Table 1.
An example of 5-way-1-shot task setting of FSRC.
Fig 1.
The general framework of KEPN.
Fig 2.
The illustration of knowledge enhancement mechanism.
(a) Original prototype distribution, the query instance q closest to the prototype pa. (b) The distribution of the prototype has modified through prototype enhancement module, where r represent the relation information. (c) Updated prototype distribution, the query instance q is closest to the prototype pb, which is correctly classified now.
Fig 3.
Structure of the prototype enhancement module.
The rectangle, diamond, and triangle represent relationship information, basic prototype representation, and enhanced prototype representation, respectively.
Fig 4.
Illustration of class cluster loss.
The (a), (b), and (c) represent the sampling strategies of triplet loss, triplet loss with hard sample mining, and class cluster loss, respectively. The (d), (e), and (f) respectively represent the updated metric space under the corresponding sampling strategy.
Table 2.
Hyperparameter of the modes built in our experiments.
Table 3.
Experimental results of different models on the FewRel 1.0 validation and test set.
Table 4.
Experimental results on FewRel 2.0 domain adaptation test set.
Fig 5.
The performance on different tasks.
Table 5.
Paired-sample t-test on FewRel 1.0 validation set.
Table 6.
Ablation study on FewRel 1.0 validation set.
Fig 6.
Visualization of instance embeddings of a 5-way-30-shot task on the FewRel 1.0 validation set.
Table 7.
Case study on FewRel 1.0 validation set.
Fig 7.
Hyperparameter study on FewRel 1.0 validation set.