Multi-task Sequence to Sequence Learning

Posted by Kaiyuan Chen on November 26, 2017

Multi-task Sequence to Sequence Learning

By sequence to sequence model, we have three basic settings

  • one-to-many settings
  • many-to-one setting
  • many-to-many settings

for multi-task learning.


have one encoder in common, such as translation and parsing and have multiple decoders

for example, input is a sequence of English words, a separate decoder can generate translation in German,tags in parsing and same sequence of English words.

many to one

only one decoder is shared, such as image captioning


multiple input encoder <=> multiple output decoder


autoencoder helps less in terms of perplexities but more on BLEU scores compared to skip-thought.


all of them are on machine translations