Synthesizer: Rethinking Self-Attention in Transformer Models

Benchmark Model Rank Results
dialogue-generation-on-persona-chat-1Synthesizer (R+V)#6BLEU-1: 14.7CIDr: 19.09METEOR: 6.39ROUGE-L: 14.79
document-summarization-on-cnn-daily-mailSynthesizer (R+V)#21ROUGE-1: 38.57ROUGE-2: 16.24ROUGE-L: 35.95
linguistic-acceptability-on-cola-devSynthesizer (R+V)#5Accuracy: 53.3
machine-translation-on-wmt2014-english-frenchSynthesizer (Random + Vanilla)#20BLEU score: 41.85
machine-translation-on-wmt2014-english-germanSynthesizer (Random + Vanilla)#36BLEU score: 28.47