Joint Entity and Relation Extraction: Methods and Document‑Level Approaches

Entity relation extraction is a crucial step in knowledge graph construction and information extraction, involving the identification of semantic relationships between entities.

Traditional sentence‑level extraction focuses on simple contexts, while complex contexts involve multiple triples within a sentence or cross‑sentence relations, as illustrated by examples from the DocRED dataset where over 40% of facts require joint extraction.

The talk reviews three major families of joint extraction methods: (1) sequence‑labeling approaches such as the NovelTagging scheme (ACL 2017) that encode relation tags with Begin/Inside/End/Single markers and use LSTM‑CRF models; (2) table‑filling approaches that represent entities and relations in a matrix, later extended with multi‑head selection and sigmoid‑based overlapping relation handling; (3) sequence‑to‑sequence models like CopyRE and its improvements (CopyMTL, Seq2UMTree) that treat triples as generated sequences and employ copy mechanisms to recover multi‑token entities.

Document‑level relation extraction is addressed by building graph representations of entire documents. Early methods aggregate sentence‑level predictions, while later works employ graph neural networks: GCNN (ACL 2019) constructs word‑level graphs with syntactic, coreference, and adjacency edges; EOG (EMNLP 2019) introduces heterogeneous edges among mentions, entities, and sentences; LSR (ACL 2020) learns latent graph structures end‑to‑end; and Double Graph (EMNLP 2020) separates mention‑level and entity‑level graphs, using GCN or random walks followed by MLP classification.

Experimental results on CDR, CHR, and DocRED datasets show that graph‑based models consistently outperform baselines, and that edge design (especially cross‑sentence edges) and graph refinement are critical for performance.

The concluding outlook highlights open challenges such as mitigating label‑bias in seq2seq decoding, exploring sequence‑to‑set formulations, addressing over‑smoothing in GNNs, and improving information flow among heterogeneous nodes in document‑level graphs.