TL;DR
This paper introduces Hatebusters, a web application designed to combat YouTube hate speech by empowering a community of volunteer users to report it. It leverages a machine learning classifier to identify potentially hateful comments and employs gamification to encourage user engagement, aiming to accelerate the removal of harmful content.
What Problem Does This Paper Address?
Online hate speech, particularly on platforms like YouTube, poses a significant threat to individual well-being and public discourse. While social media companies rely on user reports and AI for moderation, the sheer volume of content makes manual review impossible, and current AI struggles with nuance. Existing reporting tools often take a passive approach, requiring users to actively seek them out. This paper addresses the need for a more proactive and community-driven system to identify and report hate speech, specifically on YouTube, to expedite its removal and mitigate its negative impact.
Methods
Hatebusters is built around three core components:
- Hate Speech Classification Engine (Hatewise): This engine uses a Support Vector Machine (SVM) classifier trained on YouTube comments. Comments are represented using a bag-of-words model with tf-idf values. The classifier outputs a confidence score, which is used to present comments to users.
- Backend Service: This service handles runtime operations and scheduled tasks. It includes APIs for user management and comment processing. Key tasks include daily fetching of new YouTube comments via the YouTube Data API, scoring them with Hatewise, and checking for deleted comments to maintain data integrity.
- User-Centric Frontend Platform: This platform is designed to be engaging and encourage user participation. It incorporates gamification elements such as a point system for ranking users and achievements to motivate them to report more comments.
Datasets
The paper describes the creation and use of a custom dataset for training the Hatewise classifier.
- Initial Training Set (March 2017): Comprised 454 samples, with 334 negative and 120 positive examples of hate speech. This set was collected using the Hatebusters pipeline and manually annotated by the platform administrators.
- Extended Training Set (April 2017): Expanded to a total of 936 samples, with 358 positive examples. This larger dataset was used to train a second, improved classifier.
The authors mention leveraging the CrowdFlower's AI for Everyone Challenge for Q4 of 2017 to obtain a new quality-controlled humanly annotated sample for future improvements.
Key Insights and Ideas
- Community-Driven Reporting: The core idea is to shift from passive reporting to an active, community-driven approach, making it easier for ordinary citizens to contribute to combating hate speech.
- Gamification for Engagement: Utilizing gamification elements like points and achievements is crucial for driving user engagement and encouraging consistent reporting behavior.
- Uncertainty Sampling for Administrators: For platform administrators, comments closer to the classifier's decision boundary (uncertainty sampling) are presented. This allows for targeted human review of ambiguous cases, improving the training data over time.
- Focus on YouTube Comments: The paper highlights the specific challenge of hate speech detection within YouTube comments, a domain less explored than platforms like Twitter.
Key Results
The paper reports improvements in the Hatewise classifier's performance with the expanded dataset:
- The F1-score increased from 0.459 to 0.654.
- Average precision improved from 0.556 to 0.722.
While specific quantitative results on user engagement or the impact of gamification are not detailed, the paper implies that these elements are designed to increase user reporting activity.
Limitations and Future Work
Limitations:
- The paper acknowledges that the initial classifier was trained on a relatively small dataset.
- The reliance on user-generated content for reporting means the system is dependent on user participation and their ability to correctly identify hate speech.
Future Work:
- Version 2 Development: The authors are actively working on Hatebusters version 2.
- Improved Classification: This includes exploring different comment representations, incorporating results diversification, utilizing deleted comments, and considering user information. They also aim to obtain more high-quality annotated data.
- Enhanced User Engagement: Future plans involve introducing more achievements and weekly digest emails to further boost user engagement.
Summary updated