CAREER: Machine Learning-Based Approaches Toward Combatting Abusive Behavior in Online Communities

这项研究的目的是通过计算模拟网络上的滥用行为，以建立有助于对抗它的工具，其目标是使互联网成为一个更受欢迎的地方。从早期开始，网络上就充斥着煽动性言论、网络喷子、骚扰和谩骂。这个项目将揭示多种类型的在线对话中网络滥用的结构，这是研究计算机媒介交流的重要一步。这将导致统计机器学习算法作为理论启发的社会语言学变量的函数的建模滥用，并将需要新的技术和方法进步。这项工作将使依赖于计算生成的滥用预测的新型自动化和半自动化应用程序成为可能。教育和推广计划与研究活动密切相关，重点是扩大研究的广泛影响。公共应用程序编程接口（API）将使世界各地的开发人员和在线社区管理人员能够将本研究开发的防止滥用的防御措施集成到他们自己的站点中。这项工作将包括两个主要阶段。首先，该研究将通过统计机器学习技术深入了解在线滥用行为。具体来说，这项工作将利用社会科学和语言学的理论，为鲁棒统计机器学习算法的特征创建提供信息，以预测滥用行为。这些被提议的滥用模型将使一种全新的、变革性的混合主动工件能够干预社交媒体和在线社区。在第二阶段，该项目将通过构建、部署和评估打击滥用的社会技术工具来探索这类新启用的工件。具体来说，它将探讨使用滥用预测的两类工具：屏蔽和版主工具。第一个是盾牌，它将主动阻止入境的辱骂到达人们手中。第二类工具，版主工具，将为社区版主标记和分类滥用。

This research aims to computationally model abusive online behavior to build tools that help counter it, with the goal of making the Internet a more welcoming place. Since its earliest days, flaming, trolling, harassment and abuse have plagued the Internet. This project will lay bare the structure of online abuse over many types of online conversations, a major step forward for the study of computer-mediated communication. This will result from modeling abuse with statistical machine learning algorithms as a function of theoretically inspired, sociolinguistic variables, and will entail new technical and methodological advances. This work will enable a transformative new class of automated and semi-automated applications that depend on computationally generated abuse predictions. The education and outreach plan is deeply tied to the research activities, and focuses on scaling-up the research's broader impacts. A public application programming interface (API) will enable developers and online community managers around the world to integrate into their own sites the defenses against abuse developed by this research.The work will consist of two major phases. In the first, the research will develop a deep understanding of abusive online behavior via statistical machine learning techniques. Specifically, the work will appropriate theories from social science and linguistics to inform the creation of features for robust statistical machine learning algorithms to predict abuse. These proposed abuse models will enable a brand new, transformative class of mixed-initiative artifacts capable of intervening in social media and online communities. In the second phase, this project will explore this newly enabled class of artifacts by building, deploying and evaluating sociotechnical tools for combatting abuse. Specifically, it will explore two classes of tools that use the abuse predictions: shields and moderator tools. The first, shields, will proactively block inbound abuse from reaching people. The second class of tools, moderator tools, will flag and triage abuse for community moderators.