Digital media is an essential source of content, communications, and commerce in the modern world, but harmful content has traumatized users, incited violent acts, and negatively impacted already marginalized groups. There is a need to find a balance between supporting responsible speech, bolstering public safety, and providing for responsible digital communications practices.
Transparent and standardized approaches to content categorization are essential to content moderation programs, but global content moderation is very challenging and a “one size fits all” approach to content moderation does not work. AI-powered content moderation tools can help to automate the moderation of content and also help companies adapt their content moderation programs to meet regional standards.
A common framework for content moderation: The Global Alliance for Responsible Media (GARM)
Terms of service related to content moderation rules have grown to become tens-of-thousands word documents at many organizations. The largest technology companies employ thousands of content moderators. This work has proven very complicated and most companies not do not employ enough contact moderators to filter content and many users' languages. There is a strong need to efficiently moderate content at scale and to do so in a way that is fair, effective, and transparent.
The Global Alliance for Responsible Media (GARM) addresses the challenge of harmful online content by attempting to create a framework that describes it using consistent and understandable language. GARM is a joint effort of prominent marketers, media agencies, media platforms, and industry associations attempting to safeguard digital media by stopping harmful content from being monetized through advertising.
GARM has developed common definitions to ensure that the advertising industry is categorizing harmful content in the same way across the board. GARM is motivated by an effort to improve transparency in the availability, monetization, and inclusion of content within advertising campaigns, but this framework has seen widespread adoption across all forms of digital media.
GARM organized harmful content into 11 categories:
- Adult & Explicit Sexual Content
- Arms & Ammunition
- Crime & Harmful acts to individuals and Society
- Human Right Violations
- Death, Injury, or Military Conflict
- Online piracy
- Hate speech & acts of aggression
- Obscenity and Profanity
- Illegal Drugs/Tobacco/e-cigarettes/Vaping/Alcohol
- Spam or Harmful Content
- Debated Sensitive Social Issue
Adapting content moderation standards with AI
If your company works across borders or serves customers in different languages, you are probably already familiar with the challenge of creating acceptable speech standards and applying them universally. The GARM standard provides an excellent framework for content moderation, but implementing a single, comprehensive system for managing content standards is still very challenging given the many alphabets, languages, and cultural norms that exist across the world.
The good news is that there are effective methods for content moderation that can adapt to regional cultural and linguistic differences, can learn and get more accurate over time, and can even apply these standards across multiple languages automatically. Enter the world of natural language understanding and AI. Let's take a quick look at how it all works.
Build custom models
In AI, an "embedding" is a type of model that can be learned and reused across models. By using an existing embedding model new custom models can be created quickly. This can greatly accelerate the process of customizing content moderation models to meet the needs of regional audiences.
Learn from real user data
Training data is the data you use to train an algorithm or machine learning model to predict the outcome you design your model to predict. AI models can adapt to learn to recognize moderation concepts that may be specific to a given audience by providing training data taken from real customer examples.
Keep a human in the loop
Human-in-the-loop is a branch of AI that brings together AI and human intelligence to create machine learning models. It's when humans are involved with setting up the systems, tuning and testing the model so the decision-making improves, and then actioning the decisions it suggests. Human moderators play a key role in adapting AI tools to meet regional standards.
Transfer learn
Transfer learning means taking the relevant parts of a pre-trained machine learning model and applying it to a new but similar problem. This will usually be the core information for the model to function, with new aspects added to the model to solve a specific task. Transfer learning can play a key role in adapting AI models to meet regional needs.
The future of content moderation is AI-powered
Digital platforms have to build content moderation regimes that are scalable and serve their entire customer base. Companies can effectively manage global content moderation programs by leveraging standardized categories for content moderation and then adapting the interpretation of these categories to meet the needs of regional audiences. AI-powered content moderation can learn from real users to data and be trained to moderate content based on regional norms.
