Verification-oriented comment detection
Challenge Description
Build a system that automatically retrieves comments that could be useful in the verification process.
Users tend to leave comments on Web videos expressing their personal opinion or experience in relation to what the video shows, convey public statements about the event shown in the video, and often challenge or support the credibility of the video. Several of these comments could be helpful for the verification process.
As a baseline approach, a predefined list of verification-related keywords is used to filter the video comments and create a subset with comments of relevance to the verification process. The list of verification-related comments is available in 7 languages (English, German, Greek, Arabic, French, Spanish and Farsi).
English | French | German | Arabic | Spanish | Farsi | Greek |
lies | mensonges | Lügen | أكاذيب | mentiras | دروغ | ψέματα |
fake | faux | falsch | أكاذيب | falso | جعلی | ψεύτικο |
wrong | erroné | falsch | غلطان | incorrecto | اشتباه | λάθος |
lie | mensonge | Lüge | أكذوبة، كذب | mentira | دروغ | ψέμα |
confirm | confirmé | bestätigt | مؤكد | confirmado | تأييد شده | επιβεβαιώνω |
where | où | wo | أين | donde | كجا | που |
location | lieu | Ort | مكان | lugar | محل، مكان | τοποθεσία |
lying | mentir | lügend | كذب | mitiendo | دروغ گو | ψευδής |
false | faux | falsch, fehlerhaft | خاطئ | falso | غلط، اشتباه، دروغین | εσφαλμένο |
incorrect | inexact | unrichtig | غير صحيح | incorrecto | غلط، اشتباه | λανθασμένο |
misleading | trompeur | irreführend | مضلل | engañoso | گمراه كننده | παραπλανητικό |
propaganda | propagande | Propaganda | دعاية | propaganda | تبلیغات سیاسی، پروپاگاندا | προπαγάνδα |
liar | menteur | Lügner | كاذب | mentiroso | کذاب | ψεύτης |
Input
- YouTube API responses containing the video comments.
- Verification-related keywords in 7 languages.
Output
List of verification-related comments and/or keywords
Evaluation
The approach will be evaluated qualitatively since no ground-thruth is provided. You can select up to five videos from the dataset as test videos and apply the approach on them. As a result, the extracted verification comments should provide useful information for debunking the video.
Suggested test videos:
Hints:
Experiment with weakly supervised learning using the verification-keywords to label the comments as verification related ones.