A General Framework for Information Extraction using Dynamic Span Graphs

TL;DR

This paper presents a versatile framework for information extraction that uses dynamically built span graphs to improve entity, relation, and coreference detection, outperforming previous methods across various datasets.

Contribution

The authors propose a novel dynamic span graph approach that propagates confidence scores to refine span representations, enhancing multi-task information extraction performance.

Findings

Significant outperformance of state-of-the-art methods on multiple datasets.

Effective detection of nested span entities with improved F1 scores.

Dynamic propagation of confidence scores improves span representation quality.

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.