Models Accuracy¶
All y-axis been distorted purposely and this can cause misrepresents data and incorrect conclusion.
Abstractive Summarization¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/summarization
Below chart is ROUGE-1.
[1]:
from IPython.core.display import Image, display
display(Image('abstractive-accuracy.png', width=500))
Full score at malaya.summarization.abstractive._t5_availability.
_t5_availability = {
'small': [
'122MB',
'ROUGE-1: 0.33854',
'ROUGE-2: 0.14588',
'ROUGE-L: 0.23528',
],
'base': [
'448MB',
'ROUGE-1: 0.34103',
'ROUGE-2: 0.14994',
'ROUGE-L: 0.23655',
],
}
Dependency parsing¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/dependency
Below chart is F1 accuracy for dependency tagging.
[2]:
display(Image('dependency-accuracy.png', width=500))
bert-base¶
arc accuracy: 0.8554239102233114
types accuracy: 0.8481064607232274
root accuracy: 0.9203253968253969
precision recall f1-score support
PAD 0.99996 1.00000 0.99998 877864
X 1.00000 0.99986 0.99993 145204
acl 0.96111 0.96190 0.96150 6037
advcl 0.94287 0.93895 0.94091 2408
advmod 0.97171 0.96904 0.97037 9464
amod 0.96283 0.94008 0.95132 8128
appos 0.97426 0.95940 0.96677 4852
aux 1.00000 0.50000 0.66667 4
case 0.98907 0.98834 0.98870 21519
cc 0.98089 0.98708 0.98397 6500
ccomp 0.95515 0.92164 0.93810 855
compound 0.95432 0.96565 0.95995 13479
compound:plur 0.96507 0.97778 0.97138 1215
conj 0.96943 0.98036 0.97486 8604
cop 0.96407 0.98531 0.97457 1906
csubj 0.92157 0.85455 0.88679 55
csubj:pass 0.93750 0.78947 0.85714 19
dep 0.95199 0.93574 0.94380 996
det 0.97043 0.96678 0.96860 8248
fixed 0.94176 0.93672 0.93923 1122
flat 0.98010 0.98217 0.98113 20755
iobj 0.87500 0.80000 0.83582 35
mark 0.94507 0.97448 0.95955 2860
nmod 0.96363 0.95912 0.96137 8121
nsubj 0.97076 0.97091 0.97083 12788
nsubj:pass 0.95192 0.96362 0.95774 3986
nummod 0.98563 0.97942 0.98251 7773
obj 0.96915 0.97071 0.96993 10551
obl 0.97549 0.97164 0.97356 11389
parataxis 0.95038 0.90415 0.92669 699
punct 0.99752 0.99773 0.99762 33438
root 0.98046 0.98124 0.98085 10073
xcomp 0.95153 0.94749 0.94951 2590
accuracy 0.99562 1243537
macro avg 0.96396 0.93822 0.94823 1243537
weighted avg 0.99562 0.99562 0.99562 1243537
tiny-bert¶
arc accuracy: 0.7189048051328787
types accuracy: 0.6942783162846734
root accuracy: 0.8860992063492065
precision recall f1-score support
PAD 0.99996 1.00000 0.99998 943088
X 0.99999 0.99981 0.99990 145797
acl 0.85006 0.80040 0.82448 6042
advcl 0.61783 0.60566 0.61169 2437
advmod 0.86865 0.86755 0.86810 9513
amod 0.82596 0.78837 0.80672 8217
appos 0.84113 0.79100 0.81530 5000
aux 0.80000 0.50000 0.61538 8
case 0.94714 0.95046 0.94879 21376
cc 0.92151 0.94487 0.93304 6349
ccomp 0.59326 0.26201 0.36349 874
compound 0.85764 0.83530 0.84632 13667
compound:plur 0.83743 0.91349 0.87381 1156
conj 0.87306 0.90624 0.88934 8500
cop 0.90592 0.93670 0.92105 1943
csubj 0.75000 0.05263 0.09836 57
csubj:pass 0.00000 0.00000 0.00000 16
dep 0.66704 0.55176 0.60395 1082
det 0.89147 0.84818 0.86929 7970
fixed 0.80819 0.61696 0.69975 1120
flat 0.90396 0.93947 0.92137 21129
iobj 0.00000 0.00000 0.00000 25
mark 0.74718 0.83845 0.79019 2767
nmod 0.86083 0.78159 0.81930 8017
nsubj 0.85174 0.89750 0.87402 12712
nsubj:pass 0.78514 0.82246 0.80337 4061
nummod 0.88943 0.93509 0.91169 8026
obj 0.89982 0.84423 0.87114 10618
obl 0.84081 0.88283 0.86131 11385
parataxis 0.48635 0.26667 0.34446 735
punct 0.98350 0.99126 0.98736 33736
root 0.91085 0.93726 0.92387 10073
xcomp 0.69305 0.76415 0.72686 2544
accuracy 0.98102 1310040
macro avg 0.77906 0.72946 0.74011 1310040
weighted avg 0.98076 0.98102 0.98073 1310040
albert-base¶
arc accuracy: 0.8118309576064845
types accuracy: 0.7931625589721538
root accuracy: 0.879281746031746
precision recall f1-score support
PAD 1.00000 1.00000 1.00000 905035
X 0.99997 0.99998 0.99998 159607
acl 0.89111 0.88994 0.89052 6051
advcl 0.75213 0.78003 0.76583 2373
advmod 0.89975 0.92642 0.91289 9378
amod 0.86607 0.87808 0.87204 8145
appos 0.87914 0.89496 0.88698 4779
aux 1.00000 0.37500 0.54545 8
case 0.96890 0.97142 0.97016 21521
cc 0.96049 0.96393 0.96221 6405
ccomp 0.70574 0.67583 0.69046 873
compound 0.88800 0.89660 0.89228 13530
compound:plur 0.93381 0.93981 0.93680 1246
conj 0.94147 0.93436 0.93790 8608
cop 0.94652 0.96651 0.95641 1941
csubj 0.75000 0.39623 0.51852 53
csubj:pass 0.77778 0.77778 0.77778 9
dep 0.81778 0.72871 0.77068 1010
det 0.91665 0.90606 0.91132 8314
fixed 0.87862 0.80565 0.84055 1168
flat 0.96177 0.93608 0.94875 20400
iobj 0.71429 0.42857 0.53571 35
mark 0.88640 0.88577 0.88608 2854
nmod 0.86857 0.90150 0.88473 8020
nsubj 0.89466 0.93382 0.91382 12633
nsubj:pass 0.91977 0.81904 0.86648 4045
nummod 0.95316 0.95864 0.95589 8003
obj 0.90795 0.92092 0.91439 10357
obl 0.93016 0.90607 0.91796 11466
parataxis 0.72669 0.62953 0.67463 718
punct 0.99482 0.99724 0.99603 33312
root 0.93869 0.94093 0.93981 10073
xcomp 0.85300 0.80468 0.82813 2524
accuracy 0.98785 1284494
macro avg 0.88860 0.84152 0.85761 1284494
weighted avg 0.98786 0.98785 0.98782 1284494
albert-tiny¶
arc accuracy: 0.7087220659183397
types accuracy: 0.6735055899028873
root accuracy: 0.8178452380952382
precision recall f1-score support
PAD 1.00000 1.00000 1.00000 901404
X 0.99997 0.99998 0.99997 158217
acl 0.74523 0.72259 0.73374 6056
advcl 0.44763 0.44416 0.44589 2319
advmod 0.80839 0.80245 0.80541 9537
amod 0.74481 0.69167 0.71726 8144
appos 0.71137 0.68084 0.69577 4963
aux 0.00000 0.00000 0.00000 9
case 0.90625 0.93745 0.92159 21056
cc 0.92435 0.90888 0.91655 6453
ccomp 0.32162 0.13918 0.19429 855
compound 0.76535 0.75323 0.75924 13008
compound:plur 0.76103 0.77066 0.76581 1186
conj 0.79454 0.78507 0.78978 8640
cop 0.87581 0.90736 0.89130 1943
csubj 0.66667 0.04082 0.07692 49
csubj:pass 0.00000 0.00000 0.00000 18
dep 0.41637 0.38321 0.39910 929
det 0.81424 0.77924 0.79636 7909
fixed 0.63932 0.41054 0.50000 1101
flat 0.85963 0.91321 0.88561 20856
iobj 1.00000 0.03333 0.06452 30
mark 0.69997 0.72039 0.71003 2879
nmod 0.71129 0.68985 0.70041 7964
nsubj 0.74144 0.81233 0.77527 12719
nsubj:pass 0.68649 0.56466 0.61964 3905
nummod 0.84427 0.87244 0.85813 7581
obj 0.79591 0.78073 0.78825 10380
obl 0.75820 0.78392 0.77085 11144
parataxis 0.25150 0.06231 0.09988 674
punct 0.98207 0.98323 0.98265 33034
root 0.84186 0.87362 0.85745 10073
xcomp 0.62652 0.63961 0.63300 2489
accuracy 0.96997 1277524
macro avg 0.70128 0.63294 0.64105 1277524
weighted avg 0.96929 0.96997 0.96946 1277524
xlnet-base¶
arc accuracy: 0.9310084738376598
types accuracy: 0.9258795751889828
root accuracy: 0.9474206349206349
precision recall f1-score support
PAD 0.99998 1.00000 0.99999 632972
X 1.00000 0.99997 0.99999 143586
acl 0.98091 0.98226 0.98158 5806
advcl 0.97098 0.95161 0.96120 2356
advmod 0.98802 0.97806 0.98302 9527
amod 0.95966 0.97100 0.96530 8208
appos 0.98846 0.98947 0.98896 4936
aux 1.00000 1.00000 1.00000 10
case 0.99454 0.99110 0.99282 21128
cc 0.98704 0.99518 0.99109 6429
ccomp 0.89091 0.97313 0.93021 856
compound 0.98091 0.96643 0.97362 13079
compound:plur 0.99068 0.98401 0.98733 1188
conj 0.98303 0.99214 0.98756 8524
cop 0.98664 0.99071 0.98867 1938
csubj 0.96000 0.96000 0.96000 50
csubj:pass 0.95652 0.91667 0.93617 24
dep 0.98182 0.96716 0.97444 1005
det 0.98698 0.97756 0.98225 8065
fixed 0.96071 0.97162 0.96613 1057
flat 0.98389 0.99064 0.98726 20411
iobj 0.96154 0.80645 0.87719 31
mark 0.96611 0.98539 0.97565 2806
nmod 0.97956 0.97285 0.97619 8030
nsubj 0.98317 0.98402 0.98359 12701
nsubj:pass 0.96930 0.97858 0.97392 3969
nummod 0.99113 0.99327 0.99220 7879
obj 0.98266 0.98076 0.98171 10342
obl 0.98468 0.98256 0.98362 11183
parataxis 0.95595 0.95455 0.95525 682
punct 0.99952 0.99949 0.99950 33107
root 0.98888 0.98888 0.98888 10073
xcomp 0.95951 0.96027 0.95989 2517
accuracy 0.99678 994475
macro avg 0.97738 0.97381 0.97531 994475
weighted avg 0.99679 0.99678 0.99678 994475
alxlnet-base¶
arc accuracy: 0.8943757029483008
types accuracy: 0.88690168487317
root accuracy: 0.9425595238095238
precision recall f1-score support
PAD 0.99999 1.00000 0.99999 644667
X 0.99998 0.99999 0.99998 144988
acl 0.95995 0.96137 0.96066 6058
advcl 0.91687 0.93839 0.92751 2386
advmod 0.97160 0.97620 0.97389 9496
amod 0.95264 0.94761 0.95012 8342
appos 0.97560 0.97638 0.97599 4995
aux 1.00000 1.00000 1.00000 6
case 0.99147 0.98685 0.98916 21680
cc 0.97523 0.99377 0.98441 6418
ccomp 0.95249 0.90112 0.92610 890
compound 0.95478 0.95656 0.95567 13399
compound:plur 0.97575 0.98067 0.97821 1190
conj 0.96575 0.98929 0.97738 8494
cop 0.98201 0.98708 0.98454 1935
csubj 1.00000 0.90476 0.95000 42
csubj:pass 0.91667 0.91667 0.91667 12
dep 0.96490 0.94781 0.95628 1073
det 0.96461 0.97375 0.96916 8230
fixed 0.95762 0.92188 0.93941 1152
flat 0.98208 0.98030 0.98119 20967
iobj 1.00000 0.82927 0.90667 41
mark 0.96463 0.95609 0.96034 2824
nmod 0.96933 0.95492 0.96207 8207
nsubj 0.97533 0.97086 0.97309 12867
nsubj:pass 0.95811 0.94145 0.94970 3911
nummod 0.98952 0.98590 0.98770 7659
obj 0.97249 0.96839 0.97044 10440
obl 0.97129 0.97222 0.97175 11483
parataxis 0.95691 0.91348 0.93469 705
punct 0.99883 0.99955 0.99919 33252
root 0.98284 0.98372 0.98328 10073
xcomp 0.92520 0.94988 0.93738 2474
accuracy 0.99475 1010356
macro avg 0.97044 0.95958 0.96462 1010356
weighted avg 0.99476 0.99475 0.99475 1010356
Emotion Analysis¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/emotion
Graph based on F1-score.
[3]:
from IPython.core.display import Image, display
display(Image('emotion-accuracy.png', width=500))
multinomial¶
precision recall f1-score support
anger 0.88832 0.90889 0.89848 5872
fear 0.89515 0.88078 0.88791 4110
happy 0.88992 0.92776 0.90845 6091
love 0.92420 0.90616 0.91509 4252
sadness 0.91943 0.87356 0.89591 5212
surprise 0.92340 0.92838 0.92588 2597
accuracy 0.90371 28134
macro avg 0.90674 0.90426 0.90529 28134
weighted avg 0.90409 0.90371 0.90366 28134
bert-base¶
precision recall f1-score support
anger 0.99712 0.99763 0.99737 5895
fear 0.99687 0.99759 0.99723 4150
happy 0.99900 0.99900 0.99900 6017
love 0.99855 0.99615 0.99735 4154
sadness 0.99793 0.99906 0.99849 5307
surprise 0.99770 0.99694 0.99732 2612
accuracy 0.99790 28135
macro avg 0.99786 0.99773 0.99779 28135
weighted avg 0.99790 0.99790 0.99790 28135
tiny-bert¶
precision recall f1-score support
anger 0.99765 0.99481 0.99623 5970
fear 0.99607 0.99656 0.99631 4068
happy 0.99671 0.99918 0.99794 6062
love 0.99758 0.99638 0.99698 4145
sadness 0.99736 0.99793 0.99764 5303
surprise 0.99614 0.99691 0.99652 2587
accuracy 0.99701 28135
macro avg 0.99692 0.99696 0.99694 28135
weighted avg 0.99702 0.99701 0.99701 28135
albert-base¶
precision recall f1-score support
anger 0.99785 0.99472 0.99628 6062
fear 0.99582 0.99926 0.99754 4056
happy 0.99866 0.99866 0.99866 5988
love 0.99712 0.99760 0.99736 4162
sadness 0.99813 0.99813 0.99813 5334
surprise 0.99685 0.99803 0.99744 2533
accuracy 0.99758 28135
macro avg 0.99740 0.99773 0.99757 28135
weighted avg 0.99758 0.99758 0.99758 28135
albert-tiny¶
precision recall f1-score support
anger 0.99396 0.98603 0.98998 6012
fear 0.99390 0.99512 0.99451 4096
happy 0.99652 0.99652 0.99652 6030
love 0.99114 0.99187 0.99150 4059
sadness 0.99121 0.99699 0.99409 5316
surprise 0.99278 0.99619 0.99448 2622
accuracy 0.99346 28135
macro avg 0.99325 0.99378 0.99351 28135
weighted avg 0.99346 0.99346 0.99346 28135
xlnet-base¶
precision recall f1-score support
anger 0.99699 0.99733 0.99716 5983
fear 0.99778 0.99827 0.99802 4045
happy 0.99883 0.99850 0.99867 6005
love 0.99718 0.99625 0.99671 4261
sadness 0.99754 0.99773 0.99764 5288
surprise 0.99804 0.99843 0.99824 2553
accuracy 0.99773 28135
macro avg 0.99773 0.99775 0.99774 28135
weighted avg 0.99773 0.99773 0.99773 28135
alxlnet-base¶
precision recall f1-score support
anger 0.99669 0.99439 0.99554 6065
fear 0.99702 0.99727 0.99714 4027
happy 0.99764 0.99949 0.99857 5918
love 0.99554 0.99694 0.99624 4250
sadness 0.99867 0.99641 0.99754 5286
surprise 0.99422 0.99730 0.99576 2589
accuracy 0.99691 28135
macro avg 0.99663 0.99697 0.99680 28135
weighted avg 0.99691 0.99691 0.99691 28135
Entities Recognition¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/entities
Graph based on F1-score.
[4]:
from IPython.core.display import Image, display
display(Image('ner-accuracy.png', width=500))
bert-base¶
precision recall f1-score support
OTHER 0.99224 0.99931 0.99576 5160854
PAD 1.00000 1.00000 1.00000 877767
X 0.99995 1.00000 0.99998 2921053
event 0.99911 0.88679 0.93961 143787
law 0.99704 0.97040 0.98354 146950
location 0.98677 0.98420 0.98548 428869
organization 0.99335 0.95355 0.97304 694150
person 0.97636 0.99476 0.98547 507960
quantity 0.99965 0.99803 0.99884 88200
time 0.98462 0.99938 0.99194 179880
accuracy 0.99406 11149470
macro avg 0.99291 0.97864 0.98537 11149470
weighted avg 0.99409 0.99406 0.99400 11149470
tiny-bert¶
precision recall f1-score support
OTHER 0.98178 0.99946 0.99054 5160854
PAD 1.00000 1.00000 1.00000 1673627
X 1.00000 1.00000 1.00000 2921053
event 0.99666 0.70215 0.82388 143787
law 0.99522 0.94921 0.97167 146950
location 0.96753 0.96547 0.96650 428869
organization 0.99403 0.87009 0.92794 694150
person 0.92771 0.99283 0.95917 507960
quantity 0.99643 0.99762 0.99703 88200
time 0.95574 0.99855 0.97668 179880
accuracy 0.98642 11945330
macro avg 0.98151 0.94754 0.96134 11945330
weighted avg 0.98675 0.98642 0.98594 11945330
albert-base¶
precision recall f1-score support
OTHER 0.98087 0.99948 0.99008 5160854
PAD 1.00000 1.00000 1.00000 881183
X 0.99996 1.00000 0.99998 2933007
event 0.99021 0.80012 0.88507 143787
law 0.96373 0.94234 0.95291 146950
location 0.97388 0.96256 0.96819 428869
organization 0.99506 0.83927 0.91055 694150
person 0.91340 0.99378 0.95189 507960
quantity 0.99636 0.99704 0.99670 88200
time 0.98911 0.99859 0.99383 179880
accuracy 0.98466 11164840
macro avg 0.98026 0.95332 0.96492 11164840
weighted avg 0.98509 0.98466 0.98421 11164840
albert-tiny¶
precision recall f1-score support
OTHER 0.96614 0.99651 0.98109 5160854
PAD 1.00000 1.00000 1.00000 881183
X 0.99984 1.00000 0.99992 2933007
event 0.97661 0.52453 0.68250 143787
law 0.97992 0.89007 0.93284 146950
location 0.92117 0.91206 0.91659 428869
organization 0.96821 0.76413 0.85414 694150
person 0.87211 0.97366 0.92009 507960
quantity 0.98545 0.99220 0.98881 88200
time 0.94056 0.98312 0.96137 179880
accuracy 0.97124 11164840
macro avg 0.96100 0.90363 0.92374 11164840
weighted avg 0.97185 0.97124 0.96965 11164840
xlnet-base¶
precision recall f1-score support
OTHER 0.98873 0.99965 0.99416 5160854
PAD 1.00000 1.00000 1.00000 877767
X 0.99999 1.00000 0.99999 2921053
event 0.99404 0.93677 0.96456 143787
law 0.99734 0.98832 0.99281 146950
location 0.99189 0.97927 0.98554 428869
organization 0.99785 0.92433 0.95968 694150
person 0.97446 0.98956 0.98195 507960
quantity 0.99861 0.99875 0.99868 88200
time 0.99153 0.99872 0.99511 179880
accuracy 0.99285 11149470
macro avg 0.99344 0.98154 0.98725 11149470
weighted avg 0.99291 0.99285 0.99276 11149470
alxlnet-base¶
precision recall f1-score support
OTHER 0.99124 0.99962 0.99541 5160854
PAD 1.00000 1.00000 1.00000 877767
X 1.00000 1.00000 1.00000 2921053
event 0.99766 0.86900 0.92890 143787
law 0.99837 0.97023 0.98410 146950
location 0.99004 0.98249 0.98625 428869
organization 0.99584 0.94088 0.96758 694150
person 0.96062 0.99571 0.97785 507960
quantity 0.99920 0.99976 0.99948 88200
time 0.98851 0.99976 0.99410 179880
accuracy 0.99319 11149470
macro avg 0.99215 0.97575 0.98337 11149470
weighted avg 0.99327 0.99319 0.99309 11149470
Language Detection¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/language-detection
Graph based on F1-score.
[5]:
display(Image('language-detection-accuracy.png', width=500))
fast-text¶
precision recall f1-score support
eng 0.94014 0.96750 0.95362 553739
ind 0.97290 0.97316 0.97303 576059
malay 0.98674 0.95262 0.96938 1800649
manglish 0.96595 0.98417 0.97498 181442
other 0.98454 0.99698 0.99072 1428083
rojak 0.81149 0.91650 0.86080 189678
accuracy 0.97002 4729650
macro avg 0.94363 0.96515 0.95375 4729650
weighted avg 0.97111 0.97002 0.97028 4729650
Deep learning¶
precision recall f1-score support
eng 0.96760 0.97401 0.97080 553739
ind 0.97635 0.96131 0.96877 576059
malay 0.96985 0.98498 0.97736 1800649
manglish 0.98036 0.96569 0.97297 181442
other 0.99641 0.99627 0.99634 1428083
rojak 0.94221 0.84302 0.88986 189678
accuracy 0.97779 4729650
macro avg 0.97213 0.95421 0.96268 4729650
weighted avg 0.97769 0.97779 0.97760 4729650
POS Recognition¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/pos
Graph based on F1-score.
[6]:
display(Image('pos-accuracy.png', width=500))
bert-base¶
precision recall f1-score support
ADJ 0.79261 0.80819 0.80033 45666
ADP 0.95551 0.96155 0.95852 119589
ADV 0.86824 0.83832 0.85302 47760
AUX 0.99362 0.99710 0.99536 10000
CCONJ 0.97639 0.92470 0.94984 37171
DET 0.93663 0.92556 0.93107 38839
NOUN 0.91335 0.89454 0.90385 268329
NUM 0.91883 0.94521 0.93183 41211
PAD 0.98980 1.00000 0.99487 147445
PART 0.91225 0.91291 0.91258 5500
PRON 0.97505 0.94047 0.95745 48835
PROPN 0.91824 0.94054 0.92926 227608
PUNCT 0.99829 0.99853 0.99841 182824
SCONJ 0.76934 0.84297 0.80447 15150
SYM 0.99711 0.95722 0.97676 3600
VERB 0.94284 0.94533 0.94408 124518
X 0.99947 0.99882 0.99914 413549
accuracy 0.95254 1777594
macro avg 0.93280 0.93129 0.93181 1777594
weighted avg 0.95272 0.95254 0.95254 1777594
tiny-bert¶
precision recall f1-score support
ADJ 0.78068 0.79622 0.78837 45666
ADP 0.95356 0.96107 0.95730 119589
ADV 0.85048 0.83499 0.84266 47760
AUX 0.99502 0.99850 0.99676 10000
CCONJ 0.96900 0.91986 0.94379 37171
DET 0.93853 0.94263 0.94058 38839
NOUN 0.89955 0.89812 0.89883 268329
NUM 0.93685 0.93740 0.93712 41211
PAD 0.99445 1.00000 0.99722 272341
PART 0.91302 0.91418 0.91360 5500
PRON 0.97478 0.93785 0.95596 48835
PROPN 0.92504 0.92239 0.92371 227608
PUNCT 0.99776 0.99815 0.99796 182824
SCONJ 0.75747 0.84376 0.79829 15150
SYM 0.95358 0.90167 0.92690 3600
VERB 0.93816 0.94470 0.94142 124518
X 0.99974 0.99879 0.99926 413549
accuracy 0.95343 1902490
macro avg 0.92810 0.92649 0.92704 1902490
weighted avg 0.95364 0.95343 0.95349 1902490
albert-base¶
precision recall f1-score support
ADJ 0.81706 0.76324 0.78923 45666
ADP 0.95181 0.96143 0.95660 119589
ADV 0.84898 0.84148 0.84521 47760
AUX 0.99502 1.00000 0.99751 10000
CCONJ 0.93370 0.94071 0.93719 37171
DET 0.93324 0.92824 0.93073 38839
NOUN 0.90102 0.89915 0.90008 268329
NUM 0.93291 0.94002 0.93645 41211
PAD 1.00000 1.00000 1.00000 147215
PART 0.91795 0.89909 0.90842 5500
PRON 0.97728 0.93198 0.95409 48835
PROPN 0.91565 0.93866 0.92701 227608
PUNCT 0.99818 0.99890 0.99854 182824
SCONJ 0.79499 0.74330 0.76828 15150
SYM 0.98485 0.90278 0.94203 3600
VERB 0.94143 0.94251 0.94197 124518
X 0.99972 0.99975 0.99973 414899
accuracy 0.95105 1778714
macro avg 0.93199 0.91948 0.92547 1778714
weighted avg 0.95085 0.95105 0.95088 1778714
albert-tiny¶
precision recall f1-score support
ADJ 0.71343 0.69192 0.70251 45666
ADP 0.94552 0.92892 0.93715 119589
ADV 0.82394 0.77969 0.80120 47760
AUX 0.99502 0.99930 0.99716 10000
CCONJ 0.95223 0.92397 0.93789 37171
DET 0.92886 0.89495 0.91159 38839
NOUN 0.85984 0.87755 0.86860 268329
NUM 0.90365 0.90240 0.90303 41211
PAD 1.00000 1.00000 1.00000 147215
PART 0.88633 0.82509 0.85461 5500
PRON 0.94693 0.93722 0.94205 48835
PROPN 0.90464 0.89602 0.90031 227608
PUNCT 0.98900 0.99757 0.99327 182824
SCONJ 0.70104 0.77234 0.73496 15150
SYM 0.94761 0.86417 0.90397 3600
VERB 0.90093 0.92448 0.91255 124518
X 0.99946 0.99954 0.99950 414899
accuracy 0.93335 1778714
macro avg 0.90579 0.89501 0.90002 1778714
weighted avg 0.93344 0.93335 0.93331 1778714
xlnet-base¶
precision recall f1-score support
ADJ 0.83194 0.77563 0.80280 45666
ADP 0.96501 0.95786 0.96142 119589
ADV 0.85073 0.84144 0.84606 47760
AUX 0.99502 0.99950 0.99726 10000
CCONJ 0.96564 0.92473 0.94474 37171
DET 0.94985 0.93192 0.94080 38839
NOUN 0.89484 0.92123 0.90784 268329
NUM 0.94009 0.94511 0.94260 41211
PAD 0.99816 1.00000 0.99908 146373
PART 0.91259 0.94345 0.92777 5500
PRON 0.96988 0.94223 0.95586 48835
PROPN 0.93581 0.92557 0.93066 227608
PUNCT 0.99831 0.99933 0.99882 182824
SCONJ 0.73907 0.82376 0.77912 15150
SYM 0.96944 0.96917 0.96930 3600
VERB 0.94517 0.94727 0.94622 124518
X 0.99992 0.99957 0.99975 410749
accuracy 0.95410 1773722
macro avg 0.93303 0.93222 0.93236 1773722
weighted avg 0.95433 0.95410 0.95411 1773722
alxlnet-base¶
precision recall f1-score support
ADJ 0.79153 0.79396 0.79275 45666
ADP 0.95941 0.96102 0.96021 119589
ADV 0.85117 0.82073 0.83567 47760
AUX 0.99641 0.99860 0.99750 10000
CCONJ 0.96687 0.92793 0.94700 37171
DET 0.91526 0.93156 0.92334 38839
NOUN 0.91155 0.89253 0.90194 268329
NUM 0.92871 0.93635 0.93252 41211
PAD 0.99816 1.00000 0.99908 146373
PART 0.91285 0.92364 0.91821 5500
PRON 0.97040 0.94404 0.95704 48835
PROPN 0.90899 0.94301 0.92569 227608
PUNCT 0.99887 0.99928 0.99908 182824
SCONJ 0.69691 0.86964 0.77375 15150
SYM 0.99941 0.94556 0.97174 3600
VERB 0.95809 0.93052 0.94411 124518
X 0.99985 0.99945 0.99965 410749
accuracy 0.95109 1773722
macro avg 0.92732 0.93046 0.92819 1773722
weighted avg 0.95168 0.95109 0.95121 1773722
Paraphrase¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/paraphrase
Graph based on BLEU score from ``tensor2tensor.utils.bleu_hook``.
[7]:
display(Image('paraphrase-accuracy.png', width=500))
Relevancy¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/relevancy
Graph based on F1-score.
[8]:
display(Image('relevancy-accuracy.png', width=500))
bert-base¶
precision recall f1-score support
not relevant 0.87625 0.73478 0.79930 5946
relevant 0.87117 0.94531 0.90673 11281
accuracy 0.87264 17227
macro avg 0.87371 0.84004 0.85302 17227
weighted avg 0.87293 0.87264 0.86965 17227
tiny-bert¶
precision recall f1-score support
not relevant 0.95455 0.00353 0.00704 5946
relevant 0.65562 0.99991 0.79197 11281
accuracy 0.65601 17227
macro avg 0.80508 0.50172 0.39950 17227
weighted avg 0.75880 0.65601 0.52104 17227
albert-base¶
precision recall f1-score support
not relevant 0.81807 0.80844 0.81323 5946
relevant 0.89966 0.90524 0.90244 11281
accuracy 0.87183 17227
macro avg 0.85886 0.85684 0.85783 17227
weighted avg 0.87150 0.87183 0.87165 17227
albert-tiny¶
precision recall f1-score support
not relevant 0.84793 0.66768 0.74708 5946
relevant 0.84249 0.93689 0.88718 11281
accuracy 0.84397 17227
macro avg 0.84521 0.80228 0.81713 17227
weighted avg 0.84437 0.84397 0.83883 17227
xlnet-base¶
precision recall f1-score support
not relevant 0.85676 0.80272 0.82886 5946
relevant 0.89937 0.92926 0.91407 11281
accuracy 0.88559 17227
macro avg 0.87806 0.86599 0.87147 17227
weighted avg 0.88466 0.88559 0.88466 17227
alxlnet-base¶
precision recall f1-score support
not relevant 0.89878 0.71678 0.79753 5946
relevant 0.86512 0.95745 0.90895 11281
accuracy 0.87438 17227
macro avg 0.88195 0.83712 0.85324 17227
weighted avg 0.87674 0.87438 0.87049 17227
Sentiment Analysis¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/sentiment
Graph based on F1-score.
[9]:
display(Image('sentiment-accuracy.png', width=500))
multinomial¶
precision recall f1-score support
negative 0.86962 0.92076 0.89446 101375
positive 0.91632 0.86274 0.88872 101959
accuracy 0.89167 203334
macro avg 0.89297 0.89175 0.89159 203334
weighted avg 0.89304 0.89167 0.89158 203334
bert-base¶
precision recall f1-score support
negative 0.99213 0.99444 0.99328 101421
positive 0.99445 0.99215 0.99330 101924
accuracy 0.99329 203345
macro avg 0.99329 0.99330 0.99329 203345
weighted avg 0.99329 0.99329 0.99329 203345
tiny-bert¶
precision recall f1-score support
negative 0.98953 0.98579 0.98766 101153
positive 0.98599 0.98968 0.98783 102192
accuracy 0.98774 203345
macro avg 0.98776 0.98774 0.98774 203345
weighted avg 0.98775 0.98774 0.98774 203345
albert-base¶
precision recall f1-score support
negative 0.99400 0.99051 0.99225 101699
positive 0.99054 0.99402 0.99228 101646
accuracy 0.99226 203345
macro avg 0.99227 0.99226 0.99226 203345
weighted avg 0.99227 0.99226 0.99226 203345
albert-tiny¶
precision recall f1-score support
negative 0.98825 0.98262 0.98543 101357
positive 0.98282 0.98839 0.98560 101988
accuracy 0.98551 203345
macro avg 0.98554 0.98550 0.98551 203345
weighted avg 0.98553 0.98551 0.98551 203345
xlnet-base¶
precision recall f1-score support
negative 0.99253 0.99451 0.99352 101430
positive 0.99452 0.99255 0.99354 101915
accuracy 0.99353 203345
macro avg 0.99353 0.99353 0.99353 203345
weighted avg 0.99353 0.99353 0.99353 203345
alxlnet-base¶
precision recall f1-score support
negative 0.99286 0.99090 0.99188 101710
positive 0.99091 0.99287 0.99189 101635
accuracy 0.99188 203345
macro avg 0.99188 0.99188 0.99188 203345
weighted avg 0.99188 0.99188 0.99188 203345
Similarity¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/similarity
Graph based on F1-score.
[10]:
display(Image('similarity-accuracy.png', width=500))
bert-base¶
precision recall f1-score support
not similar 0.91813 0.86843 0.89259 114935
similar 0.84816 0.90468 0.87551 93371
accuracy 0.88468 208306
macro avg 0.88315 0.88656 0.88405 208306
weighted avg 0.88677 0.88468 0.88493 208306
tiny-bert¶
precision recall f1-score support
not similar 0.90845 0.85704 0.88200 114843
similar 0.83576 0.89387 0.86384 93463
accuracy 0.87357 208306
macro avg 0.87210 0.87546 0.87292 208306
weighted avg 0.87583 0.87357 0.87385 208306
albert-base¶
precision recall f1-score support
not similar 0.88351 0.88549 0.88450 114523
similar 0.85978 0.85743 0.85860 93783
accuracy 0.87286 208306
macro avg 0.87164 0.87146 0.87155 208306
weighted avg 0.87283 0.87286 0.87284 208306
albert-tiny¶
precision recall f1-score support
not similar 0.84881 0.82946 0.83902 114914
similar 0.79588 0.81821 0.80689 93392
accuracy 0.82441 208306
macro avg 0.82234 0.82383 0.82295 208306
weighted avg 0.82508 0.82441 0.82461 208306
xlnet-base¶
precision recall f1-score support
not similar 0.74384 0.92845 0.82596 114854
similar 0.87347 0.60705 0.71629 93452
accuracy 0.78426 208306
macro avg 0.80866 0.76775 0.77112 208306
weighted avg 0.80200 0.78426 0.77676 208306
alxlnet-base¶
precision recall f1-score support
not similar 0.89614 0.90170 0.89891 114554
similar 0.87897 0.87231 0.87563 93752
accuracy 0.88847 208306
macro avg 0.88756 0.88700 0.88727 208306
weighted avg 0.88841 0.88847 0.88843 208306
Subjectivity Analysis¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/subjectivity
Graph based on F1-score.
[11]:
display(Image('subjectivity-accuracy.png', width=500))
multinomial¶
precision recall f1-score support
negative 0.91527 0.87238 0.89331 1003
positive 0.87657 0.91818 0.89689 990
accuracy 0.89513 1993
macro avg 0.89592 0.89528 0.89510 1993
weighted avg 0.89605 0.89513 0.89509 1993
bert-base¶
precision recall f1-score support
negative 0.87825 0.96429 0.91926 980
positive 0.96183 0.87068 0.91399 1013
accuracy 0.91671 1993
macro avg 0.92004 0.91748 0.91663 1993
weighted avg 0.92073 0.91671 0.91658 1993
tiny-bert¶
precision recall f1-score support
negative 0.95678 0.84086 0.89508 974
positive 0.86368 0.96369 0.91095 1019
accuracy 0.90366 1993
macro avg 0.91023 0.90228 0.90301 1993
weighted avg 0.90917 0.90366 0.90319 1993
albert-base¶
precision recall f1-score support
negative 0.87616 0.94006 0.90699 1001
positive 0.93471 0.86593 0.89901 992
accuracy 0.90316 1993
macro avg 0.90544 0.90299 0.90300 1993
weighted avg 0.90531 0.90316 0.90301 1993
albert-tiny¶
precision recall f1-score support
negative 0.90070 0.89184 0.89625 1017
positive 0.88844 0.89754 0.89297 976
accuracy 0.89463 1993
macro avg 0.89457 0.89469 0.89461 1993
weighted avg 0.89469 0.89463 0.89464 1993
xlnet-base¶
precision recall f1-score support
negative 0.89613 0.94616 0.92047 1003
positive 0.94218 0.88889 0.91476 990
accuracy 0.91771 1993
macro avg 0.91916 0.91753 0.91761 1993
weighted avg 0.91901 0.91771 0.91763 1993
alxlnet-base¶
precision recall f1-score support
negative 0.89258 0.92604 0.90900 987
positive 0.92466 0.89066 0.90734 1006
accuracy 0.90818 1993
macro avg 0.90862 0.90835 0.90817 1993
weighted avg 0.90877 0.90818 0.90816 1993
Toxicity Analysis¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/toxic
Graph based on F1-score.
[12]:
display(Image('toxic-accuracy.png', width=500))
multinomial¶
precision recall f1-score support
severe toxic 0.32096 0.99468 0.48532 9955
obscene 0.06031 0.68096 0.11081 2799
identity attack 0.03312 0.60086 0.06277 1393
insult 0.15655 0.69002 0.25519 12575
threat 0.00661 0.11058 0.01247 416
asian 0.00087 0.01799 0.00166 389
atheist 0.00137 0.04494 0.00266 178
bisexual 0.00052 0.08333 0.00104 24
buddhist 0.00000 0.00000 0.00000 45
christian 0.13652 0.86153 0.23570 4622
female 0.12714 0.78073 0.21867 6891
heterosexual 0.00153 0.06299 0.00299 127
indian 0.14732 0.97509 0.25597 4014
homosexual, gay or lesbian 0.04442 0.45581 0.08095 1369
intellectual or learning disability 0.00000 0.00000 0.00000 6
male 0.08106 0.58298 0.14233 4947
muslim 0.07845 0.59531 0.13863 2602
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 2
other race or ethnicity 0.00000 0.00000 0.00000 7
other religion 0.00000 0.00000 0.00000 8
other sexual orientation 0.00000 0.00000 0.00000 1
physical disability 0.00000 0.00000 0.00000 2
psychiatric or mental illness 0.00720 0.09651 0.01340 601
transgender 0.00249 0.06608 0.00481 227
malay 0.54919 0.99337 0.70733 17044
chinese 0.29545 0.99079 0.45517 8793
micro avg 0.14989 0.82799 0.25383 79037
macro avg 0.07597 0.35869 0.11807 79037
weighted avg 0.25444 0.82799 0.37086 79037
samples avg 0.07772 0.16003 0.09295 79037
bert-base¶
precision recall f1-score support
severe toxic 0.85194 0.65179 0.73854 9790
obscene 0.63710 0.41623 0.50351 2847
identity attack 0.63238 0.29603 0.40328 1412
insult 0.71381 0.56111 0.62832 12673
threat 0.56707 0.22850 0.32574 407
asian 0.54394 0.56965 0.55650 402
atheist 0.80097 0.96491 0.87533 171
bisexual 1.00000 0.51852 0.68293 27
buddhist 0.60938 0.90698 0.72897 43
christian 0.86376 0.86044 0.86210 4679
female 0.88242 0.90816 0.89510 6925
heterosexual 0.67073 0.81481 0.73579 135
indian 0.95325 0.88580 0.91829 4028
homosexual, gay or lesbian 0.88355 0.92161 0.90218 1416
intellectual or learning disability 0.00000 0.00000 0.00000 6
male 0.75975 0.59414 0.66682 5019
muslim 0.87416 0.89385 0.88390 2619
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 0
other race or ethnicity 0.00000 0.00000 0.00000 11
other religion 0.14286 0.09091 0.11111 11
other sexual orientation 0.00000 0.00000 0.00000 0
physical disability 0.00000 0.00000 0.00000 6
psychiatric or mental illness 0.60000 0.81588 0.69148 592
transgender 0.79012 0.87671 0.83117 219
malay 0.96219 0.96486 0.96352 16987
chinese 0.94062 0.90214 0.92098 8727
micro avg 0.86098 0.77313 0.81469 79152
macro avg 0.58074 0.54233 0.54909 79152
weighted avg 0.84966 0.77313 0.80502 79152
samples avg 0.15924 0.15441 0.15445 79152
tiny-bert¶
precision recall f1-score support
severe toxic 0.77495 0.77346 0.77421 9857
obscene 0.62343 0.41033 0.49492 2788
identity attack 0.55057 0.34761 0.42616 1378
insult 0.69412 0.56324 0.62187 12659
threat 0.60825 0.13170 0.21651 448
asian 0.66667 0.47478 0.55459 337
atheist 0.85784 0.92593 0.89059 189
bisexual 1.00000 0.05263 0.10000 19
buddhist 0.63043 0.67442 0.65169 43
christian 0.79541 0.89441 0.84201 4612
female 0.85257 0.92515 0.88738 6907
heterosexual 0.67785 0.78295 0.72662 129
indian 0.94898 0.87673 0.91143 3967
homosexual, gay or lesbian 0.88188 0.92275 0.90185 1424
intellectual or learning disability 0.00000 0.00000 0.00000 5
male 0.70644 0.64640 0.67509 4918
muslim 0.81178 0.94261 0.87232 2544
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 0
other race or ethnicity 0.00000 0.00000 0.00000 7
other religion 0.00000 0.00000 0.00000 9
other sexual orientation 0.00000 0.00000 0.00000 2
physical disability 0.00000 0.00000 0.00000 4
psychiatric or mental illness 0.67727 0.76410 0.71807 585
transgender 0.80090 0.84689 0.82326 209
malay 0.95652 0.97334 0.96486 16839
chinese 0.96350 0.88984 0.92521 8869
micro avg 0.83535 0.79611 0.81526 78748
macro avg 0.57331 0.51182 0.51773 78748
weighted avg 0.82603 0.79611 0.80692 78748
samples avg 0.15765 0.15682 0.15490 78748
albert-base¶
precision recall f1-score support
severe toxic 0.79715 0.71003 0.75107 9863
obscene 0.64770 0.38489 0.48285 2780
identity attack 0.65517 0.27496 0.38736 1382
insult 0.73404 0.49344 0.59016 12652
threat 0.68478 0.14754 0.24277 427
asian 0.67557 0.48361 0.56369 366
atheist 0.85149 0.91489 0.88205 188
bisexual 0.93750 0.62500 0.75000 24
buddhist 0.55556 0.33333 0.41667 45
christian 0.84738 0.87439 0.86068 4737
female 0.88191 0.91253 0.89696 6997
heterosexual 0.76812 0.76812 0.76812 138
indian 0.92663 0.91164 0.91907 4142
homosexual, gay or lesbian 0.89547 0.92446 0.90973 1390
intellectual or learning disability 0.00000 0.00000 0.00000 7
male 0.73157 0.61368 0.66746 5014
muslim 0.86958 0.87620 0.87288 2496
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 1
other race or ethnicity 0.00000 0.00000 0.00000 11
other religion 0.00000 0.00000 0.00000 9
other sexual orientation 0.00000 0.00000 0.00000 1
physical disability 0.00000 0.00000 0.00000 1
psychiatric or mental illness 0.65781 0.72131 0.68810 549
transgender 0.76995 0.84536 0.80590 194
malay 0.98510 0.94072 0.96240 16869
chinese 0.90845 0.95077 0.92913 8694
micro avg 0.86054 0.76973 0.81261 78977
macro avg 0.58448 0.50766 0.53137 78977
weighted avg 0.84634 0.76973 0.79982 78977
samples avg 0.15569 0.15257 0.15179 78977
albert-tiny¶
precision recall f1-score support
severe toxic 0.78533 0.72620 0.75460 9788
obscene 0.67641 0.33796 0.45072 2808
identity attack 0.66042 0.22988 0.34104 1379
insult 0.74085 0.47457 0.57854 12662
threat 0.52941 0.02153 0.04138 418
asian 0.65027 0.29975 0.41034 397
atheist 0.85882 0.82022 0.83908 178
bisexual 1.00000 0.03125 0.06061 32
buddhist 0.73333 0.26190 0.38596 42
christian 0.87017 0.84438 0.85708 4723
female 0.85865 0.92302 0.88967 6963
heterosexual 0.76147 0.70339 0.73128 118
indian 0.93209 0.90115 0.91636 4097
homosexual, gay or lesbian 0.89625 0.89690 0.89658 1387
intellectual or learning disability 0.00000 0.00000 0.00000 7
male 0.68679 0.62619 0.65509 4941
muslim 0.86187 0.87102 0.86642 2543
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 0
other race or ethnicity 0.00000 0.00000 0.00000 8
other religion 0.00000 0.00000 0.00000 8
other sexual orientation 0.00000 0.00000 0.00000 0
physical disability 0.00000 0.00000 0.00000 1
psychiatric or mental illness 0.74208 0.57243 0.64631 573
transgender 0.79327 0.76037 0.77647 217
malay 0.99392 0.93774 0.96501 16896
chinese 0.89948 0.96317 0.93024 8770
micro avg 0.85977 0.76240 0.80816 78956
macro avg 0.59003 0.45196 0.48121 78956
weighted avg 0.84448 0.76240 0.79239 78956
samples avg 0.15465 0.15102 0.15056 78956
xlnet-base¶
precision recall f1-score support
severe toxic 0.76274 0.78363 0.77305 10006
obscene 0.50862 0.52366 0.51603 2874
identity attack 0.40349 0.52707 0.45707 1404
insult 0.58435 0.70709 0.63989 12717
threat 0.29885 0.46547 0.36400 391
asian 0.41160 0.74425 0.53005 391
atheist 0.78571 0.96175 0.86486 183
bisexual 0.54545 0.72000 0.62069 25
buddhist 0.54054 0.80000 0.64516 50
christian 0.73638 0.92561 0.82022 4584
female 0.87304 0.92314 0.89739 6935
heterosexual 0.70130 0.81818 0.75524 132
indian 0.92564 0.91477 0.92018 4001
homosexual, gay or lesbian 0.84066 0.93236 0.88414 1375
intellectual or learning disability 0.10526 0.50000 0.17391 4
male 0.71216 0.65484 0.68230 5044
muslim 0.83993 0.92537 0.88058 2546
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 0
other race or ethnicity 0.00000 0.00000 0.00000 9
other religion 0.15625 0.71429 0.25641 7
other sexual orientation 0.00000 0.00000 0.00000 0
physical disability 0.05556 0.33333 0.09524 3
psychiatric or mental illness 0.57323 0.86678 0.69008 578
transgender 0.76557 0.90086 0.82772 232
malay 0.95376 0.97807 0.96576 17103
chinese 0.94832 0.90540 0.92636 8837
micro avg 0.77904 0.83829 0.80758 79431
macro avg 0.51957 0.64911 0.56246 79431
weighted avg 0.79150 0.83829 0.81220 79431
samples avg 0.16354 0.16681 0.16247 79431
alxlnet-base¶
precision recall f1-score support
severe toxic 0.81360 0.70005 0.75257 9795
obscene 0.55653 0.49517 0.52406 2694
identity attack 0.50581 0.34938 0.41329 1371
insult 0.68040 0.60330 0.63953 12672
threat 0.40360 0.35123 0.37560 447
asian 0.66192 0.49077 0.56364 379
atheist 0.87151 0.92308 0.89655 169
bisexual 0.87500 0.60870 0.71795 23
buddhist 0.60417 0.63043 0.61702 46
christian 0.88051 0.84715 0.86351 4619
female 0.86958 0.92578 0.89680 6979
heterosexual 0.75000 0.82051 0.78367 117
indian 0.96407 0.87242 0.91596 4029
homosexual, gay or lesbian 0.88719 0.93265 0.90935 1366
intellectual or learning disability 0.21429 0.50000 0.30000 6
male 0.72205 0.65407 0.68638 4897
muslim 0.82114 0.94992 0.88085 2576
other disability 0.00000 0.00000 0.00000 0
other gender 0.00000 0.00000 0.00000 1
other race or ethnicity 0.00000 0.00000 0.00000 9
other religion 0.00000 0.00000 0.00000 6
other sexual orientation 0.00000 0.00000 0.00000 2
physical disability 0.00000 0.00000 0.00000 1
psychiatric or mental illness 0.57044 0.87589 0.69091 564
transgender 0.71756 0.88679 0.79325 212
malay 0.95141 0.97619 0.96364 17051
chinese 0.92615 0.92417 0.92516 8888
micro avg 0.83376 0.80221 0.81768 78919
macro avg 0.56470 0.56732 0.55962 78919
weighted avg 0.82757 0.80221 0.81282 78919
samples avg 0.16116 0.15980 0.15799 78919
Translation EN-MS¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/translation/en-ms
Graph based on BLEU score from ``tensor2tensor.utils.bleu_hook``.
[13]:
display(Image('enms-accuracy.png', width=500))
Translation MS-EN¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/translation/ms-en
Graph based on BLEU score from ``tensor2tensor.utils.bleu_hook``.
[14]:
display(Image('msen-accuracy.png', width=500))
Constituency Parsing¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/constituency
Graph based on FScore.
[15]:
display(Image('constituency-accuracy.png', width=500))
Entities Recognition Ontonotes5¶
Trained on 80% of dataset, tested on 20% of dataset. Link to download dataset available inside the notebooks. All training sessions stored in session/entities-ontonotes5
Graph based on F1-score.
[16]:
from IPython.core.display import Image, display
display(Image('ner-ontonotes5-accuracy.png', width=500))
bert-base¶
precision recall f1-score support
ADDRESS 0.99858 0.99974 0.99916 93446
CARDINAL 0.93840 0.90631 0.92207 48255
DATE 0.95490 0.93656 0.94564 126548
EVENT 0.92876 0.93591 0.93232 5711
FAC 0.93271 0.92658 0.92964 27392
GPE 0.93437 0.94852 0.94139 101357
LANGUAGE 0.93478 0.96389 0.94911 803
LAW 0.94824 0.95744 0.95281 24834
LOC 0.94148 0.93213 0.93678 34538
MONEY 0.87803 0.87563 0.87683 30032
NORP 0.95516 0.90446 0.92912 57014
ORDINAL 0.91510 0.91083 0.91296 6213
ORG 0.92453 0.95354 0.93881 219533
OTHER 0.99135 0.99308 0.99221 3553350
PAD 0.99956 1.00000 0.99978 1292421
PERCENT 0.96287 0.96814 0.96550 21722
PERSON 0.97376 0.93891 0.95602 101981
PRODUCT 0.87537 0.81769 0.84555 11124
QUANTITY 0.94385 0.92483 0.93424 11614
TIME 0.91912 0.90170 0.91033 9502
WORK_OF_ART 0.93126 0.81978 0.87197 13800
X 0.99906 0.99792 0.99849 1350434
accuracy 0.98821 7141624
macro avg 0.94460 0.93244 0.93822 7141624
weighted avg 0.98821 0.98821 0.98818 7141624
tiny-bert¶
precision recall f1-score support
ADDRESS 0.99501 0.99981 0.99740 93446
CARDINAL 0.93442 0.87581 0.90416 48255
DATE 0.93723 0.92710 0.93214 126548
EVENT 0.78758 0.93942 0.85682 5711
FAC 0.91859 0.91403 0.91630 27392
GPE 0.92833 0.93455 0.93143 101357
LANGUAGE 0.90220 0.81569 0.85677 803
LAW 0.92771 0.95289 0.94013 24834
LOC 0.92497 0.91983 0.92239 34538
MONEY 0.84986 0.85362 0.85174 30032
NORP 0.93555 0.89741 0.91608 57014
ORDINAL 0.86050 0.92435 0.89129 6213
ORG 0.93290 0.93551 0.93420 219533
OTHER 0.99018 0.99121 0.99070 3553350
PAD 0.99956 1.00000 0.99978 1292421
PERCENT 0.95852 0.96165 0.96008 21722
PERSON 0.93958 0.95846 0.94893 101981
PRODUCT 0.86273 0.77742 0.81786 11124
QUANTITY 0.90690 0.90839 0.90764 11614
TIME 0.89077 0.89339 0.89208 9502
WORK_OF_ART 0.83798 0.78145 0.80873 13800
X 0.99872 0.99767 0.99819 1350434
accuracy 0.98592 7141624
macro avg 0.91908 0.91635 0.91704 7141624
weighted avg 0.98590 0.98592 0.98589 7141624
albert-base¶
precision recall f1-score support
ADDRESS 0.99832 0.99969 0.99901 93446
CARDINAL 0.93291 0.89046 0.91119 48255
DATE 0.94941 0.93032 0.93977 126548
EVENT 0.89330 0.92506 0.90890 5711
FAC 0.92540 0.91257 0.91894 27392
GPE 0.93484 0.93404 0.93444 101357
LANGUAGE 0.87207 0.92528 0.89789 803
LAW 0.95567 0.95140 0.95353 24834
LOC 0.91754 0.92362 0.92057 34538
MONEY 0.85349 0.87696 0.86507 30032
NORP 0.91698 0.91416 0.91557 57014
ORDINAL 0.89159 0.93320 0.91192 6213
ORG 0.95070 0.92537 0.93786 219533
OTHER 0.99000 0.99315 0.99157 3553350
PAD 1.00000 1.00000 1.00000 1291289
PERCENT 0.96746 0.95953 0.96348 21722
PERSON 0.93748 0.96710 0.95206 101981
PRODUCT 0.84749 0.77832 0.81143 11124
QUANTITY 0.92798 0.92079 0.92437 11614
TIME 0.91058 0.88413 0.89716 9502
WORK_OF_ART 0.88983 0.77196 0.82671 13800
X 0.99917 0.99799 0.99858 1351758
accuracy 0.98714 7141816
macro avg 0.93010 0.92341 0.92636 7141816
weighted avg 0.98707 0.98714 0.98707 7141816
albert-tiny¶
precision recall f1-score support
ADDRESS 0.99322 0.99706 0.99513 93446
CARDINAL 0.89708 0.84719 0.87142 48255
DATE 0.92805 0.90984 0.91886 126548
EVENT 0.82915 0.90247 0.86426 5711
FAC 0.93366 0.84317 0.88611 27392
GPE 0.90096 0.91216 0.90653 101357
LANGUAGE 0.87336 0.74720 0.80537 803
LAW 0.91511 0.91677 0.91594 24834
LOC 0.90699 0.89105 0.89895 34538
MONEY 0.83930 0.84377 0.84152 30032
NORP 0.88694 0.85367 0.86999 57014
ORDINAL 0.82854 0.88041 0.85369 6213
ORG 0.91629 0.89916 0.90764 219533
OTHER 0.98521 0.98937 0.98728 3553350
PAD 1.00000 1.00000 1.00000 1291289
PERCENT 0.95564 0.95898 0.95731 21722
PERSON 0.90445 0.93733 0.92060 101981
PRODUCT 0.79770 0.72456 0.75937 11124
QUANTITY 0.88216 0.86310 0.87252 11614
TIME 0.84045 0.86203 0.85110 9502
WORK_OF_ART 0.85491 0.64130 0.73286 13800
X 0.99631 0.99466 0.99549 1351758
accuracy 0.98068 7141816
macro avg 0.90298 0.88251 0.89145 7141816
weighted avg 0.98053 0.98068 0.98054 7141816
xlnet-base¶
precision recall f1-score support
ADDRESS 0.99908 0.99993 0.99950 93446
CARDINAL 0.93228 0.92861 0.93044 48255
DATE 0.95220 0.95546 0.95383 126548
EVENT 0.90646 0.95535 0.93026 5711
FAC 0.94217 0.93432 0.93823 27392
GPE 0.95861 0.94860 0.95358 101357
LANGUAGE 0.91076 0.99128 0.94931 803
LAW 0.93475 0.96392 0.94911 24834
LOC 0.92387 0.94305 0.93336 34538
MONEY 0.85448 0.93027 0.89077 30032
NORP 0.95467 0.92540 0.93981 57014
ORDINAL 0.89995 0.95847 0.92829 6213
ORG 0.94905 0.95571 0.95237 219533
OTHER 0.99394 0.99254 0.99324 3553350
PAD 0.99992 1.00000 0.99996 1292031
PERCENT 0.97215 0.96423 0.96817 21722
PERSON 0.96138 0.97204 0.96668 101981
PRODUCT 0.88197 0.83028 0.85534 11124
QUANTITY 0.93301 0.95695 0.94483 11614
TIME 0.90852 0.91454 0.91152 9502
WORK_OF_ART 0.87106 0.88457 0.87776 13800
X 0.99879 0.99898 0.99889 1349384
accuracy 0.98994 7140184
macro avg 0.93814 0.95021 0.94388 7140184
weighted avg 0.99001 0.98994 0.98996 7140184
alxlnet-base¶
precision recall f1-score support
ADDRESS 0.99949 0.99981 0.99965 93446
CARDINAL 0.92765 0.91402 0.92078 48255
DATE 0.95309 0.93386 0.94338 126548
EVENT 0.88426 0.93241 0.90770 5711
FAC 0.92367 0.92991 0.92678 27392
GPE 0.93880 0.95315 0.94592 101357
LANGUAGE 0.84296 0.90909 0.87478 803
LAW 0.95472 0.95091 0.95281 24834
LOC 0.92551 0.92953 0.92751 34538
MONEY 0.86719 0.87403 0.87060 30032
NORP 0.95470 0.89518 0.92398 57014
ORDINAL 0.86582 0.94721 0.90469 6213
ORG 0.95300 0.93138 0.94206 219533
OTHER 0.99080 0.99334 0.99206 3553350
PAD 0.99992 1.00000 0.99996 1292031
PERCENT 0.96856 0.96851 0.96853 21722
PERSON 0.94616 0.96716 0.95655 101981
PRODUCT 0.87820 0.79333 0.83361 11124
QUANTITY 0.94752 0.91872 0.93290 11614
TIME 0.90322 0.90949 0.90635 9502
WORK_OF_ART 0.88971 0.79732 0.84098 13800
X 0.99883 0.99891 0.99887 1349384
accuracy 0.98816 7140184
macro avg 0.93244 0.92942 0.93047 7140184
weighted avg 0.98810 0.98816 0.98810 7140184
[ ]: