Research
My work develops retrieval-augmented and synthetic-data methods for natural language processing (NLP) and artificial intelligence (AI). The same core idea runs through several areas. A model is given relevant prior examples, retrieved by similarity, and learns to adapt them to the case at hand. Alongside this, I explore several directions for creating and using synthetic data across tasks. These include back-translation and LLM-based generation to expand training data for machine translation, neurosymbolic contrastive data for studying fairness, and a closer look at which properties of synthetic data actually improve downstream quality. The result is that smaller, specialized models can reach or exceed much larger general-purpose models on specialized and low-resource tasks.
Machine translation
Fuzzy-match augmentation retrieves similar past translations from a translation memory and adds them to a model’s training data and prompts. My work shows that this is effective for both neural MT and large language models, that it combines well with back-translation and synthetic data, and that it holds under domain shift and severe low-resource conditions.
Educational NLP
The same retrieval idea transfers to automated grading of short-answer language-learning exercises. Instead of retrieving similar translations, the system retrieves how similar student answers were graded before. This line produced the ShAnEL-2 dataset and a study of fuzzy semantic retrieval strategies for grading.
Fairness in language models
I work on fairness in language models, in particular how they choose gender for ambiguous referents in translation and how synthetic contrastive data can reduce gender bias.
Construction grammar
A separate line evaluates computational construction grammars on semantic frame extraction and semantic role labeling, using standard benchmarks. It studies how different parsing heuristics affect the quality of the extracted meaning.
Infrastructure and compute
Across these areas I work with retrieval and similarity search, synthetic data generation, parameter-efficient fine-tuning, and large-scale training and inference.
I run experiments on the Flemish Tier-1 supercomputers, Hortense and the newer Sofia cluster, on NVIDIA A100 and H200 GPUs. My approved allocations include Tier-1 project grants and Starting Grants, held as principal applicant and as a collaborator on several projects across the department.