compare_mt.py Analysis Report

Aggregate Scores

BLEU
PBMT NMT Win?
BLEU 0.2243 0.2403 s2>s1
[0.2170,0.2315] [0.2331,0.2479] p=0.0000


Score Comparison

length ratio
PBMT NMT Win?
length ratio 0.9479 (ref=48183, out=45672) 0.9382 (ref=48183, out=45207) s1>s2
[0.9409,0.9550] [0.9296,0.9467] p=0.0030


Score Comparison

Word Accuracies

Word fmeas by frequency bucket
frequency PBMT NMT
<1 0.1005 0.0493
1 0.2226 0.0850
2 0.3430 0.1735
3 0.3644 0.2408
4 0.4364 0.1667
[5,10) 0.3664 0.2012
[10,100) 0.4843 0.3880
[100,1000) 0.5482 0.5160
>=1000 0.6377 0.6485


Word Accuracy Analysis

Word fmeas by labels bucket
labels PBMT NMT
CC 0.7877 0.8043
DT 0.5252 0.5634
IN 0.5360 0.5441
JJ 0.4991 0.4620
NN 0.5485 0.4828
NNP 0.6297 0.4930
NNS 0.5591 0.4992
PRP 0.6274 0.6584
RB 0.4861 0.4662
TO 0.6246 0.6360
VB 0.4720 0.4366
VBP 0.4339 0.4533
VBZ 0.4814 0.5031
other 0.6716 0.6661


Word Accuracy Analysis

Word fmeas by numerical labels bucket
numerical labels PBMT NMT
<0.25 0.6103 0.6139
[0.25,0.5) 0.5848 0.5773
[0.5,0.75) 0.5438 0.5189
>=0.75 0.5831 0.5627


Word Accuracy Analysis

Source Word Accuracies

Word fmeas by frequency bucket
frequency PBMT NMT
<1 0.0000 0.0000
1 0.4460 0.4100
2 0.5981 0.5658
3 0.5887 0.5652
4 0.6518 0.6298
[5,10) 0.6447 0.6380
[10,100) 0.6407 0.6508
[100,1000) 0.6000 0.6058
>=1000 0.0000 0.0000


Source Word Accuracy Analysis

Sentence Buckets

bucket type: length, statistic type: BLEU
length PBMT NMT
<10 0.2231 0.2532
[10,20) 0.2098 0.2467
[20,30) 0.2316 0.2379
[30,40) 0.2175 0.2318
[40,50) 0.2219 0.2190
[50,60) 0.2331 0.2554
>=60 0.2726 0.2470


Sentence Bucket Analysis

bucket type: reference-output length difference, statistic type: count
lengthdiff PBMT NMT
<-20 0 0
[-20,-10) 8 12
[-10,-5) 66 67
-5 41 44
-4 73 58
-3 115 97
-2 173 187
-1 263 256
0 394 423
1 330 344
2 296 275
3 219 194
4 127 153
5 107 93
[6,11) 194 182
[11,21) 39 51
>=21 0 9


Sentence Bucket Analysis

bucket type: sentence-level BLEU, statistic type: count
sentbleu PBMT NMT
<0.1 128 116
[0.1,0.2) 832 798
[0.2,0.3) 651 604
[0.3,0.4) 391 396
[0.4,0.5) 191 212
[0.5,0.6) 106 129
[0.6,0.7) 52 63
[0.7,0.8) 32 47
[0.8,0.9) 31 36
>=0.9 31 44


Sentence Bucket Analysis

Characteristic N-grams

min_ngram_length=1, max_ngram_length=4

report_length=50, alpha=1.0, compare_type=match

50 n-grams that PBMT had higher match
n-gram match PBMT NMT
phantom 0.9459 34 1
Amy 0.9091 9 0
, who 0.9000 8 0
my mother 0.8889 7 0
And so 0.8571 5 0
else happened 0.8571 5 0
my mother . 0.8571 5 0
something else happened 0.8571 5 0
Avelile 0.8571 5 0
so on 0.8333 4 0
file 0.8333 4 0
rightness 0.8333 4 0
and so on 0.8333 4 0
the fusiform 0.8333 4 0
mother . " 0.8333 4 0
my mother . " 0.8333 4 0
the phantom 0.8333 4 0
Center 0.8333 4 0
fine 0.8333 4 0
fusiform 0.8333 4 0
centers 0.8333 4 0
Oz 0.8333 4 0
mother . " " 0.8333 4 0
viewer 0.8333 4 0
instead 0.8182 8 1
exchange 0.8125 12 2
creativity can 0.8000 3 0
found that 0.8000 3 0
billion pixels 0.8000 3 0
Israel 0.8000 3 0
archive 0.8000 3 0
lose the ability 0.8000 3 0
the Bible 0.8000 3 0
Limpopo 0.8000 3 0
fusiform gyrus 0.8000 3 0
Yeah 0.8000 7 1
pixels 0.8000 3 0
lose the ability to 0.8000 3 0
The point 0.8000 3 0
command 0.8000 3 0
and so on . 0.8000 3 0
excited 0.8000 3 0
like my mother 0.8000 3 0
so on . 0.8000 3 0
exactly like 0.8000 3 0
phantom limbs 0.8000 3 0
Beth 0.8000 3 0
amygdala 0.8000 3 0
address 0.8000 3 0
Galton 0.8000 3 0


50 n-grams that NMT had higher match
n-gram match PBMT NMT
going to show you 0.1250 0 6
going to show 0.1250 0 6
Is 0.1429 0 5
, because the 0.1429 0 5
hemisphere 0.1429 0 5
't even 0.1429 0 5
Is it 0.1429 0 5
'm going to show 0.1429 0 5
: Okay , 0.1667 0 4
the process 0.1667 0 4
the light 0.1667 0 4
presence of 0.1667 0 4
the camera 0.1667 0 4
the present 0.1667 0 4
process of 0.1667 0 4
: Okay 0.1667 0 4
left hemisphere 0.1667 0 4
years old 0.1667 0 4
the presence of 0.1667 0 4
yet 0.1667 0 4
how do 0.1818 1 8
with a 0.1818 1 8
don 't have to 0.2000 0 3
size of 0.2000 0 3
, all the 0.2000 0 3
it doesn 't 0.2000 0 3
it doesn 0.2000 0 3
, he 0.2000 0 3
AG : Okay , 0.2000 0 3
I was a 0.2000 0 3
don 't even 0.2000 0 3
if they 0.2000 0 3
not a 0.2000 0 3
core of 0.2000 0 3
core of the 0.2000 0 3
a very early 0.2000 0 3
a little bit 0.2000 0 3
little bit 0.2000 0 3
to show you 0.2000 1 7
penalty 0.2000 0 3
a computer mouse 0.2000 0 3
Everything 0.2000 0 3
, there 's a 0.2000 0 3
death penalty 0.2000 0 3
, as if 0.2000 0 3
the surgeon 0.2000 0 3
AG : Okay 0.2000 0 3
he was 0.2000 0 3
to show 0.2000 1 7
that . 0.2000 0 3


min_ngram_length=1, max_ngram_length=4

report_length=50, alpha=1.0, compare_type=match

ref_labels=example/ted.ref.eng.tag, out0_labels=example/ted.sys1.eng.tag, out1_labels=example/ted.sys2.eng.tag,

50 n-grams that PBMT had higher match
n-gram match PBMT NMT
, RB RB 0.8571 5 0
WDT VBZ IN 0.8571 5 0
NN RB VBD 0.8571 5 0
WDT NN 0.8571 5 0
VBG NN 0.8571 5 0
RP DT 0.8333 4 0
NN . DT 0.8333 4 0
NN NN , 0.8333 9 1
RB IN . 0.8333 4 0
NNS VBD JJ 0.8000 3 0
NNP , PRP$ NN 0.8000 3 0
NN IN NNS RB 0.8000 3 0
NNS , RB RB 0.8000 3 0
WP , 0.8000 3 0
RB IN DT JJ 0.8000 3 0
VBD PRP , 0.8000 3 0
, WDT VBZ IN 0.8000 3 0
NNS CC VB 0.8000 3 0
JJS NN IN 0.8000 3 0
VBP DT NN TO 0.8000 3 0
JJ MD 0.8000 3 0
NN , NN CC 0.8000 3 0
VBD DT JJ NN 0.8000 3 0
JJR JJ 0.8000 3 0
NN VBD DT NN 0.8000 3 0
IN DT NNS . 0.8000 3 0
, NN , NN 0.8000 3 0
IN NNS RB 0.8000 3 0
PRP$ NN . CC 0.8000 3 0
VB . VB 0.8000 3 0
JJ NN RB 0.8000 3 0
CD CD NNS . 0.8000 3 0
VBG IN NNS 0.8000 3 0
CC RB IN . 0.8000 3 0
RP DT NN 0.8000 3 0
. RB 0.7778 6 1
VBD NNP 0.7778 6 1
NN CC NN . 0.7778 6 1
DT NN IN VBG 0.7500 2 0
CD NNP 0.7500 2 0
CD NN WDT 0.7500 2 0
NNS WDT MD 0.7500 2 0
NN VBZ WP 0.7500 2 0
, WP VBP 0.7500 2 0
NN : DT NN 0.7500 2 0
VBP JJ NNP 0.7500 2 0
PRP VBP NN . 0.7500 2 0
IN IN NNP 0.7500 2 0
IN TO VB 0.7500 2 0
, PRP RB 0.7500 2 0


50 n-grams that NMT had higher match
n-gram match PBMT NMT
'' RB 0.1111 0 7
PRP VBP '' RB 0.1250 0 6
VBP '' RB 0.1250 0 6
VBG TO VB PRP 0.1250 0 6
IN NN VBZ 0.1429 0 5
WP VBZ VBG 0.1429 0 5
NN WRB PRP 0.1429 0 5
PRP VBD DT NN 0.1429 0 5
NNS PRP 0.1667 0 4
PRP , CC 0.1667 0 4
TO JJ 0.1667 0 4
IN PRP VBP VBG 0.1667 0 4
WRB JJ NNS 0.1667 0 4
'' RB VBP 0.1667 0 4
WP VBZ VBG IN 0.1667 0 4
VBP '' RB VBP 0.1667 0 4
IN DT NN VBD 0.1667 0 4
NN PRP MD 0.1667 0 4
JJ NNP NN 0.1667 0 4
VBP NNS VBP TO 0.2000 0 3
DT NN , WDT 0.2000 0 3
VBD DT RB 0.2000 0 3
PRP VBD PRP$ 0.2000 0 3
RB PRP VBP VBG 0.2000 0 3
NN IN PRP VBZ 0.2000 0 3
VBN PRP 0.2000 0 3
DT NNS , PRP 0.2000 0 3
NNS IN NN . 0.2000 0 3
NNP : NNP , 0.2000 0 3
JJ , PRP 0.2000 0 3
CD , PRP 0.2000 1 7
IN PRP VBD PRP 0.2000 0 3
NNS CC NNS IN 0.2000 0 3
NNP : NNP 0.2000 0 3
VBD PRP DT 0.2000 0 3
VBD DT RB JJ 0.2000 0 3
VBG TO VB DT 0.2000 0 3
VBD VBG 0.2000 2 11
IN CD NNS IN 0.2000 0 3
IN PRP VBD DT 0.2000 0 3
NNS PRP VBP 0.2000 0 3
PRP MD RB VB 0.2000 0 3
NN VBZ PRP 0.2000 0 3
, PDT DT 0.2000 0 3
JJ , CC PRP 0.2000 0 3
PDT DT JJ 0.2000 0 3
NNS VBP VBN 0.2000 0 3
JJ NNP NN PRP 0.2000 0 3
NNP NN PRP 0.2000 0 3
VB DT NN TO 0.2000 0 3


Sentence Examples

10 sentences where PBMT>NMT at sentence-level BLEU

Output sentbleu
Src To je vzrušujúce samo o sebe.
Ref That 's exciting in itself .
PBMT That 's exciting in itself . 1.0000
NMT This is a thrill of self . 0.1858


Output sentbleu
Src A žili spolu šťastne až do smrti.
Ref And they lived together happily ever after .
PBMT And they lived together happily ever after . 1.0000
NMT And he lived with happily to death . 0.2016


Output sentbleu
Src Beth Israel je v Bostone.
Ref Beth Israel 's in Boston .
PBMT Beth Israel 's in Boston . 1.0000
NMT Beat Isaill is in Boston . 0.4111


Output sentbleu
Src (Potlesk) Posledná otázka, Julian.
Ref ( Applause ) Last question , Julian .
PBMT ( Applause ) Last question , Julian . 1.0000
NMT ( Applause ) The last question , Julian . Julan . 0.4240


Output sentbleu
Src Zistili sme, že aj bohaté deti z predmestia naozaj chcú DryBath. (smiech) Aspoň raz za týždeň.
Ref We realized that even rich kids from the suburbs really want DryBath . ( Laughter ) At least once a week .
PBMT And we found that even rich kids a suburbs really want DryBath . ( Laughter ) At least once a week . 0.7378
NMT We 've learned that the rich kids have actually wanted DryBothth , ( Laughter ) at the time for a week . 0.1786


Output sentbleu
Src JA: Nie som si istý o incidente.
Ref JA : I 'm not sure about the incident .
PBMT JA : I 'm not sure about the incident . 1.0000
NMT J : I 'm not sure of you . 0.4808


Output sentbleu
Src Toto bolo opísané nádherne v roku 2006 v knihe, ktorú napísali Michael Porter a Elizabeth Teisberg.
Ref This was described beautifully in a book in 2006 by Michael Porter and Elizabeth Teisberg .
PBMT This was described beautifully in 2006 in a book that wrote Michael Porter and Elizabeth Teisberg . 0.6254
NMT This was originally gorgeous in 2006 in 2006 in 2006 , and they wrote Michael Porrter and Edzabet Tode . 0.1171


Output sentbleu
Src Prečo je to tak?
Ref Why is that ?
PBMT Why is that ? 1.0000
NMT Why is this ? 0.5081


Output sentbleu
Src Nemyslím tým doslova 24 hodín, 7 dní v týždni.
Ref I don 't literally mean 24 hours , seven days a week .
PBMT I don 't mean literally 24 hours , seven days a week . 0.7272
NMT I don 't mean literally 24 hours in the week . 0.2525


Output sentbleu
Src Rozdiely boli dramatické.
Ref The differences were dramatic .
PBMT The differences were dramatic . 1.0000
NMT The variations were dramatic . 0.5373


10 sentences where NMT>PBMT at sentence-level BLEU

Output sentbleu
Src Drobné vychytávky môžu viesť k veľkým zmenám.
Ref Tiny tweaks can lead to big changes .
PBMT Now , this idea of marginal raisestas can lead to a big change to happen . 0.1584
NMT Facferences can lead to big changes . 0.7290


Output sentbleu
Src Takže tu je video.
Ref So here 's a video .
PBMT So here 's the video . " 0.4111
NMT So here 's a video . 1.0000


Output sentbleu
Src Nejde o tie čísla.
Ref It 's not the numbers .
PBMT It 's not about those numbers . 0.4111
NMT It 's not the numbers . 1.0000


Output sentbleu
Src Tu vám teda ukážem dva zábery.
Ref So here I 'm going to show you two tandem clips .
PBMT Here you 're going to two footage . 0.1522
NMT So here I 'm going to show you two footage . 0.7547


Output sentbleu
Src Čo to je?
Ref What is this ?
PBMT " What is that ? 0.3861
NMT What is this ? 1.0000


Output sentbleu
Src (potlesk) Je to tak.
Ref ( Applause ) That 's true .
PBMT ( Applause ) It 's so . 0.3826
NMT ( Applause ) That 's true . 1.0000


Output sentbleu
Src Svetlo je dobré.
Ref Light is good .
PBMT The light 's good . 0.3593
NMT Light is good . 1.0000


Output sentbleu
Src A to o čom hovorím je toto.
Ref And what I 'm talking about is this .
PBMT And that 's what I 'm saying is this . 0.3508
NMT And what I 'm talking about is this . 1.0000


Output sentbleu
Src Zatiaľ sa len učia počítať.
Ref They 're just learning how to count .
PBMT Meanwhile just teach to count . 0.3115
NMT They 're just learning how to count . 1.0000


Output sentbleu
Src Chlapec: Preto.
Ref Boy : That 's why .
PBMT Boy : So . 0.3082
NMT Boy : That 's why . 1.0000