mdeberta-v3-base-subjectivity-sentiment-multilingual-no-arabic

This model is a fine-tuned version of microsoft/mdeberta-v3-base for Subjectivity Detection in News Articles, as presented in the paper AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles. It was developed by AI Wizards for their participation in the CLEF 2025 CheckThat! Lab Task 1.

Code: https://github.com/MatteoFasulo/clef2025-checkthat

It achieves the following results on the evaluation set:

• Loss: 0.8900
• Macro F1: 0.7969
• Macro P: 0.7988
• Macro R: 0.7953
• Subj F1: 0.7450
• Subj P: 0.7560
• Subj R: 0.7342
• Accuracy: 0.8102

Model description

This model is designed to classify sentences as subjective (opinion-laden) or objective in news articles. This task is a key component in combating misinformation, improving fact-checking pipelines, and supporting journalists.

The primary strategy of this model, based on mDeBERTaV3-base, involves enhancing transformer-based embeddings by integrating sentiment scores derived from an auxiliary model. This sentiment-augmented architecture significantly boosts performance, especially the subjective F1 score. The model also employs decision threshold calibration, optimized on the development set, to effectively address class imbalance prevalent across the datasets.

Key Contributions from the paper:

1. Sentiment-Augmented Fine-Tuning: Enriches embedding-based models by integrating sentiment scores from an auxiliary model, notably improving subjective sentence detection.
2. Diverse Model Coverage: Benchmarked across multilingual BERT variants (like mDeBERTaV3-base) for all official CLEF languages.
3. Threshold Calibration for Imbalance: A simple yet effective method to tune decision thresholds on each language's development data to enhance macro-F1 performance.

Intended uses & limitations

Intended Uses: This model is intended for subjectivity detection in news articles, classifying sentences as subjective or objective. It has been evaluated across:

• Monolingual settings: Arabic, German, English, Italian, and Bulgarian.
• Zero-shot transfer: Tested on unseen languages such as Greek, Polish, Romanian, and Ukrainian.
• Multilingual training.

Its capabilities make it suitable for applications that require discerning factual statements from opinions, such as:

• Combating misinformation by identifying biased or opinionated content.
• Improving automated fact-checking pipelines.
• Assisting journalists in content analysis and report generation.

Limitations:

• The model's performance is optimized for news articles; its effectiveness might vary on text from other domains (e.g., social media, conversational text) with different linguistic characteristics or content styles.
• While decision threshold calibration is used to mitigate class imbalance, extreme imbalances in specific datasets might still affect performance.
• The paper notes an initial submission error where an incorrect train/dev mix was used, leading to under-calibrated thresholds and lower initial scores in the multilingual track. While corrected results are reported, this highlights sensitivity to data preparation and calibration.

Training and evaluation data

The model was fine-tuned on datasets provided for the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles.

• Training and development datasets were provided for monolingual settings in Arabic, German, English, Italian, and Bulgarian.
• Final evaluation included additional unseen languages such as Greek, Romanian, Polish, and Ukrainian to assess the model's generalization capabilities.

The training process involved enhancing sentence representations by integrating sentiment scores from an auxiliary model. Class imbalance, which was a notable characteristic across these languages, was addressed through decision threshold calibration during training.

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	249	0.4619	0.7825	0.7790	0.7920	0.7421	0.6919	0.8	0.7901
No log	2.0	498	0.4544	0.7877	0.7856	0.8021	0.7537	0.6846	0.8384	0.7932
0.4852	3.0	747	0.5219	0.7931	0.7899	0.8052	0.7572	0.6970	0.8288	0.7994
0.4852	4.0	996	0.7055	0.7935	0.7975	0.7904	0.7385	0.7605	0.7178	0.8082
0.2503	5.0	1245	0.8536	0.7883	0.7937	0.7844	0.7306	0.7592	0.7041	0.8040
0.2503	6.0	1494	0.8900	0.7969	0.7988	0.7953	0.7450	0.7560	0.7342	0.8102

Framework versions

• Transformers 4.47.0
• Pytorch 2.5.1+cu121
• Datasets 3.3.1
• Tokenizers 0.21.0

How to use

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

import torch
import torch.nn as nn
from transformers import DebertaV2Model, DebertaV2Config, AutoTokenizer, PreTrainedModel, pipeline, AutoModelForSequenceClassification 
from transformers.models.deberta.modeling_deberta import ContextPooler

sent_pipe = pipeline(
    "sentiment-analysis",
    model="cardiffnlp/twitter-xlm-roberta-base-sentiment",
    tokenizer="cardiffnlp/twitter-xlm-roberta-base-sentiment",
    top_k=None,  # return all 3 sentiment scores
)

class CustomModel(PreTrainedModel):
    config_class = DebertaV2Config
    def __init__(self, config, sentiment_dim=3, num_labels=2, *args, **kwargs):
        super().__init__(config, *args, **kwargs)
        self.deberta = DebertaV2Model(config)
        self.pooler = ContextPooler(config)
        output_dim = self.pooler.output_dim
        self.dropout = nn.Dropout(0.1)
        self.classifier = nn.Linear(output_dim + sentiment_dim, num_labels)

    def forward(self, input_ids, positive, neutral, negative, token_type_ids=None, attention_mask=None, labels=None):
        outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
        encoder_layer = outputs[0]
        pooled_output = self.pooler(encoder_layer)
        sentiment_features = torch.stack((positive, neutral, negative), dim=1).to(pooled_output.dtype)
        combined_features = torch.cat((pooled_output, sentiment_features), dim=1)
        logits = self.classifier(self.dropout(combined_features))
        return {'logits': logits}

model_name = "MatteoFasulo/mdeberta-v3-base-subjectivity-sentiment-multilingual-no-arabic"
tokenizer = AutoTokenizer.from_pretrained("microsoft/mdeberta-v3-base")
config = DebertaV2Config.from_pretrained(
    model_name, 
    num_labels=2, 
    id2label={0: 'OBJ', 1: 'SUBJ'}, 
    label2id={'OBJ': 0, 'SUBJ': 1},
    output_attentions=False, 
    output_hidden_states=False
)
model = CustomModel(config=config, sentiment_dim=3, num_labels=2).from_pretrained(model_name)

def classify_subjectivity(text: str):
    # get full sentiment distribution
    dist = sent_pipe(text)[0]
    pos = next(d["score"] for d in dist if d["label"] == "positive")
    neu = next(d["score"] for d in dist if d["label"] == "neutral")
    neg = next(d["score"] for d in dist if d["label"] == "negative")

    # tokenize the text
    inputs = tokenizer(text, padding=True, truncation=True, max_length=256, return_tensors='pt')

    # feeding in the three sentiment scores
    with torch.no_grad():
        outputs = model(
            input_ids=inputs["input_ids"],
            attention_mask=inputs["attention_mask"],
            positive=torch.tensor(pos).unsqueeze(0).float(),
            neutral=torch.tensor(neu).unsqueeze(0).float(),
            negative=torch.tensor(neg).unsqueeze(0).float()
        )

    # compute probabilities and pick the top label
    probs = torch.softmax(outputs.get('logits')[0], dim=-1)
    label = model.config.id2label[int(probs.argmax())]
    score = probs.max().item()

    return {"label": label, "score": score}

examples = [
    "The company reported a 10% increase in revenue for the last quarter.",
    "Die angegebenen Fehlerquoten können daher nur für symptomatische Patienten gelten.",
    "Si smonta qui definitivamente la narrazione per cui le scelte energetiche possono essere frutto esclusivo di valutazioni “tecniche” e non politiche.",
]
for text in examples:
    result = classify_subjectivity(text)
    print(f"Text: {text}")
    print(f"→ Subjectivity: {result['label']} (score={result['score']:.2f})\n")

For more detailed usage, including training and evaluation scripts, please refer to the GitHub repository.

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}, 
}