visualizing and understanding neural machine translation
play

Visualizing and Understanding Neural Machine Translation Yanzhuo - PowerPoint PPT Presentation

Visualizing and Understanding Neural Machine Translation Yanzhuo Ding, Yang Liu, Huanbo Luan, Maosong Sun Tsinghua University Presented by: Yuchen He Layer-wise relevance propagation (LRP) Can calculate the relevance between two arbitrary


  1. Visualizing and Understanding Neural Machine Translation Yanzhuo Ding, Yang Liu, Huanbo Luan, Maosong Sun Tsinghua University Presented by: Yuchen He

  2. Layer-wise relevance propagation (LRP) Can calculate the relevance between two arbitrary neurons Measures/visualizes how much each pixel is related to the final classification

  3. Goal ● To quantify and visualize the relevance between a neural network layer and contextual word vectors(source & target word embeddings) Offers more insights in interpreting how target words are generated

  4. Relevance vector

  5. Calculating Neuron-Level Relevance Base case: (relevance of v to itself) for any neuron v v Recursive case: (relevance of u to v ) u for any neurons u, v z v OUT( u ) comprises all u ’s directly connected descendant neurons in the network.

  6. Calculating Weight Ratios u for any neurons u, v v is the weight of u to v in the existing neural network IN( u ) comprises all u ’s directly connected ancestor neurons in the network.

  7. Putting things together Sum up and get vector-level relevance Generate and normalize relevance vector as a sequence of for all related contextual word vectors

  8. Relevance vector

  9. Application Help debug attention-based NMT systems ● Word omission ● Word repetition ● Unrelated words ● Negation reversion

  10. “Relevance matrix” attention weights source context vector target hidden state target word embedding I visit is to pray

  11. “Relevance matrix” attention weights source context vector target hidden state target word embedding I visit is to pray

  12. Word omission Input 巴基斯坦 总统 穆沙拉夫 赢 得参众两院信任投票 Reference Pakistani president Musharraf wins votes of confidence in senate and house Pakistani president win over democratic vote of Output confidence in senate (missing words)

  13. senate house confidence vote

  14. senate house confidence vote

  15. Word repetition Input 美国人 历 史上有 讲诚 信的 传统 ,有犯 错认错 的 传统 Reference In history, Americans have the tradition of honesty and would not hesitate to admit their mistakes In the history of the history of the history of the Output Americans, there is a tradition of faith in the history of mistakes

  16. Americans history have

  17. Americans history have

  18. Unrelated words Input 此次会 议 的一个重要 议题 是跨大西洋关系 Reference One of the top agendas of the meeting is to discuss the transatlantic relations Output A key topic of the meeting is to forge ahead

  19. is transatlantic relations

  20. is transatlantic relations

  21. N egation reversion Input 不解决生存 问题 ,就 谈 不上 发 展,更 谈 不上可持 续发 展 Without solving the issue of subsistence, there will be Reference no development to speak of , let alone sustainable development If we do not solve the problem of living , we will talk Output about development and still less can we talk about sustainable development

  22. talke not development

  23. talke not development

  24. Thank you

Recommend


More recommend