Community Reading: Graph-DPEP for Few-Shot DocRE

Community research spotlight

We did not author this paper. We're sharing it because it is relevant to graph data, information extraction, and the problems Curate Labs studies.

Graph-DPEP: Decomposed Plug and Ensemble Play for Few-Shot Document Relation Extraction with Graph-of-Thoughts Reasoning is a good example of LLM-era IE becoming more structured.

The method represents document-level relation outputs as graph-style triplets, decomposes prompting by relation type, uses a verifier to identify missed entity pairs, and then performs an ensemble-style second pass using subgraph reasoning.

Why it matters

Few-shot DocRE is hard because documents contain many candidate entity pairs and relation labels. Asking an LLM to solve the whole space in one prompt is brittle. Graph-DPEP reduces that burden by decomposing the type space and using graph-shaped intermediate objects.

Our community read

The system is operationally heavier than a single model call, but the design is honest about why prompting fails: too many labels, too many candidates, and too little structure.

For low-resource domains, the lesson is useful. If an LLM is doing extraction, give it smaller decision spaces and explicit graph objects to reason over.

Source

• arXiv: 2411.02864