As part of our research project on trustworthiness of AI and machine learning, we have successfully handed over our cross-domain event extraction system to our partners at Atos SE.
The event extraction system is designed to transform unstructured text into structured event representations. Our cross-domain event extraction system effectively handles heterogeneous datasets while providing interpretable and flexible workflows. Rather than relying on separate models for each domain event scheme, our system uses a single model that is trained jointly across multiple datasets. This enables our system to operate on six diverse datasets without the need to persist domain-specific models. Secondly, our system supports both pipeline and end-to-end configurations. Users can run sequential trigger identification and classification, or joint detection; visualise results; compare extraction modes; and export structured outputs. We are looking forward to extending this framework to multimodal scenarios and additional downstream tasks in multimodal knowledge reasoning.
The system is open source and available on GitHub. Looking ahead, we plan to extend the framework to multimodal scenarios and additional tasks in multimodal knowledge reasoning.
