Applying optimization-based meta-learning, especially Reptile, to GLUE tasks yields small but consistent improvements over strong baselines on low-resource NLU tasks, with better transfer to a held-out task.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2019 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Investigating Meta-Learning Algorithms for Low-Resource Natural Language Understanding Tasks
Applying optimization-based meta-learning, especially Reptile, to GLUE tasks yields small but consistent improvements over strong baselines on low-resource NLU tasks, with better transfer to a held-out task.