“NLP Status Report 2016-12-19”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"|2016/12/12 |Yang Feng || *s2smn: installed tensorflow and ran a toy example (solv...”为内容创建页面）
 
 
（6位用户的12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"|2016/12/12
+
| rowspan="6"|2016/12/19
 
|Yang Feng ||
 
|Yang Feng ||
−
*[[s2smn:]] installed tensorflow and ran a toy example (solved problems: version conflict and memory exhausted)
+
*[[s2smn:]] wrote the manual of s2s with tensorflow [[http://cslt.riit.tsinghua.edu.cn/mediawiki/images/5/51/Nmt-tensorflow-mannua-yfeng.pdf nmt-manual]]
−
*wrote the code of the memory  network part
+
*wrote part of the code of mn.
−
*[[Huilan:]] prepared for periodical report and system submission.
+
*wrote the manual of Moses [[http://cslt.riit.tsinghua.edu.cn/mediawiki/images/9/92/Moses%E6%93%8D%E4%BD%9C%E6%89%8B%E5%86%8C--%E5%86%AF%E6%B4%8B.pdf moses-manual]]
 +
*[[Huilan:]] fixed the problem of syntax-based translation.
 +
*sort out the system and corresponding documents.
 
||
 
||
−
*[[s2smn:]] finish the manual of nmt tensorflow
+
*[[s2smn:]] finish the code of adding mn.
−
*[[Huilan:]] system submission
+
*[[Huilan:]] handover.
 
|-
 
|-
 
|Jiyuan Zhang ||
 
|Jiyuan Zhang ||
−
*attempted to use memory model to improve the atten model of bad effect
+
*coded tone_model,but had some trouble
−
*With the vernacular as the input，generated poem by local atten model[http://cslt.riit.tsinghua.edu.cn/mediawiki/images/2/2f/Local_atten_resluts.pdf]
+
*run global_attention_model that decodes four sentences, [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/d/d5/Four_local_atten.pdf four][http://cslt.riit.tsinghua.edu.cn/mediawiki/images/0/05/Five_local_attention.pdf five]generated by local_attention model
−
*Modified working mechanism of memory model(top1 to average)
+
−
*help andi
+
 
||  
 
||  
 
*improve poem model   
 
*improve poem model   
 
|-
 
|-
 
|Andi Zhang ||
 
|Andi Zhang ||
−
*prepared a paraphrase data set that is enumerated from a previous one (ignoring words like "啊呀哈")
+
*coded to output encoder outputs and correspoding source & target sentences(ids in dictionaries)
−
*worked on coding bidirectional model under tensorflow, met with NAN problem
+
*coded a script for bleu scoring, which tests the five checkpoints auto created by training process and save the one with best performance
 
||
 
||
−
*ignore NAN problem for now, run it on the same data set used in Theano
+
*extract encoder outputs
 
|-
 
|-
 
|Shiyue Zhang ||  
 
|Shiyue Zhang ||  
−
* finished tsne pictures, and discussed with teachers
+
* changed the one-hot vector to (0, -inf, -inf...), and retied the experiments. But no improvement showed.
−
* tried experiments with 28-dim mem, but found almost all of them converged to baseline
+
* tried 1-dim gate, but converged to baseline
−
* returned to 384-dim mem, which is still slightly better than basline.
+
* tried to only train gate, but the best is taking all instance as "right"
−
* found the problem of action mem, one-hot vector is not proper.
+
* trying a model similar to attention
−
* [[http://cslt.riit.tsinghua.edu.cn/mediawiki/images/2/2f/RNNG%2Bmm%E5%AE%9E%E9%AA%8C%E6%8A%A5%E5%91%8A.pdf report]]
+
* [[http://cslt.riit.tsinghua.edu.cn/mediawiki/images/9/9f/RNNG%2Bmm_experiment_report.pdf report]]
 
||
 
||
−
* change one-hot vector to (0, -10000.0, -10000.0...)
+
* try to add true action info when training gate
−
* try 1-dim gate
+
* try different scale vectors
−
* try max cos
+
* try to change cos to only inner product
 
|-
 
|-
 
|Guli ||
 
|Guli ||
−
*install and run moses
+
*read papers about Transfer learning and solving OOV
−
*prepare thesis report
+
*conducted comparative test
 +
*writing survey
 
||
 
||
−
*read papers about Transfer learning and solving OOV
+
* complete the first draft of the survey 
 
|-
 
|-
 
|Peilun Xiao ||
 
|Peilun Xiao ||
−
*Read a paper about document classification wiht GMM distributions of word vecotrs and try to code it in python
+
*use LDA to generate 10-500 dimension document vector in the rest datasets
−
*Use LDA to reduce the dimension of the text in r52、r8 and contrast the performance of classification
+
*write a python code about a new algorithm about tf-idf
 
||
 
||
−
*Use LDA to reduce the dimension of the text in 20news and webkb
+
*debug the code
 
|}
 
|}

2016年12月26日 (一) 00:58的最后版本

Date People Last Week This Week
2016/12/19 Yang Feng
  • s2smn: wrote the manual of s2s with tensorflow [nmt-manual]
  • wrote part of the code of mn.
  • wrote the manual of Moses [moses-manual]
  • Huilan: fixed the problem of syntax-based translation.
  • sort out the system and corresponding documents.
Jiyuan Zhang
  • coded tone_model,but had some trouble
  • run global_attention_model that decodes four sentences, fourfivegenerated by local_attention model
  • improve poem model
Andi Zhang
  • coded to output encoder outputs and correspoding source & target sentences(ids in dictionaries)
  • coded a script for bleu scoring, which tests the five checkpoints auto created by training process and save the one with best performance
  • extract encoder outputs
Shiyue Zhang
  • changed the one-hot vector to (0, -inf, -inf...), and retied the experiments. But no improvement showed.
  • tried 1-dim gate, but converged to baseline
  • tried to only train gate, but the best is taking all instance as "right"
  • trying a model similar to attention
  • [report]
  • try to add true action info when training gate
  • try different scale vectors
  • try to change cos to only inner product
Guli
  • read papers about Transfer learning and solving OOV
  • conducted comparative test
  • writing survey
  • complete the first draft of the survey
Peilun Xiao
  • use LDA to generate 10-500 dimension document vector in the rest datasets
  • write a python code about a new algorithm about tf-idf
  • debug the code
取自“http://index.cslt.org/mediawiki/index.php?title=NLP_Status_Report_2016-12-19&oldid=25027”