Hive Text Moderation

Text classifier for sexual content, threats, bullying, hate, drugs and spam.

About Hive Text Moderation

Hive’s Text Moderation Model is designed to detect online language, including explicit abuse, indirect threats, coded language, and evasive phrasing. Hive also offers OCR Moderation, which it says enables Trust and Safety teams to analyze and flag any inappropriate language from media circulating their platforms.

Hive names Tango as a customer, describing it as one of the world’s largest live streaming platforms with more than 500 million registered users. Tango deployed Hive’s Visual Moderation and Text Moderation, and Text Moderation extended coverage to live chat, identifying harassment, solicitation, and harmful language in real time. Hive also states that Plato, a gaming community, was able to implement an automatic filter using Hive text models in public group chats.

Drafted from hivemoderation.com · not yet confirmed by Hive

How it works

Text Moderation combines deep learning with pattern-based methods and returns multilevel classes for sexual content, threats of violence, bullying, hate speech and drug-related text, so a team can set enforcement around its own standards. It is built to handle misspellings, leetspeak, repeated characters and character substitutions used to get past simple filters.

An OCR model extracts overlay, scene and document text, plus emojis, from images and video for the same checks. By default and at no extra cost, results include matches against the IWF's keyword and URL lists for child sexual abuse, and a separate CSE text classifier built with Thorn flags potential child sexual exploitation in chats and comments.