Enhancing Structured Meaning Representations with Aspect Classification

A research team from multiple institutions has published a groundbreaking paper introducing the first comprehensive dataset for aspect classification within structured semantic representations. The work, led by Claire Benét Post and Paul Bontempo with seven other collaborators, addresses a critical gap in how AI systems understand language by adding temporal dimension to meaning graphs. Their 15-page paper details a new annotation scheme that labels eventive predicates according to the UMR aspect lattice, providing crucial distinctions between states ("knows"), activities ("running"), and completed events ("built").

The researchers developed a sophisticated annotation pipeline with multi-step adjudication to ensure consistency across annotators, resulting in high-quality labels for English sentences over AMR graphs that previously lacked aspectual information. To demonstrate practical utility, they conducted baseline experiments using three modeling approaches, establishing the first benchmarks for automatic UMR aspect prediction. This dataset enables future AI systems to move beyond static semantic parsing to dynamic understanding of how events unfold over time—a capability essential for applications like advanced question answering, narrative understanding, and temporal reasoning in conversational AI.