Introduction
Information Extraction (IE) is a fundamental task in natural language processing, aimed at transforming unstructured text into structured representations. Modern Joint Information Extraction (JIE) models have advanced significantly by simultaneously addressing four sub-tasks: entity extraction, event trigger identification, relation extraction, and event argument extraction. By modeling these tasks jointly, systems can leverage inter-task dependencies to improve overall accuracy.
However, a significant challenge remains: domain shift. While JIE models perform exceptionally well when the test data matches the training distribution (e.g., news articles), their performance often degrades sharply when applied to out-of-domain data, such as social media or technical reports. This degradation occurs because vocabulary, linguistic style, and event descriptions vary across domains. Traditional Unsupervised Domain Adaptation (UDA) methods, which often focus on single-task alignment, are frequently insufficient for the complex, multi-faceted nature of JIE.
The DA4JIE Framework
To address these challenges, we present DA4JIE, a novel framework designed for unsupervised domain adaptation in joint information extraction. DA4JIE introduces two key components: Instance-relational Domain Adaptation (IrDA) and Context-invariant Structure Learning (CiSL).
Instance-relational Domain Adaptation (IrDA)
Standard UDA techniques typically attempt to align global feature distributions between source and target domains. In contrast, IrDA recognizes that different IE tasks have unique relational dependencies. Instead of a "fully connected" alignment that treats all tasks uniformly, IrDA utilizes a "type-relational graph" with a chain connection strategy. This approach aligns specific instance types (e.g., source entities to target entities) while respecting the natural hierarchy and dependencies between entities, triggers, and relations. This flexible alignment allows the model to bridge the domain gap without forcing overly restrictive constraints on task-specific representations.
Context-invariant Structure Learning (CiSL)
To improve transferability, a model should focus on structural linguistic features that remain stable across domains rather than relying on domain-specific vocabulary. The CiSL module employs Graph Transformer Networks (GTN) to integrate syntactic dependency trees with attention-based graphs. This process generates a "Context-invariant Graph" (CiG) that captures the underlying structural skeleton of a sentence. By conditioning the pooling layer on this CiG, the model can filter out domain-specific noise and focus on universal linguistic patterns.
Experimental Setup
The framework was evaluated using the ACE-05 dataset, a benchmark for joint information extraction. The dataset encompasses six distinct domains. In our experiments, a combination of Broadcast News (bn) and Newswire (nw) served as the labeled source data. The model was then tested on four unlabeled target domains: Broadcast Conversations (bc), Telephone Conversations (cts), Weblogs (wl), and Usenet Newsgroups (un).
Results and Analysis
DA4JIE demonstrates superior performance across all out-of-domain test sets compared to strong baselines, including standard BERT-based models and specialized JIE architectures like OneIE and FourIE.
As shown in the results table, DA4JIE achieved an average out-of-domain (aDom) score of 64.0 F1, representing a 2-point improvement over the BERT baseline. The model showed particularly notable gains in entity and relation extraction, where structural context is paramount. Ablation studies further confirmed that both the IrDA and CiSL modules are critical; removing either component resulted in a significant performance drop, highlighting the synergy between relational alignment and structural learning.
Conclusion
DA4JIE offers a robust solution for the domain shift problem in joint information extraction. By moving away from blunt, uniform alignment and focusing on task-specific relations and context-invariant structures, the framework achieves state-of-the-art results in unsupervised domain adaptation. Future research will explore automating the design of the relational graph and extending the framework to multi-source domain adaptation scenarios.
Summary updated