A Multimodal Text- and Graph-Based Approach for Open-Domain Event Extraction from Documents

The key tension the summary glosses over is the 'open-domain' qualifier: most event extraction benchmarks test on closed, predefined ontologies, so a system that handles novel event types is solving a fundamentally harder problem than what most LLM-based extraction work targets.

This paper sits in a cluster of research exploring what LLMs still can't do reliably on their own. The DiscoTrace paper from arXiv cs.CL (April 16) made a related observation: LLMs systematically lack the rhetorical selectivity that humans apply when constructing answers from documents, favoring breadth over precision. The multimodal graph approach here is essentially a structural remedy for that same weakness at the document level, using graph representations to enforce relational reasoning that pure token-sequence models tend to shortcut. The IG-Search work from the same period also reinforces this pattern, showing that step-level structural signals improve reasoning quality over trajectory-level ones.

The real test is whether the graph-augmented approach holds on genuinely out-of-distribution event types, not just held-out splits of existing benchmarks. If the authors release evaluation results on a post-2024 news corpus with event categories absent from their training ontology, that would be meaningful evidence the approach generalizes.