SelfORE: Self-supervised Relational Feature Learning for Open Relation Extraction

More: A Metric Learning Based Framework for Open-Domain Relation Extraction

“…We compare with four baselines [10,8,22,7] on two datasets. All these models are evaluated on the test set to show their Method FewRel Prec.…”

“…(1) Benefit by the rich supervision signals come from the labeled data, MORE outperforms all unsupervised or selfsupervised methods on both datasets, such as SelfORE [8] which used to achieve admirable results on NYT+FB. The results indicate the effectiveness of prior knowledge transfer, which is conducive to novel type-detection in open scenarios.…”

“…BERT Inspired by SelfORE [8], which exploits the pre-trained language model, we also choose BERT [15] as our encoder and follow the operation proposed by [16] to fit our OpenRE task better. Specifically, for a sentence S = {s 1 , .., s T }, where s indicates the token and T is the length of S. We insert four special tokens before and after each entity mentioned in a sentence and get a new sequence:…”

“…Evaluation Metrics We adopt B 3 F 1 score [21] as our metrics, which is widely used in previous works [10,8,7,22]. F 1 calculates the harmonic mean of precision and recall, its value is more affected by the lower one, which can fairly demonstrate the performance of the model.…”

“…Recently, neural networks began to be exploited in clustering-based OpenRE tasks to alleviate the above issues. For example, Hu et al [8] utilize a neural model to capture self-supervision signals and detect novel instances in an open scene. Gao et al [9] propose a siamese network, which accumulates novel types with its few-shot instances.…”