Current approaches to natural language processing remain strongly focused on the local lexical semantic level, causing them to underperform in the  automatic modeling of discourse information. In order to also enable progress in modeling long-distance relations in text,  we believe the modeling of world knowledge based on the information contained in text, speech and images is needed on the one hand side, in addition to a reappraisal of hybrid methodologies (both classical linguistics-informed ML systems and language model-based approaches). In this project, we want to build on the expertise we have accumulated in recent years within LT3 in the areas of automatic event detection, coreference resolution, aspect-based and event-based emotion detection, among others, in order to achieve better modeling of discourse, with application within automatic summarization, content-based recommendation algorithms, disinformation detection, etc.