Search

Document-level relation extraction faces two often overlooked challenges: long-tail problem and multi-label problem. Previous work focuses mainly on obtaining better contextual representations for entity pairs, hardly address the above challenges. In this paper, we analyze the co-occurrence correlation of relations, and introduce it into the document-level relation extraction task for the first time. We argue that the correlations can not only transfer knowledge between data-rich relations and data-scarce ones to assist in the training of long-tailed relations, but also reflect semantic distance guiding the classifier to identify semantically close relations for multi-label entity pairs. Specifically, we use relation embedding as a medium, and propose two co-occurrence prediction sub-tasks from both coarse- and fine-grained perspectives to capture relation correlations. Finally, the learned correlation-aware embeddings are used to guide the extraction of relational facts. Substantial experiments on two popular datasets (i.e., DocRED and DWIE) are conducted, and our method achieves superior results compared to baselines. Insightful analysis also demonstrates the potential of relation correlations to address the above challenges. The data and code are released at https://github.com/RidongHan/DocRE-Co-Occur. [ABSTRACT FROM AUTHOR]

Wrong-labeling problem and long-tail relations severely affect the performance of distantly supervised relation extraction task. Many studies mitigate the effect of wrong-labeling through selective attention mechanism and handle long-tail relations by introducing relation hierarchies to share knowledge. However, almost all existing studies ignore the fact that, in a sentence, the appearance order of two entities contributes to the understanding of its semantics. Furthermore, they only utilize each relation level of relation hierarchies separately, but do not exploit the heuristic effect between relation levels, i.e., higher-level relations can give useful information to the lower ones. Based on the above, in this paper, we design a novel Recursive Hierarchy-Interactive Attention network (RHIA) to further handle long-tail relations, which models the heuristic effect between relation levels. From the top down, it passes relation-related information layer by layer, which is the most significant difference from existing models, and generates relation-augmented sentence representations for each relation level in a recursive structure. Besides, we introduce a newfangled training objective, called Entity-Order Perception (EOP), to make the sentence encoder retain more entity appearance information. Substantial experiments on the popular New York Times (NYT) dataset are conducted. Compared to prior baselines, our RHIA-EOP achieves state-of-the-art performance in terms of precision–recall (P–R) curves, AUC, Top-N precision and other evaluation metrics. Insightful analysis also demonstrates the necessity and effectiveness of each component of RHIA-EOP. [ABSTRACT FROM AUTHOR]

To find relational facts of interest from plain texts, distantly supervised relation extraction (DSRE) has drawn significant attention. Recent works exploit relation hierarchies to mine more clues for long-tail relations and achieve good performance. However, they ignore or underutilize the correlation of relations in the hierarchical structure. Empirically, the correlation facilitates knowledge transfer between different relations to further handle long-tail relations and improves inter-relational discrimination. In this paper, we devise an end-to-end network to model the correlation of relations from two perspectives. Globally, we construct an undirected connected graph according to the relation hierarchies, and employ Graph Attention Networks (GATs) to aggregate node information and generate correlation-aware Global Hierarchy Embeddings (GHE). Locally, we assume that along the relation hierarchies, the classification results of adjacent levels should be highly interdependent , and introduce a constraint called Local Probability Constraints (LPC) to take it into account. LPC is then combined with a branch network for both sentence-level and bag-level classification. Experimental results on the popular New York Times (NYT) dataset show that, our model GHE-LPC outperforms other state-of-the-art baselines in terms of AUC, Top-N precision, accuracy of Hits@K, etc. [ABSTRACT FROM AUTHOR]