Conceptual

Knowledge-Augmented Document-Level Relation Extraction with Comprehensive Reasoning

KnowRA unifies the reasoning abilities needed for document-level relation extraction by building a heterogeneous multi-level document graph, adding co-reference resolution, and expanding it into a document knowledge graph with retrieved external knowledge. A knowledge filtration step autonomously decides whether to accept possibly wrong external knowledge, and an axial attention mechanism captures direct and indirect associations through intermediary entities for cross-sentence logical reasoning.