Curate LabsCurate Labs

Community Reading: GraphER as Structure Learning for IE

GraphER shifts joint extraction from tagging or generation toward graph structure learning.

May 9, 2024External researchInformation extractionGraph learning

External research: Curate Labs did not author this paper.

Community Reading: GraphER as Structure Learning for IE

Community Reading: GraphER as Structure Learning for IE

GraphER: A Structure-aware Text-to-Graph Model for Entity and Relation Extraction sits near ATG in the extraction landscape, but its modeling choice is different. Rather than decoding a graph autoregressively, GraphER constructs and refines a graph over candidate spans, then classifies nodes and edges.

The useful shift is to treat joint extraction as graph structure learning. Entity spans and relation candidates are not independent local predictions; they are pieces of a structure that should become coherent together.

Why we're excited

GraphER is a good example of graph inductive bias applied directly to IE. It gives the model a way to reason over candidate nodes and edges before final decisions, which is especially appealing for scientific or biomedical extraction where consistency matters.

The paper reports results on ACE05, CoNLL04, and SciERC and makes its code public.

Our community read

The practical sweet spot is supervised extraction with a known schema and meaningful structural constraints. It is less obviously suited to open-schema, zero-shot, or highly dynamic relation inventories.

The open design question is whether graph-structure learning can be combined with label-text generalization so the system can keep GraphER's structure without becoming benchmark-bound.

Source