Information Extraction for Unstructured Text: From Closed to Open

The talk introduces information extraction (IE) as the process of converting natural‑language text into structured triples , a key step for building high‑quality knowledge graphs and supporting downstream applications such as question answering and decision making.

IE tasks are divided into closed IE, where the relation set is predefined, and open IE, where relations are not fixed. Closed IE further includes relation extraction (given entity pairs) and joint extraction (extracting entities and relations together).

Relation Extraction : challenges include focusing on the correct entity pair and filtering noisy mentions. Recent work introduces segment‑level attention and global‑rationale enhancement (e.g., CRF‑based attention, auxiliary entity‑type and trigger‑word prediction) to improve accuracy on benchmarks like TACRED.

Joint Extraction : addresses overlapping triples by task decomposition (ETL) and by decoupling entity and relation prediction (TPLinker), using two‑dimensional matrices to mark entity boundaries and relation links, achieving better handling of entity‑pair overlaps.

Open Extraction : explores semi‑open and fully open IE, proposing non‑autoregressive maximal‑clique methods that model triples as maximal cliques in a fact graph built from segment nodes and edge predictions, eliminating exposure bias and cascade errors.

Experimental results on datasets such as TACRED, DialogRE, OpenIE4, and SAOKE demonstrate significant improvements over prior baselines across all three IE paradigms.

The presentation concludes that IE remains a crucial component of knowledge‑graph construction and that continued research on segment‑aware modeling, global reasoning, and graph‑based decoding is essential for further advances.