Model Details

Developed by: Faculty of Economics and Business, University of Rijeka

Funding: National Recovery and Resilience Plan, University of Rijeka (Grant No. uniri-mladi-drustv-23-52)

Models included: FastText, NBSVM, BiGRU, BERT, DistilBERT, BERTić (fine-tuned on Croatian-translated datasets)

Languages: Croatian (hr)

Intended domain: Economic news and disinformation detection

Intended Use

These models are released for research and educational purposes. They are designed to support the detection of economic disinformation in Croatian-language texts and should be used to:

Benchmark model performance across architectures.

Explore methods of adapting multilingual and regional models for specialized domains.

Support further academic research on fake news detection, economic disinformation, and AI ethics.

Limitations

Translation bias: The datasets were translated from English to Croatian. While validated by bilingual experts, subtle semantic shifts may affect results.

Annotation consistency: Original labels were preserved, but cultural and contextual interpretation in Croatian media may differ.

Generalizability: Models trained on specific corpora may not perform equally well on real-world, evolving disinformation patterns.

Resource demands: Transformer-based models (BERT, BERTić) require more computational resources compared to lighter models.

Ethical Considerations

Bias risks: Users should be aware that linguistic and cultural nuances may introduce bias into classification outcomes.

Potential misuse: These models could be misapplied for surveillance, censorship, or discrediting legitimate journalism. They are not intended for such purposes.

Responsible AI: Use of the models should follow principles of transparency, fairness, and accountability, in line with the EU AI Act.

Recommendations

Use the models only in non-commercial, academic, or public-good contexts.

Validate results with human expertise, especially in sensitive applications.

Cite this work when using the models in research outputs.

Report any observed errors, biases, or misuse to the maintainers.

Citation

If you use these models, please cite the associated paper: Buterin, V., Čišić, D., & Gržeta, I. (2025). Combating Economic Disinformation with AI: Insights from the EkonInfoChecker Project.

Downloads last month
-
Safetensors
Model size
0.1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support