Extraction de relations basée sur des prompts pour la pharmacovigilance

Extracting and maintaining up-to-date knowledge from diverse linguistic sources is imperative for the benefit of public health. While professional sources, including scientific journals and clinical notes, provide the most reliable knowledge, observations reported in patient forums and social media can bring complementary information for certain themes. Spotting entities and their relationships in these varied sources is particularly valuable. We focus on relation extraction in the medical domain. At the outset, we highlight the inconsistent terminology in the community and clarify the diverse setups used to build and evaluate relation extraction systems. To obtain reliable comparisons, we compare systems using the same setup. Additionally, we conduct a series of stratified evaluations to further investigate which data properties affect the models' performance. We show that model performance tends to decrease with relation density, relation diversity, and entity distance. Subsequently, this work explores a new training paradigm for biomedical relation extraction: prompt-based methods with masked language models. In this context, performance depends on the quality of prompt design. This requires manual efforts and domain knowledge, especially when designing the label words that link model predictions to relation classes. To overcome this overhead, we introduce an automated label word generation technique leveraging a dependency parser and training data. This approach minimizes manual intervention and enhances model performance with fewer parameters to be fine-tuned. Our approach performs on par with other verbalizer methods without additional training. Then, this work addresses information extraction from text written by laypeople about adverse drug reactions. To this end, as part of a joint effort, we have curated a tri-lingual corpus in German, French, and Japanese collected from patient forums and social media platforms. The challenge and the potential applications of the corpus are discussed. We present baseline experiments on the corpus that highlight three points: the effectiveness of a multilingual model in the cross-lingual setting, preparing negative samples for relation extraction by considering the co-reference and the distance between entities, and methods to address the highly imbalanced distribution of relations. Lastly, we integrate information from a medical knowledge base into the prompt-based approach with autoregressive language models for biomedical relation extraction. Our goal is to use external factual knowledge to enrich the context of the entities involved in the relation to be classified. We find that general models particularly benefit from external knowledge. Our experimental setup reveals that different entity markers are effective across different corpora. We show that the relevant knowledge helps, though the format of the prompt has a greater impact on performance than the additional information itself.