[Moses-support] Strange behavior of alternative weights translation

jian zhang Sun, 11 Oct 2015 15:30:38 -0700

Hi all,

I have two source files (same sentences, but in different order), something
like


File source-d1-d2:
<seg weight-setting="domain1"> 今年 前 两 月 广东 高新技术 产品 出口 37.6亿 美元 </seg>
<seg weight-setting="domain2"> 新华社 广州 3月 16日 电 ( 记者 陈冀 ) 最新 统计 数字 显示 , 今年 1
至 2月 , 广东省 高新技术 产品 出口 37.6亿 美元 , 同比 增长 34.8% , 占 全省 出口 总值 的 25.5% 。 </seg>

and source-d2-d1:
<seg weight-setting="domain2"> 新华社 广州 3月 16日 电 ( 记者 陈冀 ) 最新 统计 数字 显示 , 今年 1
至 2月 , 广东省 高新技术 产品 出口 37.6亿 美元 , 同比 增长 34.8% , 占 全省 出口 总值 的 25.5% 。 </seg>
<seg weight-setting="domain1"> 今年 前 两 月 广东 高新技术 产品 出口 37.6亿 美元 </seg>

The I run the following commands,
moses -threads 1 -f moses-domain-weights.ini < source-d2-d1 &>
 translation-d2-d1.log
moses -threads 1 -f moses-domain-weights.ini < source-d1-d2 &>
 translation-d1-d2.log



The output of first command, in translation-d1-d2.log is
Defined parameters (per moses.ini or switch):
alternate-weight-setting: id=domain1 LexicalReordering0= -0.00467702
0.0430348 0.0889973 0.0352209 0.000507535 0.48658 Distortion0= 0.0042172
LM0= 0.0350763 WordPenalty0= -0.149218 PhrasePenalty0= 0.0575948
TranslationModel0= 0.0122912 0.0178517 0.0438588 0.0208747
UnknownWordPenalty0= 1 id=domain2 LexicalReordering0= -0.0356997 0.0133961
0.00699043 0.760662 0.0874955 0.0278831 Distortion0= 0.00562152 LM0=
0.0105839 WordPenalty0= -0.0239658 PhrasePenalty0= 0.00725696
TranslationModel0= 0.00534787 0.00493235 0.00707 0.00309476
UnknownWordPenalty0= 1 id=domain3 LexicalReordering0= 0.0299658 0.0642291
0.132828 0.0099368 0.000757492 0.0317366 Distortion0= 0.00629415 LM0=
0.0543508 WordPenalty0= -0.340616 PhrasePenalty0= 0.0859599
TranslationModel0= 0.0183444 0.0266436 0.0654589 0.132879
UnknownWordPenalty0= 1 id=domain4 LexicalReordering0= 0.22727 0.0573986
0.465178 0.0491854 0.00318964 -0.00912402 Distortion0= 0.0350783 LM0=
0.0481925 WordPenalty0= 0.00113805 PhrasePenalty0= 0.00989932
TranslationModel0= 0.0232946 0.0316844 0.019047 -0.0203211
UnknownWordPenalty0= 1
config: moses-domain-weights.ini
distortion-limit: 6
feature: UnknownWordPenalty WordPenalty PhrasePenalty
PhraseDictionaryOnDisk name=TranslationModel0 num-features=4
path=/home/userfiltered/filtered.1/phrase-table.0-0.1.1.bin input-factor=0
output-factor=0 LexicalReordering name=LexicalReordering0 num-features=6
type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0
path=/home/userfiltered/filtered.1/reordering-table.1.wbe-msd-bidirectional-fe
Distortion KENLM lazyken=0 name=LM0 factor=0
path=/home/userbaseline/lm/lm.lm.1.bin order=5
input-factors: 0
mapping: 0 T 0
threads: 1 distinct
v: 0
weight: LexicalReordering0= 0.05546 0.0200971 0.212035 0.0738027 0.050521
0.0982874 Distortion0= 0.00893496 LM0= 0.0642775 WordPenalty0= -0.194264
PhrasePenalty0= 0.0867617 TranslationModel0= 0.0228662 0.0378634 0.0388449
0.0359839 UnknownWordPenalty0= 1
line=UnknownWordPenalty
FeatureFunction: UnknownWordPenalty0 start: 0 end: 0
line=WordPenalty
FeatureFunction: WordPenalty0 start: 1 end: 1
line=PhrasePenalty
FeatureFunction: PhrasePenalty0 start: 2 end: 2
line=PhraseDictionaryOnDisk name=TranslationModel0 num-features=4
path=/home/userfiltered/filtered.1/phrase-table.0-0.1.1.bin input-factor=0
output-factor=0
FeatureFunction: TranslationModel0 start: 3 end: 6
line=LexicalReordering name=LexicalReordering0 num-features=6
type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0
path=/home/userfiltered/filtered.1/reordering-table.1.wbe-msd-bidirectional-fe
FeatureFunction: LexicalReordering0 start: 7 end: 12
Initializing Lexical Reordering Feature..
line=Distortion
FeatureFunction: Distortion0 start: 13 end: 13
line=KENLM lazyken=0 name=LM0 factor=0
path=/home/userbaseline/lm/lm.lm.1.bin order=5
FeatureFunction: LM0 start: 14 end: 14
Loading UnknownWordPenalty0
Loading WordPenalty0
Loading PhrasePenalty0
Loading LexicalReordering0
binary file loaded, default OFF_T: -1
Loading Distortion0
Loading LM0
Loading TranslationModel0
alternate weight setting domain1
alternate weight setting domain2
alternate weight setting domain3
alternate weight setting domain4
Created input-output object : [0.514] seconds
Translating: 今年 前 两 月 广东 高新技术 产品 出口 37.6亿 美元
binary file loaded, default OFF_T: -1
Line 0: Initialize search took 0.529 seconds total
Line 0: Collecting options took 0.026 seconds at moses/Manager.cpp:117
Line 0: Search took 0.455 seconds
the first two months of this year guangdong &apos; s export of high @-@
tech products 37.6亿 us dollars
BEST TRANSLATION: the first two months of this year guangdong &apos; s
export of high @-@ tech products 37.6亿|UNK|UNK|UNK us dollars [1111111111]
 [total=-101.595]
core=(-100.000,-19.000,5.000,-10.157,-15.979,-14.790,-32.019,-4.474,0.000,0.000,-2.293,0.000,0.000,0.000,-83.558)

Line 0: Decision rule took 0.001 seconds total
Line 0: Additional reporting took 0.001 seconds total
Line 0: Translation took 1.011 seconds total
Translating: 新华社 广州 3月 16日 电 ( 记者 陈冀 ) 最新 统计 数字 显示 , 今年 1 至 2月 , 广东省 高新技术
产品 出口 37.6亿 美元 , 同比 增长 34.8% , 占 全省 出口 总值 的 25.5% 。
Line 1: Initialize search took 0.150 seconds total
Line 1: Collecting options took 0.082 seconds at moses/Manager.cpp:117
Line 1: Search took 1.328 seconds
xinhua news agency , guangzhou on march 16 , xinhua ( reporter chen ji ) ,
up @-@ to @-@ date on during the january @-@ february period . exports of
high @-@ tech products in guangdong 37.6亿 and 34.8 percent of the total
export value , accounting for 25.5 per cent throughout the province .
BEST TRANSLATION: xinhua news agency , guangzhou on march 16 , xinhua (
reporter chen ji ) , up @-@ to @-@ date on during the january @-@ february
period . exports of high @-@ tech products in guangdong 37.6亿|UNK|UNK|UNK
and 34.8 percent of the total export value , accounting for 25.5 per cent
throughout the province . [1111111111111111111111111111111111111]
 [total=-102.559]
core=(-100.000,-56.000,21.000,-48.972,-99.568,-54.978,-118.824,-10.884,-0.669,-6.607,-0.423,-0.336,-4.584,-50.000,-200.097)

Line 1: Decision rule took 0.000 seconds total
Line 1: Additional reporting took 0.000 seconds total
Line 1: Translation took 1.560 seconds total
Name:moses VmPeak:2003732 kB VmRSS:41416 kB RSSMax:1854592 kB user:2.884
sys:0.300 CPU:3.184 real:3.189


And the output of second command, in translation-d2-d1.log is
Defined parameters (per moses.ini or switch):
alternate-weight-setting: id=domain1 LexicalReordering0= -0.00467702
0.0430348 0.0889973 0.0352209 0.000507535 0.48658 Distortion0= 0.0042172
LM0= 0.0350763 WordPenalty0= -0.149218 PhrasePenalty0= 0.0575948
TranslationModel0= 0.0122912 0.0178517 0.0438588 0.0208747
UnknownWordPenalty0= 1 id=domain2 LexicalReordering0= -0.0356997 0.0133961
0.00699043 0.760662 0.0874955 0.0278831 Distortion0= 0.00562152 LM0=
0.0105839 WordPenalty0= -0.0239658 PhrasePenalty0= 0.00725696
TranslationModel0= 0.00534787 0.00493235 0.00707 0.00309476
UnknownWordPenalty0= 1 id=domain3 LexicalReordering0= 0.0299658 0.0642291
0.132828 0.0099368 0.000757492 0.0317366 Distortion0= 0.00629415 LM0=
0.0543508 WordPenalty0= -0.340616 PhrasePenalty0= 0.0859599
TranslationModel0= 0.0183444 0.0266436 0.0654589 0.132879
UnknownWordPenalty0= 1 id=domain4 LexicalReordering0= 0.22727 0.0573986
0.465178 0.0491854 0.00318964 -0.00912402 Distortion0= 0.0350783 LM0=
0.0481925 WordPenalty0= 0.00113805 PhrasePenalty0= 0.00989932
TranslationModel0= 0.0232946 0.0316844 0.019047 -0.0203211
UnknownWordPenalty0= 1
config: moses-domain-weights.ini
distortion-limit: 6
feature: UnknownWordPenalty WordPenalty PhrasePenalty
PhraseDictionaryOnDisk name=TranslationModel0 num-features=4
path=/home/userfiltered/filtered.1/phrase-table.0-0.1.1.bin input-factor=0
output-factor=0 LexicalReordering name=LexicalReordering0 num-features=6
type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0
path=/home/userfiltered/filtered.1/reordering-table.1.wbe-msd-bidirectional-fe
Distortion KENLM lazyken=0 name=LM0 factor=0
path=/home/userbaseline/lm/lm.lm.1.bin order=5
input-factors: 0
mapping: 0 T 0
threads: 1 distinct
v: 0
weight: LexicalReordering0= 0.05546 0.0200971 0.212035 0.0738027 0.050521
0.0982874 Distortion0= 0.00893496 LM0= 0.0642775 WordPenalty0= -0.194264
PhrasePenalty0= 0.0867617 TranslationModel0= 0.0228662 0.0378634 0.0388449
0.0359839 UnknownWordPenalty0= 1
line=UnknownWordPenalty
FeatureFunction: UnknownWordPenalty0 start: 0 end: 0
line=WordPenalty
FeatureFunction: WordPenalty0 start: 1 end: 1
line=PhrasePenalty
FeatureFunction: PhrasePenalty0 start: 2 end: 2
line=PhraseDictionaryOnDisk name=TranslationModel0 num-features=4
path=/home/userfiltered/filtered.1/phrase-table.0-0.1.1.bin input-factor=0
output-factor=0
FeatureFunction: TranslationModel0 start: 3 end: 6
line=LexicalReordering name=LexicalReordering0 num-features=6
type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0
path=/home/userfiltered/filtered.1/reordering-table.1.wbe-msd-bidirectional-fe
FeatureFunction: LexicalReordering0 start: 7 end: 12
Initializing Lexical Reordering Feature..
line=Distortion
FeatureFunction: Distortion0 start: 13 end: 13
line=KENLM lazyken=0 name=LM0 factor=0
path=/home/userbaseline/lm/lm.lm.1.bin order=5
FeatureFunction: LM0 start: 14 end: 14
Loading UnknownWordPenalty0
Loading WordPenalty0
Loading PhrasePenalty0
Loading LexicalReordering0
binary file loaded, default OFF_T: -1
Loading Distortion0
Loading LM0
Loading TranslationModel0
alternate weight setting domain1
alternate weight setting domain2
alternate weight setting domain3
alternate weight setting domain4
Created input-output object : [0.514] seconds
Translating: 新华社 广州 3月 16日 电 ( 记者 陈冀 ) 最新 统计 数字 显示 , 今年 1 至 2月 , 广东省 高新技术
产品 出口 37.6亿 美元 , 同比 增长 34.8% , 占 全省 出口 总值 的 25.5% 。
binary file loaded, default OFF_T: -1
Line 0: Initialize search took 0.276 seconds total
Line 0: Collecting options took 0.078 seconds at moses/Manager.cpp:117
Line 0: Search took 1.265 seconds
xinhua news agency , guangzhou on march 16 ( reporter chen ji ) @-@ @-@ to
@-@ date on to february , the province this year exports of high @-@ tech
products 37.6亿 and 34.8 percent of the total export value , accounting for
25.5 per cent throughout the province .
BEST TRANSLATION: xinhua news agency , guangzhou on march 16 ( reporter
chen ji ) @-@ @-@ to @-@ date on to february , the province this year
exports of high @-@ tech products 37.6亿|UNK|UNK|UNK and 34.8 percent of the
total export value , accounting for 25.5 per cent throughout the province .
[1111111111111111111111111111111111111]  [total=-102.552]
core=(-100.000,-51.000,22.000,-62.087,-100.147,-58.165,-87.079,-12.522,-0.669,-8.377,-0.298,-0.336,-5.141,-56.000,-197.660)

Line 0: Decision rule took 0.001 seconds total
Line 0: Additional reporting took 0.001 seconds total
Line 0: Translation took 1.619 seconds total
Translating: 今年 前 两 月 广东 高新技术 产品 出口 37.6亿 美元
Line 1: Initialize search took 0.141 seconds total
Line 1: Collecting options took 0.013 seconds at moses/Manager.cpp:117
Line 1: Search took 0.237 seconds
the first two months of this year guangdong &apos; s export of high @-@
tech products 37.6亿 us dollars
BEST TRANSLATION: the first two months of this year guangdong &apos; s
export of high @-@ tech products 37.6亿|UNK|UNK|UNK us dollars [1111111111]
 [total=-101.595]
core=(-100.000,-19.000,5.000,-10.157,-15.979,-14.790,-32.019,-4.474,0.000,0.000,-2.293,0.000,0.000,0.000,-83.558)

Line 1: Decision rule took 0.000 seconds total
Line 1: Additional reporting took 0.000 seconds total
Line 1: Translation took 0.391 seconds total
Name:moses VmPeak:2003732 kB VmRSS:46228 kB RSSMax:1854384 kB user:2.329
sys:0.268 CPU:2.597 real:2.603

The weights in ini are defined as:

[weight]
LexicalReordering0= 0.05546 0.0200971 0.212035 0.0738027 0.050521 0.0982874
Distortion0= 0.00893496
LM0= 0.0642775
WordPenalty0= -0.194264
PhrasePenalty0= 0.0867617
TranslationModel0= 0.0228662 0.0378634 0.0388449 0.0359839
UnknownWordPenalty0= 1


[alternate-weight-setting]
id=domain1
LexicalReordering0= -0.00467702 0.0430348 0.0889973 0.0352209 0.000507535
0.48658
Distortion0= 0.0042172
LM0= 0.0350763
WordPenalty0= -0.149218
PhrasePenalty0= 0.0575948
TranslationModel0= 0.0122912 0.0178517 0.0438588 0.0208747
UnknownWordPenalty0= 1

id=domain2
LexicalReordering0= -0.0356997 0.0133961 0.00699043 0.760662 0.0874955
0.0278831
Distortion0= 0.00562152
LM0= 0.0105839
WordPenalty0= -0.0239658
PhrasePenalty0= 0.00725696
TranslationModel0= 0.00534787 0.00493235 0.00707 0.00309476
UnknownWordPenalty0= 1

id=domain3
LexicalReordering0= 0.0299658 0.0642291 0.132828 0.0099368 0.000757492
0.0317366
Distortion0= 0.00629415
LM0= 0.0543508
WordPenalty0= -0.340616
PhrasePenalty0= 0.0859599
TranslationModel0= 0.0183444 0.0266436 0.0654589 0.132879
UnknownWordPenalty0= 1

id=domain4
LexicalReordering0= 0.22727 0.0573986 0.465178 0.0491854 0.00318964
-0.00912402
Distortion0= 0.0350783
LM0= 0.0481925
WordPenalty0= 0.00113805
PhrasePenalty0= 0.00989932
TranslationModel0= 0.0232946 0.0316844 0.019047 -0.0203211
UnknownWordPenalty0= 1


But, for the sentence, which uses the weights of domain2, two runs give me
different translations. Did I miss anything here?


Thanks,

Jian
-- 
Jian Zhang
Centre for Next Generation Localisation (CNGL)
<http://www.cngl.ie/index.html>
Dublin City University <http://www.dcu.ie/>

_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

[Moses-support] Strange behavior of alternative weights translation

Reply via email to