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

Entity‑relation extraction is a core task in knowledge‑graph construction and information extraction, aiming to automatically discover semantic links between entities. Simple contexts focus on a single sentence, while complex contexts involve multiple triples within a sentence or across sentences, with at least 40% of facts in Wikipedia requiring multi‑sentence reasoning.

Joint Extraction Methods

1. Sequence labeling : The 2017 ACL paper "Joint Extraction of Entities and Relations Based on Novel Tagging Scheme" introduces a tagging schema combining B/I/E/S tags with head/tail entity indices, yielding 2·4·|R|+1 labels. Models such as LSTM+CRF and encoder‑decoder LSTMs are used, with bias weighting for the "Other" label.

2. Table‑filling : Proposed in EMNLP 2014, this approach represents entities on the diagonal and relations in the lower triangle of a table, but cannot handle overlapping triples (EPO/SEO). Multi‑head selection extends it by using sigmoid classification to allow multiple relation labels per entity pair.

3. Seq2seq (CopyRE) : Treats triple extraction as a translation problem, using an encoder‑decoder with a copy mechanism to directly copy entity tokens from the source. Extensions like CopyMTL improve copying of multi‑token entities, and Seq2UMTree reduces decoding steps by predicting relation types first, then head and tail spans in a tree‑structured decoder.

The seq2seq paradigm suffers from label‑bias due to autoregressive decoding and struggles with overlapping triples, motivating research into sequence‑to‑set formulations.

Document‑Level Relation Extraction

Graph‑based models construct a document graph where each word is a node, with edges for syntactic dependencies, coreference links, adjacency, and self‑loops. GCNN (ACL 2019) learns node representations via graph convolution and aggregates mentions for relation classification.

EOG (EMNLP 2019) builds heterogeneous graphs with mention, entity, and sentence nodes, using weighted paths to propagate information.

LSR (ACL 2020) treats the graph structure as a latent variable, refining it iteratively during end‑to‑end training, achieving competitive results against GCNN, GAT, and AGGCN.

Double Graph (EMNLP 2020) separates mention‑level and entity‑level graphs; the former performs GCN or random walks for intra‑sentence reasoning, while the latter aggregates weighted mention representations for inter‑entity reasoning, achieving state‑of‑the‑art performance.

Experiments on CDR, CHR, and DocRED datasets demonstrate the importance of cross‑sentence edges, graph completeness, and the mitigation of over‑smoothing in GNNs.

Conclusion and Outlook

The presented works highlight progress in joint extraction and document‑level extraction, yet challenges remain such as decoding bias, handling overlapping triples, over‑smoothing in GNNs, and effective information flow in heterogeneous graphs. Future research may explore sequence‑to‑set models, better entity‑pair reasoning, and graph structure learning.