Statutory article retrieval plays a crucial role in making legal information more accessible to both laypeople and legal professionals. Multilingual countries like Belgium present unique challenges for retrieval models due to the need for handling legal issues in multiple languages. Building on the Belgian Statutory Article Retrieval Dataset (BSARD) in French, we introduce the bilingual version of this dataset, bBSARD. The dataset contains parallel Belgian statutory articles in both French and Dutch, along with legal questions from BSARD and their Dutch translation. Using bBSARD, we conduct extensive benchmarking of retrieval models available for Dutch and French. Our benchmarking setup includes lexical models, zero-shot dense models, and fine-tuned small foundation models. Our experiments show that BM25 remains a competitive baseline compared to many zero-shot dense models in both languages. We also observe that while proprietary models outperform open alternatives in the zero-shot setting, they can be matched or surpassed by fine-tuning small language-specific models. Our dataset and evaluation code are publicly available.

Conventional techniques for detecting online hate speech rely on the availability of a sufficient number of annotated instances, which can be costly and time consuming. For this reason, zero-shot or few-shot detection can offer an attractive alternative. In this paper, we explore a zero-shot detection approach based on natural language inference (NLI) models. Since the performance of the models in this approach depends heavily on the choice of a hypothesis, our goal is to determine which factors affect the quality of detection. We conducted a set of experiments with three NLI models and four hate speech datasets. We demonstrate that a zero-shot NLI-based approach is competitive with approaches that require supervised learning, yet they are highly sensitive to the choice of hypothesis. In addition, our experiments indicate that the results for a set of hypotheses on different model-data pairs are positively correlated, and that the correlation is higher for different datasets when using the same model than it is for different models when using the same dataset. These results suggest that if we find a hypothesis that works well for a specific model and domain or for a specific type of hate speech, we can use that hypothesis with the same model also within a different domain. While, another model might require different suitable hypotheses in order to demonstrate high performance.