Leveraging Open Information Extraction for More Robust Domain Transfer of Event Trigger Detection

TL;DR

This paper tackles negative domain transfer in event trigger detection by leveraging open information extraction (OIE) relations as cross-domain mediators. It introduces implicit and explicit multi-task architectures that couple TD with OIE relation extraction, and explores sequential and in-domain transfer regimes, augmented by a target-domain masked language modeling objective. Empirical results show that OIE mediation substantially improves zero- and few-shot TD transfer from Wikipedia to news domains, with even larger gains when MLM is incorporated, and the improvements prove robust to the choice of OIE system. The work suggests a practical path toward universal, domain-robust event extraction by bridging trigger semantics and predicate-argument structures learned through OIE.

Abstract

Event detection is a crucial information extraction task in many domains, such as Wikipedia or news. The task typically relies on trigger detection (TD) -- identifying token spans in the text that evoke specific events. While the notion of triggers should ideally be universal across domains, domain transfer for TD from high- to low-resource domains results in significant performance drops. We address the problem of negative transfer in TD by coupling triggers between domains using subject-object relations obtained from a rule-based open information extraction (OIE) system. We demonstrate that OIE relations injected through multi-task training can act as mediators between triggers in different domains, enhancing zero- and few-shot TD domain transfer and reducing performance drops, in particular when transferring from a high-resource source domain (Wikipedia) to a low(er)-resource target domain (news). Additionally, we combine this improved transfer with masked language modeling on the target domain, observing further TD transfer gains. Finally, we demonstrate that the gains are robust to the choice of the OIE system.

Figures (3)

• Figure 1: An example of event trigger detection and subject-relation-object extraction with an open information extraction (OIE) system. The detected trigger and extracted OIE relation often overlap to a significant degree, which can be leveraged for creating more robust trigger detection models across domains.
• Figure 2: Implicit model during training. The input sentence is fed twice: once with trigger IOB2 tags through PLM encoders and once with OIE relation IOB2 tags by indexing the corresponding label embedding matrix. At the implicit output, PLM's last hidden state embeddings are concatenated with OIE relation label embeddings per token and passed through the TD softmax classifier.
• Figure 3: TD domain transfer micro F1 scores when transferring from MAVEN as a source to ACE 2005, EDNYT, and EVEXTRA as targets (zero-shot, in-domain training, and sequential transfer, with six varying numbers of shots). The numbers in parentheses next to the target dataset are the in-domain performance test set scores when using all target training data. The upper three plots show in-domain training results -- target fine-tuning starting from PLM. The lower three plots show sequential transfer results -- target fine-tuning starting from PLM trained for TD on MAVEN source training data. Dash-dotted lines correspond to models with an auxiliary MLM objective on target domain training data. The x-axis shows the number of shots on an ordinal scale. Implicit and explicit models leverage MinIE relation labels, unlike the vanilla model. All reported results are averages of three runs. The corresponding results in tabular form with standard deviations are in Appendix \ref{['subsec:add_results']}.