User Tools

Site Tools


paper:dual_learning_for_machine_translation

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Next revision
Previous revision
paper:dual_learning_for_machine_translation [2020/11/19 02:18] – Create paper summary Dual Learning For Machine Translation chrisliu298paper:dual_learning_for_machine_translation [2023/06/15 07:36] (current) – external edit 127.0.0.1
Line 5: Line 5:
 **TLDR;** The authors cast machine translation as a game of a primal task and a dual task, which allows two agents to teach each other in a reinforcement learning fashion (i.e., policy gradient method). With this implementation, the agent only requires monolingual English <-> French text data (with 10% bilingual warmup) to achieve a comparable performance to NMT trained with pure bilingual data. **TLDR;** The authors cast machine translation as a game of a primal task and a dual task, which allows two agents to teach each other in a reinforcement learning fashion (i.e., policy gradient method). With this implementation, the agent only requires monolingual English <-> French text data (with 10% bilingual warmup) to achieve a comparable performance to NMT trained with pure bilingual data.
  
-===== Key Points =====+==== Key Points ====
  
   * The motivation is that parallel data are scarce and costly to collect.   * The motivation is that parallel data are scarce and costly to collect.
Line 19: Line 19:
   * It maximizes towards the expected sum of the two rewards over the middle translation outputs.   * It maximizes towards the expected sum of the two rewards over the middle translation outputs.
   * It uses beam search to obtain more meaningful results.   * It uses beam search to obtain more meaningful results.
-</markdown> 
paper/dual_learning_for_machine_translation.1605752284.txt.gz · Last modified: 2023/06/15 07:36 (external edit)

Donate Powered by PHP Valid HTML5 Valid CSS Driven by DokuWiki