compare-mt Analysis Report

Aggregate Scores

SacreBleuScorer
Borderline Facebook-AI UF Online-A Online-W VolcTrans-AT
SacreBleuScorer 34.92 34.00 33.80 33.97 32.90 34.44


SacreBleuScorer Wins
v s1 / s2 -> Facebook-AI UF Online-A Online-W VolcTrans-AT
Borderline - (p=0.05) - (p=0.01) - (p=0.02) s1>s2 (p=0.00) - (p=0.14)
Facebook-AI - (p=0.36) - (p=0.47) - (p=0.03) - (p=0.24)
UF - (p=0.38) - (p=0.07) - (p=0.15)
Online-A - (p=0.03) - (p=0.23)
Online-W s2>s1 (p=0.01)


Score Comparison

Word Accuracies

word fmeas by frequency bucket
frequency Borderline Facebook-AI UF Online-A Online-W VolcTrans-AT Examples
<1 0.00 0.00 0.00 0.00 0.00 0.00 Examples
1 0.55 0.55 0.53 0.54 0.54 0.54 Examples
2 0.60 0.61 0.60 0.60 0.59 0.61 Examples
3 0.61 0.61 0.62 0.61 0.60 0.60 Examples
4 0.64 0.62 0.64 0.63 0.62 0.61 Examples
[5,10) 0.66 0.66 0.65 0.66 0.65 0.66 Examples
[10,100) 0.63 0.62 0.61 0.62 0.62 0.61 Examples
[100,1000) 0.73 0.73 0.72 0.73 0.71 0.72 Examples
>=1000 0.84 0.84 0.83 0.84 0.83 0.83 Examples


Word Accuracy Analysis

Sentence Buckets

bucket type: length, statistic type: BLEU
length Borderline Facebook-AI UF Online-A Online-W VolcTrans-AT
<10 27.37 28.97 26.74 28.40 27.25 28.05
[10,20) 30.88 29.87 29.84 30.23 29.33 30.43
[20,30) 32.40 31.37 31.07 31.35 29.84 30.97
[30,40) 28.38 28.59 27.67 28.55 28.47 28.52
[40,50) 34.84 31.27 33.06 30.12 29.49 28.65
[50,60) 30.77 31.67 25.59 35.72 30.74 32.00
>=60 27.63 28.90 26.31 29.50 32.90 25.58


Sentence Bucket Analysis

bucket type: len(output)-len(reference), statistic type: count
lengthdiff Borderline Facebook-AI UF Online-A Online-W VolcTrans-AT
<-20 0 0 0 0 0 0
[-20,-10) 2 3 3 2 0 2
[-10,-5) 22 26 25 30 25 30
-5 19 21 22 19 21 26
-4 35 37 38 34 45 37
-3 66 73 76 71 76 67
-2 114 120 119 124 119 125
-1 171 182 186 165 170 165
0 225 219 208 228 211 223
1 153 136 145 150 139 149
2 97 91 90 92 95 85
3 48 50 47 46 50 47
4 28 19 21 20 26 20
5 9 10 12 9 8 8
[6,11) 9 12 7 9 14 15
[11,21) 2 1 1 1 1 1
>=21 0 0 0 0 0 0


Sentence Bucket Analysis

bucket type: sentence-level BLEU, statistic type: count
sentbleu Borderline Facebook-AI UF Online-A Online-W VolcTrans-AT
<10.0 27 20 24 22 26 28
[10.0,20.0) 220 238 241 223 236 242
[20.0,30.0) 227 236 243 251 250 230
[30.0,40.0) 207 208 202 207 218 201
[40.0,50.0) 153 132 136 135 116 130
[50.0,60.0) 82 85 74 76 77 76
[60.0,70.0) 38 40 39 43 37 50
[70.0,80.0) 24 20 21 21 20 25
[80.0,90.0) 12 7 7 10 9 5
>=90.0 10 14 13 12 11 13


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 where Borderline>Facebook-AI in match
n-gram match Borderline Facebook-AI
demonstrators 0.80 3 0
the night 0.80 3 0
That 0.80 3 0
+ 0.80 3 0
in the night 0.80 3 0
discussed 0.80 3 0
theater 0.80 3 0
Naples 0.80 3 0
FC Chelsea 0.80 3 0
That is 0.75 2 0
unclear. 0.75 2 0
department 0.75 2 0
on Saturday in 0.75 2 0
- the 0.75 2 0
was on 0.75 2 0
lying 0.75 2 0
Interior Horst 0.75 2 0
the Interior Horst 0.75 2 0
the citizens 0.75 2 0
who have 0.75 2 0
Because of the 0.75 2 0
persons 0.75 2 0
the Interior Horst Seehofer 0.75 2 0
blues rock guitarist 0.75 2 0
fast and 0.75 2 0
Interior Horst Seehofer 0.75 2 0
fire department 0.75 2 0
quarantine. 0.75 2 0
Federal Minister 0.75 2 0
of the Interior Horst 0.75 2 0
the Interior 0.75 2 0
blues rock 0.75 2 0
of the Interior 0.75 2 0
crash of 0.75 2 0
travel 0.75 2 0
Federal Minister of 0.75 2 0
among 0.75 2 0
fight against the 0.75 2 0
Because of 0.71 4 1
crash 0.71 4 1
Because 0.71 4 1
the awarding of 0.67 1 0
a holding 0.67 1 0
to drive in 0.67 1 0
have been people who 0.67 1 0
club of 0.67 1 0
Wanderers. 0.67 1 0
and rolled over 0.67 1 0
Exactly 0.67 1 0
winner FC Bayern with 0.67 1 0


50 n-grams where Facebook-AI>Borderline in match
n-gram match Borderline Facebook-AI
50+1 0.17 0 4
Festival 0.20 0 3
the 50+1 0.20 0 3
landfall 0.25 0 2
lead through a 0.25 0 2
roller coaster 0.25 0 2
coaster 0.25 0 2
do 0.25 0 2
between two groups 0.25 0 2
the 50+1 rule. 0.25 0 2
roller 0.25 0 2
towards 0.25 0 2
lead through 0.25 0 2
Covid-19 0.25 0 2
being 0.25 0 2
up in 0.25 0 2
1:1 0.25 0 2
Frankfurt/Main 0.25 0 2
two groups of 0.25 0 2
Also 0.25 0 2
groups of 0.25 0 2
the lead through 0.25 0 2
in Austria. 0.25 0 2
a roller coaster 0.25 0 2
50+1 rule. 0.25 0 2
Thole/Wickler 0.25 0 2
for at 0.25 0 2
Covid-19. 0.25 0 2
the lead through a 0.25 0 2
passenger 0.25 0 2
made landfall 0.25 0 2
between two groups of 0.25 0 2
a roller 0.25 0 2
two groups 0.25 0 2
Donnersdorf/Steinsfeld 0.33 0 1
from the economy (construction, 0.33 0 1
held by 0.33 0 1
26, 2020, 8:00 0.33 0 1
the already 0.33 0 1
routing through the German 0.33 0 1
is held 0.33 0 1
been expelled 0.33 0 1
the already existing 0.33 0 1
curfew at 1 0.33 0 1
the short lead 0.33 0 1
the curfew at 0.33 0 1
Late-night 0.33 0 1
had now returned 0.33 0 1
were the third 0.33 0 1
a relatively strict 0.33 0 1


Sentence Examples

10 sentences where Borderline>Facebook-AI at sentence-level BLEU

Output sentbleu
Ref Biker reacts lightning fast and prevents worse.
Borderline Biker reacts lightning fast and prevents worse 83.66
Facebook-AI Biker reacts with lightning speed and prevents worse 27.33


Output sentbleu
Ref This decision would neither weaken the federal party nor its right wing.
Borderline This decision would neither weaken the federal party nor the right wing. 78.97
Facebook-AI Neither the federal party nor the right wing will be weakened by this decision. 25.03


Output sentbleu
Ref Werth’s winning horse Quantaz belongs to Max-Theurer.
Borderline Werth's winning horse Quantaz belongs to Max-Theurer. 83.66
Facebook-AI Werth's winning horse Quantaz is owned by Max-Theurer. 30.74


Output sentbleu
Ref New apartments are empty
Borderline New apartments are empty 100.00
Facebook-AI New apartments stand empty 50.00


Output sentbleu
Ref In the middle of the Blues Boom, Green founded Fleetwood Mac in the summer of 1967 together with his Bluesbreakers colleagues Mick Fleetwood and John McVie.
Borderline In the middle of the blues boom, Green founded Fleetwood Mac in the summer of 1967 together with his Bluesbreakers colleagues Mick Fleetwood and John McVie. 85.84
Facebook-AI In the midst of the blues boom, Green formed Fleetwood Mac in the summer of 1967 with fellow Bluesbreakers Mick Fleetwood and John McVie. 44.23


Output sentbleu
Ref The blues rock guitarist suffered from drug abuse and schizophrenia, as biographers report.
Borderline The blues rock guitarist suffered from drug abuse and schizophrenia, as biographers reported. 91.80
Facebook-AI The blues-rock guitarist suffered from drug abuse and schizophrenia, according to biographers. 52.50


Output sentbleu
Ref The numbers are still inconspicuous, said Terzenbach.
Borderline The figures are still inconspicuous, said Terzenbach. 70.35
Facebook-AI The figures are still unremarkable, said Terzenbach. 31.78


Output sentbleu
Ref In order to free the unfortunate driver, the fire department had to cut open the roof of the car.
Borderline In order to save the victims, the fire department had to cut open the roof of the car. 70.30
Facebook-AI In order to save the victim, the fire brigade had to cut the roof of the car free. 32.11


Output sentbleu
Ref From July 31 on the 33 super sports cars will race through the Lausitz.
Borderline From 31 July, the 33 super sports cars will race through Lusatia. 57.45
Facebook-AI From 31 July, the 33 supercars will race through Lusatia. 19.91


Output sentbleu
Ref Only 104 drivers started the final stage in the evening.
Borderline Only 104 riders had entered the final stage in the evening. 56.63
Facebook-AI Only 104 riders were left in the evening for the final stage. 19.73


10 sentences where Facebook-AI>Borderline at sentence-level BLEU

Output sentbleu
Ref The person had defected to South Korea three years ago and had now returned across the borders.
Borderline The person fled to South Korea three years ago and has now returned across the border. 56.15
Facebook-AI The person had defected to South Korea three years ago and had now returned across the border. 93.83


Output sentbleu
Ref “We all liked to do that in our youth,” he said to the Ö1 radio broadcaster.
Borderline "We all enjoyed doing this in our youth," he told the radio station Ö1. 12.20
Facebook-AI "We all liked to do that in our youth," he told the Ö1 radio station. 50.46


Output sentbleu
Ref Report: Seehofer pushes for longer telecommunications data retention
Borderline Report: Seehofer urges longer data retention 27.22
Facebook-AI Report: Seehofer pushes for longer data retention 66.33


Output sentbleu
Ref Baden-Württemberg: More than 250 illegal car races reported
Borderline Baden-Württemberg: Over 250 illegal car races recorded 46.34
Facebook-AI Baden-Württemberg: More than 250 illegal car races recorded 85.99


Output sentbleu
Ref There was no danger of a passing train.
Borderline Danger by an incoming train did not exist. 0
Facebook-AI There was no danger from an incoming train. 45.97


Output sentbleu
Ref The record is held by Manchester City with 100 points in the season 2017/18.
Borderline Manchester City holds the record with 100 points in the 2017 / 18 season. 37.05
Facebook-AI The record is held by Manchester City with 100 points in the 2017/18 season. 84.84


Output sentbleu
Ref A security service monitors the curfew.
Borderline A security service monitors the ban on going out. 50.71
Facebook-AI A security service monitors the curfew. 100.00


Output sentbleu
Ref Climate activist Greta Thunberg was a guest via video connection in a popular US late-night show.
Borderline Klimaatkivistin Greta Thunberg war per Videoschalte zu Gast in einer beliebten US-Latenacht-Show. 10.00
Facebook-AI Climate activist Greta Thunberg was a guest on a popular US late-night show via video link. 59.78


Output sentbleu
Ref FC Bayern Munich: Pavard injured
Borderline FC Bayern München: Pavard injured 44.72
Facebook-AI FC Bayern Munich: Pavard injured 100.00


Output sentbleu
Ref 700 dead from monsoon rains in Southeast Asia
Borderline 700 Deaths from Monsoon Rainfall in Southeast Asia 32.50
Facebook-AI 700 dead from monsoon rains in Southeast Asia 100.00