Source-linked AI summary
Jointly Multiple Events Extraction via Attention-based Graph Information Aggregation
Xiao Liu, Zhunchen Luo, Heyan Huang
TL;DR
JMEE targets the difficulty of extracting multiple associated events in the same sentence, where sequential modeling is inefficient at capturing long-range dependencies. It introduces syntactic shortcut arcs, GCNs, and self-attention for joint trigger-and-argument extraction, and reports state-of-the-art performance on widely used event-extraction datasets.
Problem
Multiple events commonly occur in one sentence, making extraction harder because event types are often associated and sequential models inefficiently capture long-range dependencies.
Method
JMEE introduces syntactic shortcut arcs, graph convolutional networks, and self-attention to jointly extract event triggers and arguments while modeling associations between events.
Results
JMEE achieves state-of-the-art performance on widely used event-extraction datasets.
Takeaways & Limitations
The framework provides a joint event-extraction approach that enhances information flow and aggregates associations between multiple events.
Takeaways & Limitations
The approach extracts events mainly at the sentence level, relying on similar document-level and sentence-level event co-occurrence distributions in ACE 2005.
Abstract
from arXiv · showhide
Event extraction is of practical utility in natural language processing. In the real world, it is a common phenomenon that multiple events existing in the same sentence, where extracting them are more difficult than extracting a single event. Previous works on modeling the associations between events by sequential modeling methods suffer a lot from the low efficiency in capturing very long-range dependencies. In this paper, we propose a novel Jointly Multiple Events Extraction (JMEE) framework to jointly extract multiple event triggers and arguments by introducing syntactic shortcut arcs to enhance information flow and attention-based graph convolution networks to model graph information. The experiment results demonstrate that our proposed framework achieves competitive results compared with state-of-the-art methods.
1 Introduction
JMEE addresses the difficulty of extracting associated multiple events in one sentence by combining syntactic shortcut arcs, graph convolution, and self-attention for joint trigger and argument extraction.
- Task and motivation: Event extraction identifies and classifies triggers, then identifies arguments and labels their roles.The paper frames these as event detection and argument extraction under ACE 2005.
- Task and motivation: Multiple events in one sentence are harder to extract because event types and triggers are often associated.Considering a later trigger can disambiguate an earlier trigger’s event type.
- Prior approaches: Prior sentence-level sequential models use contextual features and memory mechanisms but are inefficient at capturing long-range dependencies.The paper motivates syntactic shortcut arcs as an alternative information path.
- Syntactic shortcuts: Dependency-based shortcut paths reduce the hops needed for information to travel between event triggers.In Figure 1, the path from killed to barrage requires three dependency-tree hops instead of six sequential hops.
- Proposed framework: JMEE uses GCNs to model syntactic graph information and self-attention to preserve associations while jointly extracting triggers and arguments.The framework introduces syntactic shortcut arcs to enhance information flow.
- Evaluation and contributions: The framework is evaluated on ACE 2005 and is reported to achieve state-of-the-art performance on widely used event-extraction datasets.The paper also releases implementation code and notes that shortcut paths are no worse than sequential modeling in the worst tree-path case.
2 Approach
JMEE jointly extracts triggers and arguments by combining sentence-level modeling with syntactic graph structure and self-attention to capture associations among multiple events.
- The joint approach extracts event triggers and arguments simultaneously, avoiding propagated errors from a trigger-then-argument pipeline.
- 2 Approach: The framework uses four modules: word representation, syntactic graph convolution, self-attention trigger classification, and argument classification.
- 2.1 Word Representation: JMEE represents tokens with concatenated word, POS-tag, positional, and entity-type embeddings before later event-extraction modules.The resulting sequence X is fed into subsequent modules for representation learning.
- 2.2 Syntactic Graph Convolution Network: Syntactic parsing adds reversed edges, self-loops, and shortcut arcs so graph convolutions aggregate information from structurally related tokens.Edge types are reduced to control parameter count, and gated edges weight their individual importance.
- 2.2 Syntactic Graph Convolution Network: The graph convolution network initializes node states from token representations and aggregates transformed neighbor representations according to edge types.The neighborhood includes each node itself through self-loops.
- 2.2 Syntactic Graph Convolution Network: Bi-LSTM encoding complements GCNs by supplying local sequential context, while highway units allow information to flow across stacked GCN layers.The design targets dependencies along shortcut arcs without excessive information over-propagation.
3 Experiments
Experiments evaluate JMEE on ACE 2005 against established systems and examine its performance on single- and multiple-event sentences. JMEE achieves the strongest reported overall and multiple-event results, while attention visualizations illustrate cross-event interactions.
- Experimental Setup: JMEE is evaluated on ACE 2005 using the same data split and compared with six state-of-the-art methods.The comparison includes Cross-Event, JointBeam, DMCNN, PSL, JRNN, and dbRNN.
- Overall Performance: JMEE achieves the best F1 scores for trigger classification and argument-related subtasks with golden-standard entities.The reported gains are 2% over the best-reported models for trigger classification and argument role labeling.
- Effect on Extracting Multiple Events: The multiple-event evaluation separates test sentences into 1/1 single-event and 1/N multiple-event cases.A 1/1 sentence has one trigger or one argument role, whereas 1/N contains multiple events or roles.
- Effect on Extracting Multiple Events: 7.9% better trigger-classification performance than JRNN is reported for JMEE on the 1/N split.The authors attribute this result to leveraging syntactic shortcut arcs and self-attention aggregation.
- Self-Attention Analysis: Attention scores show strong connections between each event trigger, itself, the other event trigger, and relevant arguments.In the example sentence, arrested connects strongly with four people and killings, while killings also connects with arrested.
4 Related Work
Related work models event associations with feature-based, document-level, RNN, and CNN approaches. The paper identifies inefficient long-range dependency capture and incomplete event-association modeling as limitations of sentence-level sequential methods.
- Existing Event-Association Methods: Prior approaches exploit sentence-level features, document-level information, and explicit associations between triggers and arguments.Examples include ranking dependencies, combinational features, probabilistic soft logic, relational features, and document-level co-occurrence information.
- Deep Learning Methods: Deep learning approaches use RNNs and CNNs to capture associations between events.The cited methods include JRNN, dbRNN, DMCNN, and other recurrent or convolutional systems.
- Limitations of Prior Work: Sentence-level sequential modeling methods capture very long-range dependencies inefficiently and do not fully model associations between events.This limitation motivates graph-based information flow and attention-based aggregation in JMEE.
5 Conclusion and Future Work
JMEE addresses multiple-event extraction with syntactic shortcut arcs, graph convolution, self-attention, and joint trigger-argument prediction. The authors report effective performance and identify cross-role argument information as future work.
- Conclusion: JMEE introduces syntactic shortcut arcs and graph convolution to enhance information flow and capture syntactic representations.The framework then applies self-attention aggregation to model associations between events.
- Conclusion: JMEE jointly extracts event triggers and arguments using a biased loss function for dataset class imbalance.The conclusion presents this joint framework as effective for alleviating the multiple-event phenomenon.
- Future Work: Future work will use information from an argument that plays different roles in various events.