3つの要点 ✔️その1 高速・高精度な自然言語処理モデルELECTRAが登場 ✔️その2 低精度なGeneratorにより入力を置換することで、文全体から効率的に学習を行う ✔️その3 RoBERTaの約1/4の学習量で同等の性能を発揮 ELECTRA: PRE-TRAINING TEXT ENCODERS AS DISCRIMINATORS RATHER THAN GENERATORS written by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning (26 Sep 2019 (modified: 10 Mar 2020)) Comments: accepted by ICLR 2020 Subjects: Machine Learning (cs.LG); Machine Learning (sta