TY - GEN
T1 - Dynamic Prototype Selection by Fusing Attention Mechanism for Few-Shot Relation Classification
AU - Wu, Linfang
AU - Zhang, Hua Ping
AU - Yang, Yaofei
AU - Liu, Xin
AU - Gao, Kai
N1 - Publisher Copyright:
© 2020, Springer Nature Switzerland AG.
PY - 2020
Y1 - 2020
N2 - In a relation classification task, few-shot learning is an effective method when the number of training instances decreases. The prototypical network is a few-shot classification model that generates a point to represent each class, and this point is called a prototype. The mean is used to select prototypes for each class from a support set in a prototypical network. This method is fixed and static, and will lose some information at the sentence level. Therefore, we treat the mean selection as a special attention mechanism, then we expand the mean selection to dynamic prototype selection by fusing a self-attention mechanism. We also propose a query-attention mechanism to more accurately select prototypes. Experimental results on the FewRel dataset show that our model achieves significant and consistent improvements to baselines on few-shot relation classification.
AB - In a relation classification task, few-shot learning is an effective method when the number of training instances decreases. The prototypical network is a few-shot classification model that generates a point to represent each class, and this point is called a prototype. The mean is used to select prototypes for each class from a support set in a prototypical network. This method is fixed and static, and will lose some information at the sentence level. Therefore, we treat the mean selection as a special attention mechanism, then we expand the mean selection to dynamic prototype selection by fusing a self-attention mechanism. We also propose a query-attention mechanism to more accurately select prototypes. Experimental results on the FewRel dataset show that our model achieves significant and consistent improvements to baselines on few-shot relation classification.
KW - Attention mechanism
KW - Few-shot learning
KW - Relation classification
UR - http://www.scopus.com/inward/record.url?scp=85082306632&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-41964-6_37
DO - 10.1007/978-3-030-41964-6_37
M3 - Conference contribution
AN - SCOPUS:85082306632
SN - 9783030419639
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 431
EP - 441
BT - Intelligent Information and Database Systems - 12th Asian Conference, ACIIDS 2020, Proceedings
A2 - Nguyen, Ngoc Thanh
A2 - Trawinski, Bogdan
A2 - Jearanaitanakij, Kietikul
A2 - Chittayasothorn, Suphamit
A2 - Selamat, Ali
PB - Springer
T2 - 12th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2020
Y2 - 23 March 2020 through 26 March 2020
ER -