Abstract
Few-shot relation classification describes a circumstance where a model is required to classify new-coming query instances after meeting only few support instances during testing. In this paper, we place a challenging restriction to conventional few-shot relation classification by additionally limiting the amount of training data. We also propose a few-shot learning framework for relation classification, which is particularly powerful when the training data is very small. In our framework, models not only strive to classify query instances, but also seek underlying knowledge within support instances to obtain better instance representations. The framework also includes a method for aggregating cross-domain knowledge into models by open-source task enrichment. Additionally, we construct a brand new dataset: the TinyRel-CM dataset, a few-shot relation classification dataset in health domain with limited training data and challenging relation classes. Experimental results demonstrate that our framework brings performance gains for most underlying classification models, and outperforms the state-of-the-art results given little training data (e.g., on TinyRel-CM dataset and FewRel-dataset [Han et al., 2018] with reduced training set), and achieves competitive results with sufficiently large training data (e.g., on FewRel dataset with full training data).
Abstract (translated)
URL
https://arxiv.org/abs/2004.14164
