[Paper Review] A General Framework for Information Extraction using Dynamic Span Graphs

Yi Luan, Dave Wadden+4 more|arXiv (Cornell University)|Apr 5, 2019

Topic Modeling35 references37 citations

TL;DR

DyGIE introduces dynamic span graphs to jointly extract entities, relations, and coreference, yielding state-of-the-art results across multiple domains and handling overlapping spans well.

ABSTRACT

We introduce a general framework for several information extraction tasks that share span representations using dynamically constructed span graphs. The graphs are constructed by selecting the most confident entity spans and linking these nodes with confidence-weighted relation types and coreferences. The dynamic span graph allows coreference and relation type confidences to propagate through the graph to iteratively refine the span representations. This is unlike previous multi-task frameworks for information extraction in which the only interaction between tasks is in the shared first-layer LSTM. Our framework significantly outperforms the state-of-the-art on multiple information extraction tasks across multiple datasets reflecting different domains. We further observe that the span enumeration approach is good at detecting nested span entities, with significant F1 score improvement on the ACE dataset.

Motivation & Objective

• Motivate joint information extraction across tasks that share span representations.
• Propose a general, domain-agnostic framework that refines span representations via dynamic graphs.
• Enable propagation of context through coreference and relation links to improve predictions.
• Demonstrate strong empirical performance across multiple datasets and domains.

Proposed method

• Enumerate all candidate spans up to a max length and compute initial span representations.
• Construct a dynamic span graph by selecting high-confidence spans as nodes and linking them with confidence-weighted coreference and relation edges.
• Propagate context through coreference and relation layers with iterative updates to refine span representations.
• Use gated updates to combine old and new information for each propagation step.
• Predict entity types and relation types from refined span representations and jointly train with coreference supervision.
• Train with a weighted sum of log-likelihood losses for entities, relations, and coreference.

Experimental results

Research questions

• RQ1 Can a dynamic span-graph approach improve cross-task information extraction (entities, relations, coreference) over traditional shared-encoder multi-task models?
• RQ2 Does propagating context via coreference and relation links enhance span representations and downstream predictions across diverse domains?
• RQ3 Is the framework effective at detecting overlapping spans and in domains lacking external syntactic tools?
• RQ4 What is the impact of coreference vs relation propagation on entity and relation extraction?

Key findings

• Achieves state-of-the-art performance on joint entity and relation extraction across ACE04, ACE05, SciERC, and Wet Lab Protocol Corpus.
• Demonstrates substantial improvements in overlapping entity extraction on ACE04-O, ACE05-O, and GENIA compared to prior methods.
• On ACE04 and ACE05, DyGIE yields 7.1% and 7.0% relative improvements in NER and 25.8% and 13.7% in RE, respectively.
• On SciERC, gains of 5.9% in relation extraction and 1.9% in NER over prior models.
• On WLPC, DyGIE improves relation extraction by 16.8% and NER by 2.2% without gold entity boundaries.
• Coreference propagation notably helps pronoun disambiguation and cross-sentence context for entity typing, while relation propagation robustly boosts relation extraction, especially in sentences with many entities.