Closed lucasjinreal closed 5 years ago
Hi @jinfagang I'm not aware of any tutorials, but from my experience, it is not very different from other language pairs. Depending on dataset size transformer_vaswani_wmt_en_de_big or transformer_iwslt_de_en work very well. The only big difference is preprocessing, I used jieba tokenizer to tokenize data: https://github.com/fxsjy/jieba and then I applied BPE encoding (learning different vocabularies for En and Zh) and standard preprocess.py without --joined-dictionary
Does it make sense?
@edunov Thanks, the big question is about hyper parameters. I'll take a dive into that
@edunov Thanks, the big question is about hyper parameters. I'll take a dive into that
I'm not sure whether my answer is useful for anyone coming later. For the hyper parameters of en-zh translation model, you can refer to the setup in this paper http://arxiv.org/abs/1803.05567.
Hi, is there any resources or tutorial abouth translate Chinese to English?