#Paper Reading# Dual Learning for Machine Translation
論文題目:Dual Learning for Machine Translation
論文地址:https://arxiv.org/abs/1611.00179
論文發表於:NIPS 2016(B類會議)
論文大體內容:
NMT(neural machine translation)機器翻譯模型所需標註的訓練資料量特別大的問題,但大量的人工標註的訓練資料往往意味著大量的花費,因此本文針對這一問題,提出一個對偶模型,dual-NMT,能夠使用unlabel的資料也能達到一個很好的效果。
1、主要思想體現在two-agent communication game中,過程是這樣的:
(圖片來自於MSRA[1])
①需要有的輸入:資料集DA,DB;弱翻譯器ΘAB與ΘBA;強語言模型LMA與LMB;超引數α,K, γ1,t,γ2,t;
②對於Alice(熟悉English)來說,先從DA選出一個English的句子X,然後經過弱翻譯器ΘAB將X翻譯為B語言(French),得到Xmid;
③對於Bob(熟悉French)來說,看到Xmid,然後使用強語言模型LMB檢查Xmid的好壞,從而得到對弱翻譯器ΘAB的反饋;
④Bob再將Xmid使用弱翻譯器ΘBA翻譯為A語言(English),得到X’;
⑤Alice使用強語言模型LMA檢查X’以及對比X與X’的差距,從而得到對弱翻譯器ΘBA的反饋;
⑥接著對DB的句子也執行②-⑤操作,Alice和Bob交替玩這個game,從而不斷修正弱翻譯器ΘAB與ΘBA,得到強翻譯器;
2、實驗部分
①預處理:將包含非常用的30K個詞的句子去掉,每個詞用620維向量表示;
②評測方法:BLEU[2];
③baseline:傳統的NMT,pseudo-NMT;
3、最終效果
比baseline有較為明顯的提高;
4、思考
作者提出的這種對偶學習方法,確實能夠很好地克服label資料的不足。同時,作者也提到,只要能形成一個閉環系統,那麼就可以使用這種對偶學習的方法,畢竟,每個人(結點)都能夠判斷模型生成的效果,同時加入到NMT的反饋中,從而不斷把弱翻譯器改進為強翻譯器。這種方法其實類似於迭代式學習方法,通過反饋改進模型,還是挺有意思的。
參考資料:
[1]、http://www.msra.cn/zh-cn/news/blogs/2016/12/dual-learning-20161207.aspx
[2]、https://en.wikipedia.org/wiki/BLEU
以上均為個人見解,因本人水平有限,如發現有所錯漏,敬請指出,謝謝!
相關文章
- 【論文筆記】Neural machine translation by jointly learning to align and translate筆記Mac
- 《machine learning》引言Mac
- Machine Learning with SklearnMac
- Perceptron, Support Vector Machine and Dual Optimization Problem (3)Mac
- Perceptron, Support Vector Machine and Dual Optimization Problem (1)Mac
- Perceptron, Support Vector Machine and Dual Optimization Problem (2)Mac
- Machine Learning (12) - Support Vector Machine (SVM)Mac
- Machine Learning-IntroductionMac
- Machine Learning - Basic pointsMac
- Machine Learning (1) - Linear RegressionMac
- Extreme Learning Machine 翻譯REMMac
- pages bookmarks for machine learning domainMacAI
- Machine Learning(13)- Random ForestMacrandomREST
- Machine Learning (10) - Decision TreeMac
- Machine learning terms_01Mac
- Paper Reading: Random Balance ensembles for multiclass imbalance learningrandom
- Machine Learning (5) - Training and Testing DataMacAI
- SciTech-BigDataAIML-Machine Learning TutorialsAIMac
- [Paper Reading] VQ-VAE: Neural Discrete Representation Learning
- Paper Reading:A Survey of Deep Learning-based Object DetectionObject
- 《深度學習》PDF Deep Learning: Adaptive Computation and Machine Learning series深度學習APTMac
- Machine Learning Yearning 要點筆記Mac筆記
- Machine Learning(14) - K Fold Cross ValidationMacROS
- Machine Learning (6) - Logistic Regression (Binary Classification)Mac
- Machine Learning (8) - Logistic Regression (Multiclass Classification)Mac
- MATH38161 Multivariate Statistics and Machine LearningMac
- MPHY0041 Machine Learning in Medical ImagingMac
- Machine Learning(機器學習)之二Mac機器學習
- Machine Learning(機器學習)之一Mac機器學習
- 使用Octave來學習Machine Learning(二)Mac
- Machine Learning 機器學習筆記Mac機器學習筆記
- Machine Learning With Go 第4章:迴歸MacGo
- Monetizing Machine Learning.pdf 免費下載Mac
- machine learning model(algorithm model) .vs. statistical modelMacGo
- Matlab機器學習3(Machine Learning Onramp)Matlab機器學習Mac
- 論文閱讀:《Learning by abstraction: The neural state machine》Mac
- Paper Reading: Imbalanced ensemble learning leveraging a novel data-level diversity metric
- Proj. CDeepFuzz Paper Reading: Checker Bug Detection and Repair in Deep Learning LibrariesAI
- Coursera 吳恩達《Machine Learning》視訊 + 作業吳恩達Mac