Postdoc position in Efficient NLP for Information Extraction from News

Descrizione

Testo originale dell'annuncio (in inglese)

News from around the world offers early signals of crime trends — from drug-trafficking hotspots to new modi operandi — but extracting the relevant entities, timelines and relations is challenging, as news sources are scattered, asynchronous, multilingual and multimodal. Furthermore, law enforcement agencies cannot afford to share sensitive data with external parties, so methods must avoid external cloud APIs and be light enough to run on limited hardware. Your job This 24-month postdoctoral position is a close collaboration between the National Police Lab AI ( Floris Bex and Jelte Mense ) and the Natural Language Processing Group ( Albert Gatt ) at Utrecht University. Together, we want to explore how small models can extract useful information on specific crime themes, such as drug trafficking, primarily from publicly available news sources. Relevant tasks include identifying relevant entities (who did what, where), and inferring timelines (when, and in what order) and relations (e.g. links to specific countries or groups). A promising direction is to train or fine-tune small-scale models via distillation from larger models, including through the use of synthetic data. This in turn raises interesting questions about the robustness of the resulting models, their ability to generalise, and their susceptibility to known issues related to synthetic data training, such as model collapse. As a postdoctoral researcher on this project, you will: Design and evaluate methods for multimodal entity, timeline and relation extraction from multilingual news sources. Fine-tune and distil small-scale language models from larger models, including via synthetic data generation. Investigate the robustness, generalisability and limitations (e.g. model collapse) of the resulting models. Apply and validate the approach on real-world use cases, such as drug trafficking, together with the Netherlands police. Publish and present your findings at relevant conferences and in scientific journals.