“NLP Status Report 2017-4-5”版本间的差异

• Got the sampled 100w good data and ran Moses (BLEU: 30.6)
• Reimplemented the idea of ACL (added some optimization to the previous code) and check the performance in the following gradual steps: 1. use s_i-1 as memory query; 2. use s_i-1+c_i

as memory query; 3. use y as the memory states for attention; 4. use y + smt_attentions * h as memory states for attention.

• ran experiments for the above steps but the loss was inf. I am looking for reasons.

• do experiments and write the paper

• I did keyword expansion on the qx's model
• fixed some bugs
• read two papers

• improve the effect of the qx's model

• revise the original oov model so that it can automatically detect oov words and translate them
• deal with the situation that source word is oov but target word is not oov first
• it didn't predict right

• make the model work as what we wanted
• deal with the situation that source word is oov and target word is also oov, then other situations

• got a reasonable baseline on big zhen data

• implement mem model on this baseline, and test on big data