Some translation errors in the real system

(Netxiao) #1

I have applied openNMT to my translation system as the core of the translation engine, It is very stable and effect is very good, I have translated tens of thousands of sentences, found some typical translation errors, I do not know why, please give me some advice.

My data set has 18 million sentences,vocab size:150k, seq_length:35,last PPL:3.82, BLEU:50

example 1:

ori:last day of month calculation

tgt:计算 的 最后 一天

lost word: month

example 2:
ori:What happened to mosedee in 2008?


lost: What happened to

example 3:
ori:my following code for typeid does not output the correct type name.

tgt:下面的_&typeid my typeid & for for_输出的类型名称不正确。

wrong translate: &typeid my typeid & for for

example 4:
ori:i do not understand name1 ,name1 name3

tgt:我 不 理解 name1,. **name2 name2**

repeat : name2 name2