Source-linked AI summary
A General Framework for Information Extraction using Dynamic Span Graphs
Yi Luan, Dave Wadden, Luheng He, Amy Shah, Mari Ostendorf, Hannaneh Hajishirzi
TL;DR
Information extraction systems need to coordinate entity, relation, and coreference decisions while handling spans that may overlap. DYGIE uses dynamically constructed span graphs to propagate confidence-weighted relational and coreference context through shared span representations, achieving state-of-the-art results across tasks and domains, with reported gains on ACE05 and overlapping entities.
Problem
Prior IE systems commonly use pipelines or limited multi-task sharing, which constrains how entity, relation, and coreference information interact.
Method
DYGIE dynamically selects confident mention spans as graph nodes and propagates confidence-weighted relation and coreference information to refine shared span representations.
Results
DYGIE achieves state-of-the-art performance across tasks and domains, including 5.7% and 9.9% relative improvements on ACE05 entity and relation extraction and 11.3% on overlapping entity extraction.
Takeaways & Limitations
The framework couples related IE tasks to learn richer span representations and improves detection of overlapping entities across multiple benchmarks.
Takeaways & Limitations
Coreference propagation can misclassify an entity when the most confident antecedent belongs to the wrong entity category.
Abstract
from arXiv · showhide
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.
1 Introduction
Information extraction tasks identify spans, relations, and coreference links, but pipeline and prior multi-task systems provide limited cross-task interaction. DYGIE dynamically propagates contextual information through span graphs and improves performance across domains and overlapping-entity settings.
- IE tasks identify and categorize phrase spans, assign relation types between span pairs, and cluster spans that refer to the same entity.
- Pipeline systems detect entities before using those spans for relation extraction and coreference resolution, risking cascading errors.
- DYGIE couples entity, relation, and coreference extraction through dynamically constructed span graphs with shared span representations refined by contextual propagation.Highly confident mentions become graph nodes, while relation and coreference links provide confidence-weighted edges.
- Cross-sentence entity and relation contexts can resolve otherwise ambiguous spans, such as linking “This thing” and “it” to the antecedent “car.”Relations involving Tom, Starbucks, and Mike similarly support an inferred location relation for Mike.
- DYGIE achieves state-of-the-art performance across tasks and domains, including relative improvements of 5.7% and 9.9% on ACE05 entity and relation extraction.It also reports an 11.3% relative improvement on ACE05 overlapping entity extraction.
- The framework improves overlapping-entity detection by up to 8 F1 points on ACE04, ACE05, and GENIA.
2 Related Work
Related work shares representations across information extraction tasks, but often depends on indirect interaction, sequence labeling, or domain-specific syntactic tools. DYGIE instead explicitly propagates contextual information through dynamic graphs without external syntactic preprocessing.
- Prior IE research uses joint modeling and multi-task learning to share representational strength across related extraction tasks.
- The closest prior multi-task model shares span representations through gradients to the LSTM, whereas DYGIE explicitly incorporates context through dynamic graph propagation.
- Sequence-labeling entity recognizers commonly struggle to extract entities with overlapping spans.
- Graph-based relation extraction has often incorporated pipelined syntactic or discourse features rather than jointly propagating entity, relation, and coreference information.
- Domain-specific syntactic tools can introduce cascading errors and make relation extraction systems difficult to generalize across domains.
3 Model
DYGIE jointly predicts entities, relations, and coreferences by refining shared span representations with a dynamically constructed span graph. It combines local span encoding with confidence-weighted propagation through coreference and relation links before final predictions.
- Model overview: DYGIE derives entity, relation, and coreference predictions from shared span representations in a multi-task framework.The auxiliary coreference task predicts the best antecedent for each span.
- Dynamic span graph: At each training step, DYGIE selects highly confident entity spans as graph nodes and assigns confidence-based edges for coreference and relation links.Coreference edges connect candidate antecedents, while relation edges represent different relation types among entity spans in a relation beam.
- Model architecture: The model enumerates spans and represents each using BiLSTM endpoint outputs, an attention-based headword, and an embedded span-width feature.Token inputs combine character representations, GloVe embeddings, and ELMo embeddings before the bidirectional LSTM.
- Dynamic span graph: Graph propagation refines span representations by aggregating neighboring information through coreference links and relation-type scores across repeated iterations.Coreference propagation shares context across likely antecedents; relation propagation uses a ReLU function to suppress unlikely relations and a trainable projection for relation information.
- Prediction and refinement: Gated updates decide whether each span retains its current representation or integrates new information from coreference and relation update vectors.The refined representations feed entity, relation, and coreference prediction layers, with entity and relation scores normalized over their label spaces.
- Training: Training minimizes a weighted sum of log-likelihoods for entity types, relations, and coreferences, with task weights tuned as hyperparameters.The implementation also tunes propagation iterations and beam size separately for each dataset using development data.
4 Experiments
The experiments evaluate DYGIE across joint entity-relation extraction and overlapping entity extraction in multiple datasets and domains. DYGIE reports substantial gains over prior systems, including settings without gold entity information.
- Experimental setup: DYGIE is evaluated on joint entity-relation extraction across ACE04, ACE05, SciERC, and the Wet Lab Protocol Corpus.The framework uses relation propagation for all datasets and coreference propagation where coreference annotations are available.
- Joint entity and relation extraction: 7.1% and 7.0% relative NER improvements over state of the art are reported on ACE04 and ACE05, respectively.For relation extraction, the corresponding relative improvements are 25.8% on ACE04 and 13.7% on ACE05.
- Joint entity and relation extraction: 5.9% and 1.9% advances over state of the art are reported on SciERC for relation extraction and NER, respectively.The authors associate the improvement with coreference and relation propagation constructing rich contextualized representations.
- Joint entity and relation extraction: 16.8% and 2.2% improvements over WLPC baselines are reported for relation extraction and NER without using gold entity information.The referenced WLPC baseline performs relation extraction using gold entity boundaries as input.
- Overlapping entity extraction: 11.6%, 11.3%, and 1.5% improvements over state of the art are reported on ACE04-O, ACE05-O, and GENIA, respectively.These overlapping-entity experiments use coreference propagation but not relation propagation because relation annotations are unavailable.
5 Analysis of Graph Propagation
Graph propagation improves different extraction tasks in different ways: coreference propagation mainly helps entity extraction, while relation propagation benefits both entities and relations. The strongest coreference gains occur for difficult pronouns and common entity confusions, while both propagation mechanisms perform best after two iterations.
- Coreference and Relation Graph Layers: Coreference propagation mainly helps entity extraction, whereas relation propagation significantly benefits both entity and relation extraction across ACE05 and SciERC.The ablation analysis reports that coreference propagation can hurt relation extraction on ACE05 but gives a small benefit on both SciERC tasks.
- Coreference Propagation and Entities: 6.6% improvement on pronoun performance is achieved with coreference propagation in ACE05.The result concerns hard-to-disambiguate pronouns requiring cross-sentence context.
- Coreference Propagation and Entities: Coreference propagation greatly reduces frequent GPE/PER and PER/ORG confusions, while improving most entity categories.GPE denotes a geopolitical entity, and the analysis attributes the frequent confusions to pronoun-associated errors.
- Coreference Propagation and Entities: Coreference propagation can introduce entity errors when the most confident antecedent is not the pronoun’s true antecedent.In one example, the model linked “They” to companies rather than earlier fund managers, producing an ORG instead of PER prediction.
- Coreference and Relation Graph Layers: The coreference and relation layers each achieve their best performance at the second iteration, with N = 2 and M = 2, respectively.The iteration analyses are reported for entity extraction with coreference propagation and relation extraction with relation propagation.
- Coreference and Relation Graph Layers: Relation propagation yields significant improvements for sentences containing more entities on ACE05.The analysis compares DYGIE with and without relation propagation and connects broader context with performance in entity-dense sentences.
6 Conclusion
DYGIE achieves state-of-the-art results across diverse information extraction domains through dynamic span graphs that improve interaction across tasks. The framework avoids external syntactic preprocessing, adds small inference costs, and supports future extension to structural tasks such as event extraction.
- Conclusion: DYGIE achieves state-of-the-art results on entity recognition and relation extraction across diverse domains.The conclusion identifies dynamic span graphs as the key contribution enabling richer interaction across tasks.
- Conclusion: Dynamic graph propagation enhances interaction across tasks by incorporating broader contextual information into span representations.The approach is described as learning useful information from entity, relation, and coreference contexts.
- Conclusion: DYGIE requires no preprocessing with syntactic tools and improves entity, relation, and overlapping entity extraction.The conclusion contrasts this design with IE frameworks that rely on external syntactic preprocessing.
- Conclusion: Cross-sentence coreference and relation propagation add only a small inference computation cost, while beam search controls memory cost.The added costs are described as small relative to the baseline span-based model.
- Conclusion: Future work includes extending the framework to structural information extraction tasks such as event extraction.