BLEU score of word features model is lower than Base model

Support
1

Hi, i’'m using OpenNMT-py to create Eng-Viet neural machine translation use word features as improvement. Word features are pos tag, lemma, cluster and named entity tag. I trained my model with same params with Base model (without word features) but BLEU score of improved model is lower than base model. Did i do something wrong ?
Here my preprocess and training script:
+Base model: BLEU = 24.20, 58.7/32.8/19.3/11.6
image
image

+Word features model: 22.44, 57.3/30.9/17.7/10.4
image
image

2

Hi,

Did you try to train for longer?

3

Hi, I met the same question . Have you solved this problem?

4

I end up with using fairse and custom features concatenation in word embedding level