How to handle the 'repeat pattern' in the output


(Cnglen) #1

The output sequence of my model has lots of repeat pattern, such as

  • (at sentence level): <眼睛是心灵的窗户>,<眼睛是心灵的窗户>,<眼睛是心灵的窗户>,<眼睛是心灵的窗户>,
  • (at word level): 拯救干燥肌肤,<保湿><保湿>护肤品推荐

A year later after the http://forum.opennmt.net/t/how-to-deal-with-repetition/179, Any suggestions?


(Guillaume Klein) #2

This usually happens when the network receives an uncommon input, either due to a preprocessing mismatch or a lack of training data.