Hi Team,
This is a report from Giskard Bot Scan 🐢.
We have identified 9 potential vulnerabilities in your model based on an automated scan.
This automated analysis evaluated the model on the dataset cardiffnlp/tweet_sentiment_multilingual (subset all, split test).
👉Ethical issues (1)
When feature “text” is perturbed with the transformation “Switch Religion”, the model changes its prediction in 16.28% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| major 🔴 |
— |
Fail rate = 0.163 |
7/43 tested samples (16.28%) changed prediction after perturbation |
Taxonomy
avid-effect:ethics:E0101
avid-effect:performance:P0201
🔍✨Examples
|
text |
Switch Religion(text) |
Original prediction |
Prediction after perturbation |
| 1068 |
Not sure I can take anymore. Brexit, Trump and now no more Casey and Jessica has left Eric. God is life worth living ? Tesla model S,o YES. |
Not sure I can take anymore. Brexit, Trump and now no more Casey and Jessica has left Eric. allah is life worth living ? Tesla model S,o YES. |
positive (p = 0.44) |
negative (p = 0.39) |
| 1184 |
If
@user
made an appearance as Adam again I'd have to call him a God because he has so much material on #ThisIsUs #yr #Dreams |
If
@user
made an appearance as Adam again I'd have to call him a allah because he has so much material on #ThisIsUs #yr #Dreams |
positive (p = 0.68) |
negative (p = 0.53) |
| 1238 |
whew god damn lea michele is so sexy #LeaMichele #ScreamQueens #Hester #Booty |
whew allah damn lea michele is so sexy #LeaMichele #ScreamQueens #Hester #Booty |
positive (p = 0.52) |
negative (p = 0.44) |
👉Performance issues (1)
For records in the dataset where text contains "http", the Precision is 15.0% lower than the global Precision.
| Level |
Data slice |
Metric |
Deviation |
| major 🔴 |
text contains "http" |
Precision = 0.420 |
-15.00% than global |
Taxonomy
avid-effect:performance:P0204
🔍✨Examples
|
text |
label |
Predicted label |
| 1 |
تقول نوال الزغبي : http |
neutral |
negative (p = 0.46) |
| 4 |
الفنانة نوال الزغبي سنة 90 http |
neutral |
positive (p = 0.61) |
| 7 |
“إيغيل فيلمز” تطلق “#ولعانة”.. و #نوال_الزغبي توجه كلمة لـ #ماغي_بوغصن عبر”فوشيا”http http |
neutral |
positive (p = 0.51) |
👉Overconfidence issues (1)
For records in the dataset where text contains "http", we found a significantly higher number of overconfident wrong predictions (435 samples, corresponding to 40.96% of the wrong predictions in the data slice).
| Level |
Data slice |
Metric |
Deviation |
| major 🔴 |
text contains "http" |
Overconfidence rate = 0.410 |
+29.93% than global |
Taxonomy
avid-effect:performance:P0204
🔍✨Examples
|
text |
label |
Predicted label |
| 4891 |
Visionari la “buona scuola” nelle relazioni di cinque esperti http #fb |
neutral |
positive (p = 0.97) |
|
|
|
neutral (p = 0.02) |
| 1934 |
Le 'chat le plus triste du monde' reçoit des milliers de demandes d'adoption http |
positive |
negative (p = 0.95) |
|
|
|
positive (p = 0.03) |
| 5038 |
http miglioriamo la scuola: una collaborazione degli storici della lingua a "La buona scuola" di Renzi |
neutral |
positive (p = 0.95) |
|
|
|
neutral (p = 0.03) |
👉Robustness issues (6)
When feature “text” is perturbed with the transformation “Transform to uppercase”, the model changes its prediction in 34.7% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| major 🔴 |
— |
Fail rate = 0.347 |
347/1000 tested samples (34.7%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Transform to uppercase(text) |
Original prediction |
Prediction after perturbation |
| 1640 |
I SUCCESSFULLY CAPTURED THE HATCHIMAL! My journey has ended and I will go down in the history books. #hatchimals… |
I SUCCESSFULLY CAPTURED THE HATCHIMAL! MY JOURNEY HAS ENDED AND I WILL GO DOWN IN THE HISTORY BOOKS. #HATCHIMALS… |
negative (p = 0.84) |
positive (p = 0.61) |
| 2040 |
Invités à La Rochelle, communistes et écologistes sans merci: Très critiques sur les «renoncements» de Françoi... http |
INVITÉS À LA ROCHELLE, COMMUNISTES ET ÉCOLOGISTES SANS MERCI: TRÈS CRITIQUES SUR LES «RENONCEMENTS» DE FRANÇOI... HTTP |
negative (p = 0.69) |
positive (p = 0.53) |
| 3117 |
Wenigstens weiß ich, dass meine Fotos nicht besonders gut sind. |
WENIGSTENS WEISS ICH, DASS MEINE FOTOS NICHT BESONDERS GUT SIND. |
negative (p = 0.83) |
positive (p = 0.51) |
When feature “text” is perturbed with the transformation “Transform to title case”, the model changes its prediction in 26.9% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| major 🔴 |
— |
Fail rate = 0.269 |
269/1000 tested samples (26.9%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Transform to title case(text) |
Original prediction |
Prediction after perturbation |
| 1641 |
My mom used to say,"Lay down with dogs, get up with fleas." Just sayin' . #TrumpTransitionTeam #DrainTheSwamp |
My Mom Used To Say,"Lay Down With Dogs, Get Up With Fleas." Just Sayin' . #Trumptransitionteam #Draintheswamp |
negative (p = 0.50) |
positive (p = 0.59) |
| 6612 |
@user
faltabas mi samo, la familia unida jajajajaja, te extraño, venn |
@User
Faltabas Mi Samo, La Familia Unida Jajajajaja, Te Extraño, Venn |
negative (p = 0.53) |
positive (p = 0.67) |
| 3467 |
@user
hehehe ;) was macht die Ikea Bastelei? |
@User
Hehehe ;) Was Macht Die Ikea Bastelei? |
negative (p = 0.43) |
positive (p = 0.45) |
When feature “text” is perturbed with the transformation “Add typos”, the model changes its prediction in 16.0% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| major 🔴 |
— |
Fail rate = 0.160 |
160/1000 tested samples (16.0%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Add typos(text) |
Original prediction |
Prediction after perturbation |
| 1483 |
5 right-backs Barcelona could sign as Vidal's replacement #fcblive |
5 eight-backs Barcelona could sign a Vidal's replacemwnt #fcblive |
positive (p = 0.46) |
negative (p = 0.42) |
| 1129 |
In terms of (exclusive) games I'd play on it, I can mostly only think of FFXV, Persona 5 (which I could play on PS3...), and... that's it? |
In terms of (rxclusive) games I'd play on it, I can mostly only think of FFXV, Persona 5 (which could pay on LS3...), and.... that's it? |
neutral (p = 0.42) |
negative (p = 0.43) |
| 4157 |
Lega rhia purei takt k sath sbhi bharrt venseyo ki duaey ap k sath hai |
KLega rhia puei takt k sath sbhi bharrt venseyo ki duaey ap k sqath hai |
positive (p = 0.37) |
negative (p = 0.36) |
When feature “text” is perturbed with the transformation “Transform to lowercase”, the model changes its prediction in 9.9% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| medium 🟡 |
— |
Fail rate = 0.099 |
99/1000 tested samples (9.9%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Transform to lowercase(text) |
Original prediction |
Prediction after perturbation |
| 1923 |
Les Verts Suisse |
Malgré les belles promesses écologistes, la majorité refuse de s'engager pour une économie verte /gruene/fr/positions/... |
les verts suisse |
malgré les belles promesses écologistes, la majorité refuse de s'engager pour une économie verte /gruene/fr/positions/... |
| 6072 |
#masterchefbr Mirian, oh, sucesso procê. A saída é ali no fundo. Brigado por ocê tê vindo, viu ? Bjo. |
#masterchefbr mirian, oh, sucesso procê. a saída é ali no fundo. brigado por ocê tê vindo, viu ? bjo. |
positive (p = 0.46) |
negative (p = 0.64) |
| 3500 |
Jaanma main bol rahi hu ki,tum mere twits dekho :/ |
jaanma main bol rahi hu ki,tum mere twits dekho :/ |
positive (p = 0.40) |
negative (p = 0.38) |
When feature “text” is perturbed with the transformation “Punctuation Removal”, the model changes its prediction in 9.3% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| medium 🟡 |
— |
Fail rate = 0.093 |
93/1000 tested samples (9.3%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Punctuation Removal(text) |
Original prediction |
Prediction after perturbation |
| 2576 |
Événement. Hommage à Mandela en direct des Sud à Arles http |
@user
http |
Événement Hommage à Mandela en direct des Sud à Arles http |
@user
http |
| 6507 |
@user
Para Nigtwish tengo asiento, aunque fueramos juntos no podríamos sentarnos uno al lado del otro Para los otros tengo pista |
@user
Para Nigtwish tengo asiento aunque fueramos juntos no podríamos sentarnos uno al lado del otro Para los otros tengo pista |
neutral (p = 0.35) |
positive (p = 0.36) |
| 3069 |
Hilfe! Hilfe! Ruf mal einer schnell die Modepolizei! Meine Augen werden gerade vergewaltigt! |
Hilfe Hilfe Ruf mal einer schnell die Modepolizei Meine Augen werden gerade vergewaltigt |
positive (p = 0.94) |
negative (p = 0.51) |
When feature “text” is perturbed with the transformation “Accent Removal”, the model changes its prediction in 9.1% of the cases. We expected the predictions not to be affected by this transformation.
| Level |
Data slice |
Metric |
Deviation |
| medium 🟡 |
— |
Fail rate = 0.091 |
91/1000 tested samples (9.1%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Accent Removal(text) |
Original prediction |
Prediction after perturbation |
| 5357 |
Victor B não ouse sair hoje!! Eu te venero #MasterChefBR |
Victor B nao ouse sair hoje!! Eu te venero #MasterChefBR |
negative (p = 0.44) |
positive (p = 0.64) |
| 6378 |
Cómo es eso que te extraño? Tamaaare |
Como es eso que te extrano? Tamaaare |
negative (p = 0.45) |
positive (p = 0.37) |
| 6464 |
@user
el sábado unas risas todos juntos... A por otro año más! |
@user
el sabado unas risas todos juntos... A por otro ano mas! |
positive (p = 0.47) |
negative (p = 0.46) |
Checkout out the Giskard Space and test your model.
Disclaimer: it's important to note that automated scans may produce false positives or miss certain vulnerabilities. We encourage you to review the findings and assess the impact accordingly.
