FREGE: Coreference-Aware Document-Level Knowledge Graph Extraction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Building a knowledge graph from a document requires four decisions: which spans mention entities, which mentions refer to the same entity, what type each entity has, and which relations hold between entities. End-to-end systems such as JEREX learn all four by fine-tuning an encoder. We present FREGE (Frozen REpresentations for Graph Extraction), which reads all four from the hidden states and attention of a frozen decoder-only language model with lightweight probes, without updating the model or generating the graph as text. Graph extraction is thus added to an unchanged pretrained model by training only the probes (17.8M parameters with Qwen2.5-7B, whereas JEREX fine-tunes all 0.11B parameters of its encoder), and every node is a cluster of mentions in the document, with a type and relations from the schema. On Re-DocRED, evaluated with the strict entity-level metric of JEREX, FREGE with a frozen Qwen2.5-7B reaches 43.9 F1 on relations with entity types against 45.2 for a retrained JEREX, a difference that is not significant; Mistral-7B-v0.3 and Llama-3.2-3B, with the same probe recipe, give similar results. A few-shot Qwen2.5-7B-Instruct baseline that generates the graph as text reaches 2.4 F1 under our output format and budget. In all four models tested, attention features improve the complete graph through both the relation probe and the mention linker.