“2013-10-25”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
Noisy training
 
(相同用户的一个中间修订版本未显示)
第15行: 第15行:
  
 
# The 863 clean test shows consistent performance.  
 
# The 863 clean test shows consistent performance.  
# Investigation shows that the car noise corrupted the speech patterns more seriously with the current fb-based noise adding. This leads to worse performance with car noises than with white noises. Note this observation is valid only with the fb-based noise adding
+
# Investigation shows that the car noise corrupted the speech patterns more seriously with the current fb-based noise adding. This leads to worse performance with car noises than with white noises. Note this observation is valid only with the fb-based noise adding.
  
 
=== Tencent exps ===
 
=== Tencent exps ===
第22行: 第22行:
  
 
==LM development==
 
==LM development==
===Continuous LM ===
+
 
 +
===NN LM ===
  
 
# The lattice rescore toolkit is done. The efficiency problem is solved by DP.  
 
# The lattice rescore toolkit is done. The efficiency problem is solved by DP.  
第52行: 第53行:
  
 
</pre>
 
</pre>
 +
 
===QA LM===
 
===QA LM===
  
 
#  Tencent word segmentation system ready.  
 
#  Tencent word segmentation system ready.  
 
#  Collecting data for Q-LM training.
 
#  Collecting data for Q-LM training.

2013年10月25日 (五) 01:48的最后版本

Data sharing

  • LM count files still undelivered!

AM development

Sparse DNN

  • Optimal Brain Damage(OBD).
  1. The initial observation shows that, after direct/obd cutting, obd is better than direct cutting in terms of training and testing accuracy.
  2. The sticky training goes.

Noisy training

  1. The 863 clean test shows consistent performance.
  2. Investigation shows that the car noise corrupted the speech patterns more seriously with the current fb-based noise adding. This leads to worse performance with car noises than with white noises. Note this observation is valid only with the fb-based noise adding.

Tencent exps

N/A


LM development

NN LM

  1. The lattice rescore toolkit is done. The efficiency problem is solved by DP.
  2. The perplexity of 10-NN is worse than 1-NN
  3. Initial lattices rescoring shows that the rescoring needs remove the old ngram score and then insert the new score. The 1st nn works better than the 10-nn.
  4. The CSLM does not improve the performance so far. This is conflict to the ppl results.
  5. Investigate better ways to combine multiple NNs. Investigate rescoring n-best results instead of lattices.


lattice-rescore: traing-data:500M,dict:11w,test:tencent

id        cslm(lopp) cslm(-logp)   n-gram  n-gram(replace) cslm(-old+cslm) cslm((0-1024)) cslm((0-1024),(1024-2048))
map:        53.86         38.44      34.40    37.38          35.38                34.94      34.91
2044:       47.81         32.20      27.18    31.18          28.53                27.65      27.87
notetp3:    43.01         24.61      18.78    22.34          20.18                19.64      19.81
record1900: 32.14         20.42      12.89    20.10          13.97                13.52      13.81
general:    62.86         49.17      43.23    47.50          45.55                44.28      44.62
online1:    62.79         46.76      39.49    46.40          40.26                39.91      40.01       
online2:    58.20         39.21      32.15    38.75          33.04                32.61      32.65
speedup:    52.50         38.74      31.47    37.72          34.01                32.65      32.81

note: cslm(logp) --replace lattice value1 with cslm lnp
      cslm(-logp)--replace lattice value1 with cslm -lnp
      ngram(replace)--replace lattice vaule1 with ngram -lnp
      cslm((-old+cslm)--repace lattice value1 = -old_ngram+cslm 
      cslm((0-1024))--replace lattice value1=-old_ngram+cslm (slist:0-1024,mach_num=1)
      cslm((0-1024),(1024-2048))--replace lattice value1=-old_ngram+cslm (slist:0-1024,1024-1048,mach_num=2)

QA LM

  1. Tencent word segmentation system ready.
  2. Collecting data for Q-LM training.