“NLP Status Report 2016-11-21”版本间的差异

• rnng+mn

1) ran experiments of rnng+mn [report] ; 2) used top-k for memory, under training

• sequence-to-sequence + mn

1) wrote the proposal 2) discussed the details with Andy

• intern interview
• Huilan's work

• rnng+mn

1) get the result of top-k; 2) try bigger memory;

• sequence-to-sequence + mn

1)coding work

• Huilan's work

1) try syntax-based TM

• ran decoder-memory model, but the effect is not obvious
• changed binding way of memory and atten models, can generate different style of poetry
• cleanned up our code
• wrote a techreport about poemGen [1]
• submitted two databases about poemGen and musicGen

• explored a variety of binding ways of memory and atten model

• prepare new data set for paraphrase, wiped out repetition & most of the noises
• run NMT on fr-en data set and new paraphrase set
• read through source code to find ways to modify it
• helped Guli with running NMT on our server

• decide to drop theano or not
• start to work on codes

• run rnng on MKL successfully, which can double or triple the speed.
• rerun the original model and get the final result 92.32
• rerun the wrong memory models, still running
• implement the dynamic memory model and get the result 92.54 which is 0.22 better than baseline
• try another structure of memory

• try more different models and summary the results
• publish the technical reports