05
Robust and fair language models
Research · NLP · Robustness · 18 months
The challenge
Language models are increasingly used in high-stakes settings. Before trusting them, organizations need to know how they behave with imperfect input, whether they work equally well for all users, and whether they can support regulatory requirements.
The research
- 01
AI for regulatory compliance
IJCNN 2025
Applied LLMs to check whether documentation meets regulatory requirements.
- 02
Language proficiency bias
ACII 2023, MIT Media Lab
Showed that emotion recognition systems perform less reliably for non-native speakers.
- 03
Robustness of text classification
IJCNN 2022
Compared how classical and deep learning models hold up when input text is noisy or altered.
Why it matters
This research is the foundation for how I evaluate AI systems in practice: testing for reliability and fairness before deployment, and preparing systems for requirements such as the EU AI Act.
Publications
- Bridging AI and Regulation: Large Language Models for Documentation Compliance Check →
IJCNN 2025 · Rome, Italy · doi.org/10.1109/IJCNN64981.2025.11229064
- Robustness Analysis Uncovers Language Proficiency Bias in Emotion Recognition Systems →
ACII 2023 · MIT Media Lab, Cambridge, MA · doi.org/10.1109/ACII59096.2023.10388123
- Comparing the Robustness of Classical and Deep Learning Techniques for Text Classification →
IJCNN 2022 · Padua, Italy · doi.org/10.1109/IJCNN55064.2022.9892242