Better User Interfaces for Occasionally Failing Systems

技术越来越多地使用不可靠的系统作为解释输入的中心手段。常见的示例包括移动的设备中的语音和文本输入，以及汽车中的车道偏离检测。这种依赖暴露了一个根本问题，人们通常不了解底层系统，看似微小的系统或人为错误可能导致潜在的灾难性后果。虽然技术改进部分解决了这一问题，但拟议的研究计划通过更好地理解，理论，人类交互模型以及用于不可靠系统的新用户界面（UI）技术来寻求补充方法。该计划的第一部分分析了偶尔出现故障的系统周围的人类行为。我以前研究过这种行为，并将在这里将我的实验工作扩展到其他领域，例如，语音识别此外，该计划将通过众包研究收集更大的数据集，并分析它们的行为模式。以前，我确定了一个行为变化，由于挫折，通常约95%的可靠性，在两个任务域。这种变化和人类对失败的适应将针对任务域、任务关键性、用户和环境的不同组合进行系统研究。基于所获得的见解，第二个目标着眼于新的UI机制，以减轻不可靠的自动化的影响，例如，避免重复错误的预测，减少用户的挫折感。其他工作将研究新的解决方案，通过寻找具有低预测置信度和高校正度的实例来自动识别失败，例如，眼动追踪然后将鼓励用户切换到不同的UI模态或应用其他修复。另一个线程将调查传感技术，例如，EEG可以通过提供不同的预测/反应来触发更好的系统对故障的反应。*作为对第二个目标的补充，第三个目标侧重于为技术改进提供信息。用户更正表达了关于底层技术工作不够好的事件的有价值的信息。我将设计不同的方式，以适当的匿名方式向系统开发人员传达此类情况和由此产生的用户更正。所提出的研究的独创性源于其集成方法，用于迭代开发不完善系统的UI，包括减少或防止故障再次发生的新机制。收集到的数据将使我们能够更好地了解人类在不完美的系统中所做的权衡，并指导理论，模型和系统开发。预计这些结果将改善许多应用程序的UI，每天有数百万人使用。因此，即使是看似微小的改进也会产生重大影响。这些工作将产生更好的UI，更好地理解人类在不完美系统周围的行为和理论，同时为系统改进提供信息。

Technology increasingly employs unreliable systems as a central means to interpret input. Common examples include voice and text input in mobile devices, and lane departure detection in cars. This reliance exposes a fundamental problem people do not generally understand the underlying systems, and seemingly small system or human errors can lead to potentially disastrous consequences. While technical improvements partially address this, the proposed research program pursues a complementary approach through better understanding, theories, models of human interaction with, and new user interface (UI) technologies for, unreliable systems.***The first part of the proposed program analyzes human behaviours around occasionally failing systems. I have previously studied such behaviours and will here expand my experimental work to other domains, e.g., speech recognition. Additionally, the program will gather larger data sets with crowd-sourced studies and analyze them for behavioural patterns. Previously, I identified a behavioural change due to frustration, typically around 95% reliability, in two task domains. This change and human adaptation to failures will be studied systematically for different combinations of task domain, task criticality, user, and context.***Building on insights gained, the second objective looks at new UI mechanisms to alleviate the effects of unreliable automation, e.g., by avoiding repeated wrong predictions, reducing user frustration. Other work will investigate novel solutions that automatically identify failures by looking for instances with both low prediction confidence and high correction, e.g., in eye tracking. Users will then be encouraged to switch to a different UI modality or apply other fixes. Another thread will investigate if sensing technologies, e.g., EEG, can trigger better system responses to failures, by offering different predictions/reactions.***Complementing the second objective, the third one focuses on informing technical improvements. User corrections express valuable information about incidents where the underlying technologies are not working well enough. I will devise different ways to communicate such instances and resultant user corrections to system developers in appropriately anonymized ways.***The originality of the proposed research stems from its integrative approach for iteratively developing UIs for imperfect systems, including new mechanisms to reduce or prevent failure reoccurrence. The collected data will enable us to better understand the trade-offs humans make around imperfect systems, and guide theory, model and system development. The results are expected to improve UIs for many applications, used by millions of people every day. Thus, even seemingly small improvements will have significant impact. Together the work will yield better UIs and a better understanding of, and theories for, human behaviours around imperfect systems, while simultaneously informing system improvement.**