Extracting relation triples from unseen data

Abstract

Developing a relational extraction model from unstructured text is essential for the automation of large-scale knowledge graph maintenance. To maintain the knowledge graph up-to-date, it is required for a model to extract relational triples from sentences that might contain unseen entities. Simply fine-tuning BERT on a relational triple extraction task shows excellent performance on seen entities (entities in the training data), although it does not generalize well for new entities. We find that augmentation with noisy data helps to extract relational triples between unseen entities, while it comes at a cost of performance degradation on seen entities. Since we have two experts, one on seen entities and the other one on unseen entities, we filter predictions of each experts and union them to get the best of both experts. Experiments on two standard benchmark datasets, NYT and WebNLG, show that our model outperforms current state-of-the-art model on unseen data, along with competitive results on the original seen data.