In today’s world, content moderation remains a major challenge. As platforms like online games increasingly attract an international audience, the complexity of content moderation has grown as hateful content makes its way across multiple languages and has a greater probability of passing through content moderation tools.
To tackle this challenge, we use multilingual embeddings to build a content moderation tool that works across 100+ languages and only requires training data in English.
For content moderation, we just need a handful of training examples of harmful and acceptable content in one language. For example, in English, we can then train a classifier to find the decision boundary in the vector space that helps us determine which type of content is undesirable on the platform.
Updated about 2 months ago