-
Free Generation
- [1]. GPT/GPT-2
- [2]. BERT for NLG
-
Conditional Generation
- [3]. mircosoft: MASS
- [4]. mircosoft: UNIFM
- [5]. AI2: GROVER
- [6].Pretraining for Conditional Generation with Pseudo Self Attention
-
Decoding Method
-
Autoregressive
- seq2seq, transformer
-
Non-AutoRegressive
- [8] Non-Autoregressive NMT
- [9]. InDIGO
- [10]. Insertion Transformer
- [11]. Levenshtein Transformer
- [12]. based on masked LM
-
RL+GAN
- [13, 14]. seqgan/ RAML
- [15, 16]. ScratchGAN/ COT
-
Metric
- F-score/accuracy
- BLEU for MT, ppl
- [7]. SJTU :Texygen
- [13]. train-test skewness, exposure bias
- Pre-training for NLG