Visual Analytics of Online Streaming Text

我们的全球交流方式在很大程度上依赖于非结构化文本信息的交换。随着短信、社交网络和在线新闻媒体的兴起，这些数据的频繁消费和生产已经成为我们现代社会的一个决定性因素。随着这种增长，现在有广泛的应用程序领域可以从发现共同发生的主题、分析相关用户行为和检测这些数据源中的异常内容的能力中受益。然而，与此同时，我们正面临着错误信息和谣言在全球迅速和不受控制地传播所带来的前所未有的威胁。恶意活动，如部署社交机器人网络传播虚假声明或破坏公共话语，经常被观察到。虽然在过去，自动内容传播算法的能力有限，但我们现在可以看到先进的行为复杂性，可信的反应以及复杂的精心策划的活动。为了了解内容模式的演变，检测异常信息，发现大规模的协同活动，我们必须应对实时流文本的固有挑战。虽然过去的大多数研究都是针对批处理语料库处理，但对分析实时文本数据的挑战只有有限的思考。在这个项目中，我们建议通过创建一种新颖的视觉分析方法来缩小这一差距，该方法适应并扩展了自然语言处理、机器学习和视觉界面，并将它们集成到一个交互式分析管道中。我们将首先从社交网络、新闻网络和微博客中获取一个合适的基准库，并创建一个允许模拟重播的系统，以实现真实的评估场景。利用该系统，我们将研究现有视觉隐喻和交互模式的实时适用性、适应性和可扩展性。最后，为了从抽样分析到大规模理解不断变化的主题、相关行为和抽样不确定性，我们将把我们的可视化和交互方法与专门适应的文本挖掘工具集成在一起。在这里，我们将特别阐述生成内容模型的可扩展性以及进化主题层次聚类。通过将现有工具集成为可视化交互管道的一部分，我们可以允许以任务为中心的预聚合和过滤，以减少分析人员的认知负荷。此外，通过打开已建立方法的黑盒子，我们将它们与在线可视化配置和控制相结合。基于上下文知识、专业知识和直觉，分析人员将能够收回对迭代计算过程的控制，帮助系统解释中间结果，并允许与分析推理持续一致。

Our global ways of communication are heavily based on the exchange of unstructured textual information. With the rise of short text messaging, social networks, and online news media, the frequent consumption and production of such data have become a defining element of our modern society. With this rise, there now is a broad range of application areas that could benefit from the ability to discover co-occurring topics, analyze correlated user behavior, and detect anomalous content in these data sources. However, at the same time, we are facing unprecedented threats introduced by the fast and uncontrolled global spread of misinformation and rumors. Malicious activities, such as the deployment of social bot networks to disseminate false claims or to sabotage public discourse, have been frequently observed. While in the past, automated content spreading algorithms were limited in their capabilities, we can now see advanced behavioral complexity, believable reactions, and sophisticated, orchestrated campaigns.In order to understand the evolution of content patterns, detect anomalous information, and discover large scale coordinated activities, we have to cope with the inherent challenges of real-time streaming text. While most of the past research has been directed towards batch corpus processing, only limited thought has been given to the challenge of analyzing live-streaming textual data. In this project we propose to close this gap by creating a novel Visual Analytics methodology that adapts and extends natural language processing, machine learning, and visual interfaces and integrates them into an interactive analytical pipeline. We will first acquire a suitable benchmark repository from social networks, news wire, and microblogs, and create a system that allows simulated replay to enable realistic evaluation scenarios. Using this system, we will investigate the real-time applicability, adaptability, and extensibility of existing visual metaphors and interaction patterns.Finally, to make the step from sampled analysis to large scale understanding of evolving topics, correlated behaviors, and sampling-uncertainties, we will integrate our visualization and interaction methods with specifically adapted text-mining tools. Here, we will particularly elaborate on the extensibility of generative content models as well as evolutionary topic hierarchy clustering. By integrating the existing tools as part of the visual interaction pipeline, we can allow task-centered pre-aggregation and filtering to reduce cognitive load for the analyst. Moreover, by opening the black box of established methods, we will integrate them with online visual configuration and control. Based on context knowledge, expertise, and intuition, the analyst will then be enabled to take back control of the iterative computational process, help the system to interpret intermediate results, and allow continuous alignment with analytical reasoning.