ERNIE-Gram: Pre-Training with Explicitly N-Gram Masked Language Modeling for Natural Language Understanding

Benchmark Model Rank Results
linguistic-acceptability-on-cola-glue-devERNIE-Gram#5Matthews Correlation: 68.6
natural-language-inference-on-multinli-devERNIE-Gram#8Matched: 89.1
natural-language-understanding-on-glue-devERNIE-Gram#8GLUE (dev): 88.1
natural-language-inference-on-rte-glue-devERNIE-Gram#9Accuracy: 83.8
question-answering-on-squad20-devERNIE-Gram#10F1: 87.1
reading-comprehension-on-raceERNIE-Gram#10Accuracy: 77.7
natural-language-inference-on-qnli-glue-devERNIE-Gram#11Accuracy: 93.2
question-answering-on-squad11-devERNIE-Gram#50F1: 93.2
sentiment-analysis-on-sst-2-binary-classificationERNIE-Gram#84Dev Accuracy: 95.6