Report for cardiffnlp/twitter-xlm-roberta-base-sentiment-multilingual

#195
by giskard-bot - opened
Giskard org

Hi Team,

This is a report from Giskard Bot Scan 🐢.

We have identified 11 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).

You can find a full version of scan report here.

👉Robustness issues (7)

When feature “text” is perturbed with the transformation “Transform to uppercase”, the model changes its prediction in 25.2% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
major 🔴 Fail rate = 0.252 Transform to uppercase 252/1000 tested samples (25.2%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Transform to uppercase(text) Original prediction Prediction after perturbation
6037 semana que vem é a chance de vários embustes saírem #MasterChefBR SEMANA QUE VEM É A CHANCE DE VÁRIOS EMBUSTES SAÍREM #MASTERCHEFBR negative (p = 0.91) neutral (p = 0.59)
3982 in a parallel universe IN A PARALLEL UNIVERSE neutral (p = 0.87) negative (p = 0.94)
4528 #frommywindow. Il cielo della Bassa. http #FROMMYWINDOW. IL CIELO DELLA BASSA. HTTP neutral (p = 0.98) positive (p = 0.85)

When feature “text” is perturbed with the transformation “Transform to title case”, the model changes its prediction in 18.5% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
major 🔴 Fail rate = 0.185 Transform to title case 185/1000 tested samples (18.5%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Transform to title case(text) Original prediction Prediction after perturbation
585 الرياض.. انخفاض على الحرارة وأجواء مغبرة يوم الثلاثاء بمشيئة الله #طقس_العرب http الرياض.. انخفاض على الحرارة وأجواء مغبرة يوم الثلاثاء بمشيئة الله #طقس_العرب Http negative (p = 0.41) neutral (p = 0.44)
2493 Google multiplie les erreurs pour le lancement de son écosystème Android Wear via Presse citron http Google Multiplie Les Erreurs Pour Le Lancement De Son Écosystème Android Wear Via Presse Citron Http negative (p = 0.97) neutral (p = 0.68)
4245 ye sab to bahane hai ghar ka kaam na krne ke . . . Ye Sab To Bahane Hai Ghar Ka Kaam Na Krne Ke . . . negative (p = 0.49) neutral (p = 0.41)

When feature “text” is perturbed with the transformation “Add typos”, the model changes its prediction in 16.5% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
major 🔴 Fail rate = 0.165 Add typos 165/1000 tested samples (16.5%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Add typos(text) Original prediction Prediction after perturbation
6750 @user lo se pero tengo la esperanza de que en realidad sea un personaje mucho mas oscuro @user lo se pero tengo la esperanza de que en realidad sea un personaje muchl mas oscurpo negative (p = 0.64) neutral (p = 0.86)
5966 #VideoShowAoVivo morta q faço aniversário no msm dia do otaaa #ideoShowAoVivo mota q faço anivresário no msn da do ofaaa negative (p = 0.90) neutral (p = 0.93)
2694 Den ganzen tag bei der alten Verwandtschaft gesessen, weil will ja niemanden enttäuschen. I'M A SLAAAAVE FOR YOU Den ganzen tag bei der alten Verwandtsxhaft gesessn, weil will ja nuiemanden enttäuschen. I'M A SLAAAAVE OFR YOU positive (p = 0.93) negative (p = 0.98)

When feature “text” is perturbed with the transformation “Punctuation Removal”, the model changes its prediction in 12.1% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
major 🔴 Fail rate = 0.121 Punctuation Removal 121/1000 tested samples (12.1%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Punctuation Removal(text) Original prediction Prediction after perturbation
3708 100 bhi naa hopayenge :p #indvsuae #iccworldcup2015 100 bhi naa hopayenge p #indvsuae #iccworldcup2015 negative (p = 0.51) neutral (p = 0.40)
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
2432 Autorisation unique "loi sur l'eau" : le Medde publie une fiche récapitulative @user http Autorisation unique loi sur l eau le Medde publie une fiche récapitulative @user http

When feature “text” is perturbed with the transformation “Transform to lowercase”, the model changes its prediction in 8.1% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
medium 🟡 Fail rate = 0.081 Transform to lowercase 81/1000 tested samples (8.1%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Transform to lowercase(text) Original prediction Prediction after perturbation
2791 @user ach so, hmm Wahrheit @user ach so, hmm wahrheit neutral (p = 0.59) positive (p = 0.60)
910 @user She will be hearing my voice on her hesitation to back HRC. I am a MA voter. @user @user @user @user she will be hearing my voice on her hesitation to back hrc. i am a ma voter. @user @user @user neutral (p = 0.51) positive (p = 0.60)
2467 #ecologie Document : le projet d'arrêté sur les certificats d'économie d'énergie dévoilé http #ecologie document : le projet d'arrêté sur les certificats d'économie d'énergie dévoilé http positive (p = 0.62) neutral (p = 0.65)

When feature “text” is perturbed with the transformation “Accent Removal”, the model changes its prediction in 7.8% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
medium 🟡 Fail rate = 0.078 Accent Removal 78/1000 tested samples (7.8%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Accent Removal(text) Original prediction Prediction after perturbation
2301 @user si ça continue, les chatons seront des espèces en voie de disparition :D @user si ca continue, les chatons seront des especes en voie de disparition :D positive (p = 0.59) negative (p = 0.82)
826 #الاقتصاد #أستراليا و #فرنسا ستوقعان صفقة غواصات بأكثر من 36 مليار دولار غداًhttp http #الاقتصاد #استراليا و #فرنسا ستوقعان صفقة غواصات باكثر من 36 مليار دولار غداhttp http positive (p = 0.62) neutral (p = 0.52)
472 وزير الخارجية الروسي "سيرغي لافروف" أن موسكو تأمل في إجراء حوار بناء بشأن الأزمة في سورية خصوصاً حلب خلال لقائه... http وزير الخارجية الروسي "سيرغي لافروف" ان موسكو تامل في اجراء حوار بناء بشان الازمة في سورية خصوصا حلب خلال لقايه... http positive (p = 0.85) neutral (p = 0.72)

When feature “text” is perturbed with the transformation “Transform numbers to words”, the model changes its prediction in 5.72% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
medium 🟡 Fail rate = 0.057 Transform numbers to words 43/752 tested samples (5.72%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201
🔍✨Examples
text Transform numbers to words(text) Original prediction Prediction after perturbation
120 #الاتحاد_الاوروبي يمدد رسميا العقوبات الاقتصادية على #روسيا لمدة ستة أشهر جديدة لتاريخ ٣١ تموز يوليو ٢٠١٧ #موسكو #بروكسل #عاجل #الاتحاد_الاوروبي يمدد رسميا العقوبات الاقتصادية على #روسيا لمدة ستة أشهر جديدة لتاريخ واحد و ثلاثون تموز يوليو ألفان و سبعة عشر #موسكو #بروكسل #عاجل positive (p = 0.46) neutral (p = 0.54)
135 #الاتحاد_الاوروبي يمدد العقوبات الاقتصادية على #روسيا بسبب الأزمة الاوكرانية حتى 31 يوليود#أوكرانيا #كييف #القرم… http #الاتحاد_الاوروبي يمدد العقوبات الاقتصادية على #روسيا بسبب الأزمة الاوكرانية حتى واحد و ثلاثون يوليود#أوكرانيا #كييف #القرم… http neutral (p = 0.73) negative (p = 0.92)
150 في موسكو ...إشتري 4 بكرات ورق لامؤاخذه (تواليت)و إربح رحله سياحيه لفردين ل (مصر) ... #تحيا_ماسر http في موسكو ...إشتري أربعة بكرات ورق لامؤاخذه (تواليت)و إربح رحله سياحيه لفردين ل (مصر) ... #تحيا_ماسر http positive (p = 0.57) neutral (p = 0.50)
👉Overconfidence issues (2)

For records in the dataset where text contains "user", we found a significantly higher number of overconfident wrong predictions (513 samples, corresponding to 77.14% of the wrong predictions in the data slice).

Level Data slice Metric Deviation
medium 🟡 text contains "user" Overconfidence rate = 0.771 +15.31% than global

Taxonomy

avid-effect:performance:P0204
🔍✨Examples
text label Predicted label
3038 RT @user : Niddarenaturierung Frankfurt Sossenheim mit A. Hanisch u. T. Schlimme. Gelungene Veranstaltung @user http positive neutral (p = 0.99)
negative (p = 0.00)
4734 #genitori scuola dell'infanzia costituzione* i grandi assenti de #labuonascuola per @user adesso a #fahrenheit negative neutral (p = 0.99)
positive (p = 0.01)
5013 @user ma adesso arriva la buona scuola...e siamo tutti felici :-) @user negative positive (p = 0.99)
neutral (p = 0.01)

For records in the dataset where text_length(text) < 140.500 AND text_length(text) >= 96.500, we found a significantly higher number of overconfident wrong predictions (603 samples, corresponding to 74.81% of the wrong predictions in the data slice).

Level Data slice Metric Deviation
medium 🟡 text_length(text) < 140.500 AND text_length(text) >= 96.500 Overconfidence rate = 0.748 +11.83% than global

Taxonomy

avid-effect:performance:P0204
🔍✨Examples
text text_length(text) label Predicted label
3038 RT @user : Niddarenaturierung Frankfurt Sossenheim mit A. Hanisch u. T. Schlimme. Gelungene Veranstaltung @user http 115 positive neutral (p = 0.99)
negative (p = 0.00)
3414 Toter Finanzchef: Zurich bestätigt Existenz eines Abschiedsbriefes: Der Schweizer Versicherer Zurich hat die E... http 118 negative neutral (p = 0.99)
positive (p = 0.01)
4734 #genitori scuola dell'infanzia costituzione* i grandi assenti de #labuonascuola per @user adesso a #fahrenheit 112 negative neutral (p = 0.99)
positive (p = 0.01)
👉Ethical issues (2)

When feature “text” is perturbed with the transformation “Switch countries from high- to low-income and vice versa”, the model changes its prediction in 7.62% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
medium 🟡 Fail rate = 0.076 Switch countries from high- to low-income and vice versa 16/210 tested samples (7.62%) changed prediction after perturbation

Taxonomy

avid-effect:ethics:E0101 avid-effect:performance:P0201
🔍✨Examples
text Switch countries from high- to low-income and vice versa(text) Original prediction Prediction after perturbation
955 #Syria #Hezbollah Nasrallah's bodyguard identified in #Aleppo #Kazakhstan #Hezbollah Nasrallah's bodyguard identified in #Aleppo negative (p = 0.64) neutral (p = 0.58)
1471 The UK Doctor Who Beat The British GMC By Proving That Vaccines Aren’t Necessary To Achieve Health… The Chad Doctor Who Beat The Egyptian GMC By Proving That Vaccines Aren’t Necessary To Achieve Health… negative (p = 0.90) neutral (p = 0.51)
1693 Jacob #Israel on how you are really not here right now. #Bibi #Yelev #Jerusalem #Blackfriday Jacob #Iran on how you are really not here right now. #Bibi #Yelev #Jerusalem #Blackfriday negative (p = 0.84) neutral (p = 0.77)

When feature “text” is perturbed with the transformation “Switch Gender”, the model changes its prediction in 6.78% of the cases. We expected the predictions not to be affected by this transformation.

Level Metric Transformation Deviation
medium 🟡 Fail rate = 0.068 Switch Gender 56/826 tested samples (6.78%) changed prediction after perturbation

Taxonomy

avid-effect:ethics:E0101 avid-effect:performance:P0201
🔍✨Examples
text Switch Gender(text) Original prediction Prediction after perturbation
910 @user She will be hearing my voice on her hesitation to back HRC. I am a MA voter. @user @user @user @user he will be hearing my voice on his hesitation to back HRC. I am a MA voter. @user @user @user neutral (p = 0.51) positive (p = 0.52)
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. God is life worth living ? Tesla mannequin S,o YES. positive (p = 0.69) negative (p = 0.49)
1210 Retweeted CS Monitor (@csmonitor):Bannon said the administration planned to usher in a "new political... Retweeted CS Monitor (@csmonitor):Bannon said the administration planned to usherette in a "new political... neutral (p = 0.70) negative (p = 0.90)

Checkout out the Giskard Space and Giskard Documentation to learn more about how to 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.

Sign up or log in to comment