EAGER: Collaborative: BystanderBots: Automated Bystander Intervention for Cyberbullying Mitigation

欺凌对受害者、旁观者和欺凌弱小的人都有持久的负面心理和身体影响；在线环境可以放大这些影响的规模和影响，因为匿名可以鼓励人们发表针对个人或群体的敌意帖子。该项目旨在通过在在线评论帖子中设计主动的、自动化的“旁观者干预”来减少这类帖子的流行。旁观者干预，即一个或多个欺凌事件的目击者向欺凌者施压，迫使欺凌者停止欺凌行为，这种干预在校园里通常是有效的，但人们往往不愿干预在线情景。取而代之的是，一个计算机程序可以发布包含这些干预的评论，如果欺负者认为这些帖子来自人类旁观者，如果化名下的欺凌者对旁观者的干预做出反应，就像他们在面对面对抗时一样，可能会减少最初发帖或其他可能堆积起来的人的后续攻击性。该项目将分三个主要阶段进行。第一阶段涉及通过更好地检测特定评论系统中与欺凌相关的非标准语言使用来改进网络欺凌检测。第二个阶段涉及开发一个对话系统，它的行为像一个人类旁观者，创建在给定的评论帖子的上下文中看起来合适的消息，并且包含心理上有效的旁观者干预。第三个阶段涉及在一个大型视频分享网站部署该工具，并监测其检测并通过干预缓解进一步欺凌的能力。如果成功，该项目可能会对减少社交媒体系统中的在线攻击性产生实际影响，同时减少对人类版主的需求(以及可能对其造成的伤害)；这些工具还将向社区发布，以支持围绕聊天机器人和人类如何在在线评论中互动的其他类型的研究。检测方面的工作旨在促进自然语言处理(NLP)和计算语用学，特别是在非规范语言使用方面，因为最先进的欺凌检测方案通常使用词袋方法，不考虑网络欺凌的语言和结构特征。该团队将探索如何通过基于特定主题更常与欺凌联系在一起的复杂特征来开发主题模型，以及如何通过寻找在给定上下文中使用与其在其他上下文中的位置不同的单词来识别欺凌的显性指标和隐含指标。上下文将被表示为单词的子空间，其中单词本身以低维单词嵌入的形式出现。该项目的对话生成部分将描述和展示心理学文献中有效的旁观者干预的特性。这种表示将驱动旨在自动生成旁观者响应的对话管理器，以便响应包含既可信又已知在减少在线欺凌方面有效的特征。这些组件将首先通过离线测试进行评估，使用标记为欺凌内容的评论数据和生成的对话的人类评分。一旦建立了一个相当有效的管道，它将在一系列在线实验中进行评估，在这些实验中，对评论线程进行监控，并为一些(但不是所有)被检测为包含欺凌行为的线程生成自动旁观者响应。该软件将记录受监控的线程和任何生成的响应，以及特定线程中自动旁观者响应之前和之后的行为；这些数据将允许团队评估稍后在线程中旁观者干预对欺凌事件的影响。

Bullying has lasting negative psychological and physical effects on victims, bystanders, and bullies alike; online settings can magnify both the scale and impact of these effects, as anonymity can embolden people to make hostile posts about individuals or groups. This project aims to reduce the prevalence of such posts through the design of active, automated "bystander interventions" in online comment threads. Bystander interventions, in which one or more witnesses to a bullying incident pressures the bully to stop, are often effective in schoolyards, but people are often reluctant to intervene in online scenarios. Instead, a computer program could post comments that contain these interventions, potentially reducing follow-on aggression from the original poster or others who might pile on -- if bullies perceive these posts as coming from human bystanders, and if bullies under the cover of pseudonyms react to bystander interventions as they do in in-person confrontations. The project will proceed in three main stages. The first stage involves improving cyberbullying detection through better detection of non-standard language use associated with bullying in a particular commenting system. The second stage involves developing a dialogue system that acts like a human bystander, creating messages that look appropriate in the context of given a comment thread and that contain psychologically-valid bystander interventions. The third stage involves deploying the tool in a large video sharing site and monitoring its ability to detect and, through interventions, mitigate further bullying. If successful, the project could have real impacts in reducing online aggression in social media systems while reducing the need for (and possible harms to) human moderators; the tools will also be released to the community to support other kinds of research around how chatbots and humans might interact in online comments.The work on detection aims to advance natural language processing (NLP) and computational pragmatics, particularly around non-canonical language use, because state-of-the-art bullying detection schemes typically use bag-of-words approaches that do not consider the linguistic and structural features of cyberbullying. The team will explore how to identify both explicit indicators of bullying, by developing topic models based on complex features where particular topics are more often associated with bullying, and implicit indicators, through looking for words whose use in a given context diverges from their location in other contexts. The context will be represented as a subspace of words, where the words themselves occur as low-dimensional word embeddings. The dialogue generation portion of the project will characterize and represent properties of effective bystander interventions from the psychology literature. This representation will drive a dialogue manager designed to generate bystander responses automatically so that the responses contain features that are both believable and are known to be effective in reducing bullying online. These components will first be evaluated through offline testing, using comment data labeled for bullying content and human ratings of the generated dialogue. Once a reasonably effective pipeline has been built, it will be evaluated in a series of online experiments in which comment threads are monitored and automated bystander responses generated for some, but not all, threads detected as containing bullying. The software will log the monitored threads and any generated responses, along with behavior both before and after the automated bystander response in a particular thread; these data will allow the team to evaluate the impact of the bystander intervention on bullying incidents later in the thread.