Neural Snowball for Few-Shot Relation Learning

• DOI: 10.1609/aaai.v34i05.6281
• Publication Date: 2020-06-29T19:16:16Z
• Authors: Tianyu Gao; Xu Han; Ruobing Xie; Zhiyuan Liu; Fen Lin; Leyu Lin; Maosong Sun
• Keywords: Snowball sampling; Bootstrapping (finance); Relationship extraction

Knowledge graphs typically undergo open-ended growth of new relations. This cannot be well handled by relation extraction that focuses on pre-defined relations with sufficient training data. To address few-shot instances, we propose a novel bootstrapping approach, Neural Snowball, to learn transferring semantic knowledge about existing More specifically, use Relational Siamese Networks (RSN) the metric relational similarities between instances based and their labeled Afterwards, given its RSN accumulate reliable from unlabeled corpora; these are used train classifier, which can further identify facts relation. The process is conducted iteratively like snowball. Experiments show our model gather high-quality for better learning achieves significant improvement compared baselines. Codes datasets released https://github.com/thunlp/Neural-Snowball.