection) with the transformers library:

python

from transformers import pipeline

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

# Example usage:
# A subjective sentence
result1 = classifier("Questa è una scoperta affascinante e fantastica!")
print(f"Classification: {result1}")
# Expected output: [{'label': 'SUBJ', 'score': ...}]

# An objective sentence
result2 = classifier("In particolare Volkswagen e Stellantis son o arrivati a cedere il 7% in Borsa.")
print(f"Classification: {result2}")
# Expected output: [{'label': 'OBJ', 'score': ...}]

Citation

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

bibtex

@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}, 
}