FCDS: Fusing Constituency and Dependency Syntax into Document-Level Relation Extraction argues that document-level relation extraction needs more than entity mentions and contextual embeddings. The paper combines constituency syntax and dependency syntax so the model can reason over both sentence-level structure and dependency paths.

That is a deliberately structural approach. Constituency signals help aggregate sentence information and identify useful sentences; dependency signals help model local grammatical relations and paths between entities.

Why we're excited

FCDS is valuable because it resists the temptation to treat document-level extraction as "just use a larger encoder." The model injects syntax where the task needs structure: relations can depend on how clauses, entities, and sentences compose.

The paper evaluates on DocRED, CDR, and GDA, reporting strong performance across general and biomedical relation extraction settings.

Our community read

For domains with formal prose, scientific writing, or biomedical text, syntax-aware graph modeling remains a serious tool. The likely weakness is brittleness: parser quality, noisy OCR, multilingual text, and informal writing can all weaken syntax-heavy systems.

The design lesson is durable, though. Long-context extraction should not only add context length; it should add the right structure for selecting and propagating evidence.

Source

• arXiv: 2403.01886