MaskGAN Benchmark Task

A MaskGAN Benchmark Task is an Automatic Text Generation Task that uses Generative Adversarial Network and Masked Sequences to generate text items.

• AKA: MarkGAN Task.
• Context:
  • Task Input(s): word sequences.
  • Task Output(s): automatic generated sentences.
  • Task Requirement(s):
    • Benchmark Datasets:
      • Real Datasets:
        • Penn Treebank (PTB) Datasets - a vocabulary of 10,000 unique words. Training set contains 930,000 words; validation set contains 74,000 words; and the test set contains 82,000 words.
        • IMDB Movie Datasets - a 100,000 movie reviews datasets: 25,000 labeled training instances; 25,000 labeled test instances and 50,000 unlabeled training instances.
      • Synthetic Datasets:
        • Conditional Samples generated on the PTB datasets.
        • Unconditional Samples generated on the PTB datasets.
        • Conditional Samples generated on the IMDB datasets.
        • Unconditional Samples generated on the IMDB datasets.
    • Benchmark Performance Metrics:
      • Validation Perplexity,
      • Mode Collapse - percentage of unique n-grams (bi-, tri- and quad-grams) in a set of 10 000 generated IMDB movie reviews.
      • Human Evaluation - an Amazon Mechanical Turk blind heads-up comparison between pairs of baseline models trained on IMDB reviews.
    • Baseline Models:
      • MaskGAN Model;
      • MaskMLE Model;
    • other text generation competing systems:
      • a Language Model;
      • SeqGAN Model,
  • It can be solved by a MaskGAN Training System that implements MaskGAN Algorithms.
• Example(s):
  • Validation perplexity (Tab.5 in Fedus et al., 2018):

Model	Perplexity of IMDB samples under a pretrained LM
MaskMLE	$273.1 \pm 3.5$
MaskGAN	$108.3 \pm 3.5$

• Mode Collapse (Tab.6 in Fedus et al., 2018):

Table 6: Diversity statistics within 1000 unconditional samples of PTB news snippets (20 words each).

Model	Unique bigrams	% Unique trigrams %	Unique quadgrams
LM	40.6	75.2	91.9
MaskMLE	43.6	77.4	92.6
MaskGAN	38.2	70.7	88.2

• Human evaluation of the baseline models trained on IMDB datasets (Tab.7 in Fedus et al., 2018):

Preferred Model	Grammaticality %	Topicality %	Overall %
LM	15.3	19.7	15.7
MaskGAN	59.7	58.3	58.0
LM	20.0	28.3	21.7
MaskMLE	42.7	43.7	40.3
MaskGAN	49.7	43.7	44.3
MaskMLE	18.7	20.3	18.3
Real samples	78.3	72.0	73.3
LM	6.7	7.0	6.3
Real samples	65.7	59.3	62.3
MaskGAN	18.0	20.0	16.7

• Human evaluation of the baseline models trained on PTB datasets (Tab.8 in Fedus et al., 2018):

Preferred Model	Grammaticality %	Topicality %	Overall %
LM	32.0	30.7	27.3
MaskGAN	41.0	39.0	35.3
LM	32.7	34.7	32.0
MaskMLE	37.3	33.3	31.3
MaskGAN	44.7	33.3	35.0
MaskMLE	28.0	28.3	26.3
SeqGAN	38.7	34.0	30.7
MaskMLE	33.3	28.3	27.3
SeqGAN	31.7	34.7	32.0
MaskGAN	43.3	37.3	37.0

• Counter-Example(s):
  • a CJS Neural Narrative Text Generation Benchmark Task,
  • a GSGAN Benchmark Task,
  • a LeakGAN Benchmark Task,
  • a LSPGS Wikipedia Long Sentences Summarization Task,
  • MaliGAN Benchmark Task,
  • RankGAN Benchmark Task,
  • SeqGAN Benchmark Task,
  • TextGAN Benchmark Task,
  • Texygen Benchmark Task.
• See: Neural Text Generation System, Seq2Seq Model, Neural Autoregressice Model, Professor Forcing Algorithm, Scheduled Sampling Algorithm.

---

---

References

2018

• (Fedus et al., 2018) ⇒ William Fedus, Ian Goodfellow, and Andrew M Dai. (2018). “MaskGAN: Better Text Generation via Filling in the ________". In: Proceedings of the Sixth International Conference on Learning Representations (ICLR-2018).

2015a

• (Kingma & Ba, 2015) ⇒ Diederik P. Kingma, and Jimmy Ba. (2015). “Adam: A Method for Stochastic Optimization.” In: Proceedings of the 3rd International Conference for Learning Representations (ICLR-2015).

2015b

• (Luong, Pham et al., 2015) ⇒ Minh-Thang Luong, Hieu Pham, and Christopher D. Manning. (2015). “Effective Approaches to Attention-based Neural Machine Translation". In: Proceedings of Conference on Empirical Methods in Natural Language Processing (EMNLP-2015). DOI:10.18653/v1/D15-1166.

1999

• (Sutton et al., 1999) ⇒ Richard S. Sutton, David A. McAllester, Satinder P. Singh, and Yishay Mansour (1999). "Policy Gradient Methods for Reinforcement Learning with Function Approximation". In: Advances in Neural Information Processing Systems 12 (NIPS Conference).

---