“NLP Status Report 2017-4-5”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
（以“{| class="wikitable" !Date !! People !! Last Week !! This Week |- | rowspan="6"|2017/3/27 |Yang Feng || *tested for the baseline but cannot get the reasonable result...”为内容创建页面）
 
 
（4位用户的7个中间修订版本未显示）
第2行： 第2行：
 
!Date !! People !! Last Week !! This Week
 
!Date !! People !! Last Week !! This Week
 
|-
 
|-
−
| rowspan="6"|2017/3/27
+
| rowspan="6"|2017/4/5
 
|Yang Feng ||
 
|Yang Feng ||
−
*tested for the baseline but cannot get the reasonable result.
+
* Got the sampled 100w good data and ran Moses (BLEU: 30.6)
−
*debug the baseline to try to reproduce the good result but failed.
+
* 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.
−
*fixed the problem of nan in alpha-gamma method but the result is not good.
+
* ran experiments for the above steps but the loss was inf. I am looking for reasons.
−
*changed the calculation of probability for alpha-gamma method but the result is neither good.
+
−
*ran Moses for cwmt zh-en translation, but the training data is case-sensitive, so need to rerun.
+
 
||
 
||
−
*rerun Moses for cwmt zh-en and cs-en
+
*do experiments and write the paper
−
*decide to use tensorflow or theano
+
−
*run experiments based on the chosen platform
+
 
|-
 
|-
 
|Jiyuan Zhang ||
 
|Jiyuan Zhang ||
−
* I did nothing (I found my ACL mark was at borderline, so I didn't have a mind to work)
+
* I did keyword expansion on the qx's model
 +
*fixed some bugs
 +
*read two papers
 
||  
 
||  
 
*improve the effect of the qx's model
 
*improve the effect of the qx's model
 
|-
 
|-
 
|Andi Zhang ||
 
|Andi Zhang ||
−
*fixed the bug, turns out it rises from the unfamiliarity with numpy.resize() function
+
*revise the original oov model so that it can automatically detect oov words and translate them
−
*the demo model can deal with oov problem(both source word and target word are oov)
+
*deal with the situation that source word is oov but target word is not oov first
 +
*it didn't predict right
 
||
 
||
−
*some paper work about graduation design
+
*make the model work as what we wanted
−
*run some experiments using theano on old data set and new zh2en from lihang
+
*deal with the situation that source word is oov and target word is also oov, then other situations
 
|-
 
|-
 
|Shiyue Zhang ||  
 
|Shiyue Zhang ||  
−
* got good results from gnmt
+
* got a reasonable baseline on big zhen data
−
* but haven't found the way to implement our model on gnmt
+
−
* trying to modify our code to make it work on big data
+
 
||
 
||
−
* go on trying to modify our code to make it work on big data
+
* implement mem model on this baseline, and test on big data
−
* go on looking into gnmt code
+
 
|-
 
|-
 
|Peilun Xiao ||
 
|Peilun Xiao ||

2017年4月5日 (三) 02:16的最后版本

Date People Last Week This Week
2017/4/5 Yang Feng
  • 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
Jiyuan Zhang
  • I did keyword expansion on the qx's model
  • fixed some bugs
  • read two papers
  • improve the effect of the qx's model
Andi Zhang
  • 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
Shiyue Zhang
  • got a reasonable baseline on big zhen data
  • implement mem model on this baseline, and test on big data
Peilun Xiao
取自“http://index.cslt.org/mediawiki/index.php?title=NLP_Status_Report_2017-4-5&oldid=26612”