This is an implementation seq2seq with attention for neural machine translation with MXNet.
The current code uses IWSLT 2009 Chinese-English corpus as training, development and test data. Please request this data set or use other available parallel corpus. Data statistics,
| training | dev | test |
|---|---|---|
| 81819 | 446 | 504 |
The reference number of IWSLT 2009 Ch-En is 7, for example:
在 找 给 家里 人 的 礼物 .
i 'm searching for some gifts for my family .
i want to find something for my family as presents .
i 'm about to buy some presents for my family .
i 'd like to buy my family something as a gift .
i 'm looking for a gift for my family .
i 'm looking for a present for my family .
i need a gift for my family .
有 $number 块 钱 以下 的 茶 吗 ? |||| {1 ||| 1 ||| one thousand ||| $number ||| 一千}
do you have any tea under one thousand yen ?
i 'd like to take a look at some tea cheaper than one thousand yen .
is there any tea less than one thousand yen here ?
i 'm looking for some tea under one thousand yen .
do you have any tea lower than one thousand yen ?
do you have any tea less than one thousand yen ?
i would like to buy some tea cheaper than one thousand yen .
According to my test, this code can achieve 44.18 BLEU score (with beam search) on IWSLT dev set without post-processing after 53 iteration. Specifically,
1gram=72.65% 2gram=49.63% 3gram=37.62% 4gram=28.08% BP = 1.0000 BLEU = 0.4418
- Compatibility issue. The current version will ask to use Python 3 since it is annoying to handle Chinese encoding problems for Python 2.
- In the attention part,
h.dot(U)should be pre-computed. However it seems that it won't work properly if I do so. - The BLEU evaluator, which is an exe file and not included, should be replaced by nltk evaluator in the future.
- The model can be modified to make it achieve about 50 BLEU score on this data set.