Explainable AI-Based Multi-Lingual Content Moderation System

现有的审核工具是基于团队每天维护的网站的关键字名单或黑名单。这些都是缓慢和痛苦的维护，也嵌入了版主的偏见，因为他们存在的本质，往往是对某些种族的偏见。或者，像Facebook这样的平台每年花费高达25亿GB，雇佣3万人手动检查社交媒体帖子，面临着严重的心理健康风险。大多数工具没有考虑到不同品牌、代理商和广告网络对错误信息、仇恨言论和新形式有害内容的单独政策。该项目将为广告专业人员和人类版主制作一个创新的仪表盘，监控其库存网络上和目标网站列表中的网站。该系统将检测法语、西班牙语、意大利语、葡萄牙语和德语的宣传、仇恨言论、威胁等，并为任何高或低分类分数提供解释。核心创新是构建了一个集成的零迁移学习和机器翻译人工智能模型，该模型针对特定的、文化特定的、用于当地方言宣传检测的训练和测试数据进行训练。为了获得这一数据，Factmata将与法国反种族主义和反犹太主义国际联盟等主要社区合作。该项目将帮助任何托管内容的平台建立定制的模型，以删除或标记多种语言的有害内容，并根据自己的偏好和规则自动培训系统。Factmata在自动事实核查、自然语言处理(NLP)、数据注释等领域拥有经验；并成功地构建了检测仇恨言论和宣传语言的算法。它已经在辛迪加媒体网络和节目广告空间内拥有两个主要客户，他们已经要求使用拟议的产品。该团队有向内容审核行业销售产品的经验，并成功地建立了一项仅以英语进行宣传的服务。该项目将带来显著的出口拉动型增长、可观的投资回报率、增加就业机会和进一步的研发投资机会。

Existing moderation tools are based on keyword lists or blacklists of sites maintained by teams every day. These are slow and painful to maintain, and also embed the biases of moderators by their very nature of existing, often being biased against certain races. Alternatively, platforms like Facebook spend up to £2.5bn a year and employ 30,000 individuals to manually check social media posts, facing major risks to mental health. Most tools do not account for the separate policies that different brands, agencies and advertising networks have towards misinformation, hate speech and new forms of harmful content The project will produce an innovative dashboard for advertising professionals and human moderators monitoring websites on their inventory network, and in their target site lists. The system will detect propaganda, hate speech, threats and more across French, Spanish, Italian, Portuguese and German, and provide explanations for any high or low classification scores. The core innovation is building an ensemble zero-shot transfer learning and machine translation AI model which is trained on nique, culturally specific, training and test data for propaganda detection in local dialects. To obtain this data, Factmata will work with key communities such as the French Licra (International League Against Racism and Antisemitism). The project will help any platform that hosts content to build tailored models to take down or flag harmful content in multiple languages, and auto-train systems based on their own preferences and rules. Factmata has experience in the fields of automated fact checking, natural language processing ("NLP"), data annotation and more; and has successfully built algorithms to detect hate speech and propaganda language. It already has two major customers within the syndicated media network and programmatic advertising space, who have asked to use the product proposed. The team has experience selling into the content moderation industry and has built a successful service moderating propaganda in the English language only. The project will deliver significant export-led growth, a substantial ROI, increased employment and further opportunity for R&D investment.