Research on Information Extraction from a Graph Perspective

Information extraction (IE) transforms unstructured text into structured knowledge by identifying entities, relations, and events. Traditional IE tasks include entity recognition, relation extraction, and event extraction, each with specific inputs and outputs.

The talk introduces a graph‑centric view that unifies these sub‑tasks: one‑ary IE is modeled as edge‑type prediction, two‑ary IE as ring structures, and multi‑ary IE as maximal cliques in a graph.

Key techniques discussed include cascade labeling, sequence‑to‑sequence generation, and a novel segment‑level attention mechanism enhanced by conditional random fields to focus on continuous text spans.

Experiments on Chinese and English datasets demonstrate that the graph perspective and the proposed attention‑CRF model achieve superior performance on entity overlap, nested entities, and discontinuous spans, with significant gains over baseline methods.

The framework also extends to semi‑open extraction, where knowledge related to a given head entity is extracted, and to open information extraction, which handles overlapping and discontinuous entities by constructing maximal cliques.

Overall, the work proposes a unified graph‑based modeling of seven IE tasks, bridging graph analysis and natural language processing, and shows strong empirical results across multiple benchmarks.