“NLP Status Report 2017-5-22”版本间的差异

• bleu of baseline = 43.87
• 2nd translator uses as training data the concat(Chinese, machine translated English):

 hidden_size, emb_size, lr = 500, 310, 0.001 bleu = 43.53 (best)
 hidden_size, emb_size, lr = 700, 510, 0.001 bleu = 45.21 (best) but most results are under 43.1
 hidden_size, emb_size, lr = 700, 510, 0.0005 bleu = 42.19 (best)

• double-decoder model with joint loss (final loss = 1st decoder's loss + 2nd decoder's loss):

 bleu = 40.11 (best)
 The 1st decoder's output is generally better than 2nd decoder's output.

• The training process of double-decoder model without joint loss is problematic.

• Overfitting? Train 2nd translator on large data
• Replace the forced teaching mechanism in training process with beam search mechanism.

• tried to not train embedding but use external word vectors
• most results of my attempts are bad, only 3-layer rnn + no dropout model got 25.54 bleu which about 2 points worse than original baseline
• trained original baseline on new data ( the data fixed the reverse sentence problem), got bleu=27.88; moses bleu=32.47

• try more models to get similar results as original baseline on new data
• m-nmt model on new data

• learn the implement of seq2seq model
• read tf_translate code

• understand the meaning of main code
• start writing documents