Upgrade to Pro — share decks privately, control downloads, hide ads and more …

文献紹介:Improving Back-Translation with Uncertainty-based Confidence Estimation

Taichi Aida
November 27, 2019

文献紹介:Improving Back-Translation with Uncertainty-based Confidence Estimation

Improving Back-Translation with Uncertainty-based Confidence Estimation
Shuo Wang, Yang Liu, Chao Wang, Huanbo Luan, Maosong Sun
Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 791-802, Hong Kong, 2019.

Taichi Aida

November 27, 2019
Tweet

More Decks by Taichi Aida

Other Decks in Technology

Transcript

  1. Improving Back-Translation with Uncertainty-based Confidence Estimation Shuo Wang, Yang Liu,

    Chao Wang, Huanbo Luan, Maosong Sun Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 791-802, Hong Kong, 2019.
  2. Introduction - Back-Translation は low-resource な機械翻訳 で効果的 - 限られたデータで訓練された逆方向 NMT

    で生成 した擬似データはノイズが避けられない - “モデルの不確かさ” を用いて改善を図る 3
  3. Method 9 - 確率分布→不確かさ? - Predicted translation probability (PTP) -

    Expected translation probability (EXP) - Variance of translation probability (VAR) - Combination of expectation and variance (CEV)
  4. Method - Variance of translation probability (VAR) - 分散 -

    1から引いたものを採用(α = 2) 11
  5. Experiment 1. 不確かさの比較(PTP vs EXP vs VAR vs CEV) 2.

    粒度間で不確かさの比較(word vs sentence) 3. ベースラインとの比較 13
  6. Experiment - Model: Transformer - Evaluation: BLEU - Data 14

    train dev test back-translation Chinese-English LDC(1.25M) NIST06 NIST02-05, 08 WMT17 (English, 10M) English-German WMT14 (4.47M) newstest13 newstest12, 14-15 NewsCrawl12 (German, 4.5M)
  7. Result 1. 不確かさの比較(PTP vs EXP vs VAR vs CEV) 15

    CEV(期待値と分 散の組み合わせ) が最も効果あり 今後CEVを使用
  8. Reference - Dropout as a Bayesian Approximation: Representing Model Uncertainty

    in Deep Learning(ICML2016) https://arxiv.org/abs/1506.02142 - ベイジアン・ディープラーニングによる安全なAIの実現 https://qiita.com/takaaki5564/items/5ed89541d8d2a4725baa# 20