EAGER: Using Large-scale Web Data for Online Attention Models and Identification of Reading Disabilities

媒体网站现在可以从数百万读者那里获得复杂的参与度。这些测量，例如页面内滚动和视口位置，可以帮助我们了解用户注意力的模式，而不仅仅是简单的测量，例如在页面上花费的时间。这个探索性的项目将通过展示这些数据（主要是在在线新闻文章的背景下）如何更好地理解用户在线行为，并帮助捕获（和推理）大规模的用户注意力，来改变这种未被充分利用的注意力数据的使用。然而，这样的数据是非常嘈杂和具有挑战性的分析。研究人员将探索开发数据分析的新技术，以推理语言，文本和网络上的注意力之间的联系。利用这些数据和新颖的分析技术，该项目还将探索如何识别有阅读困难的用户，并在在线阅读任务中为他们提供潜在的支持。该项目的潜在成果是算法和方法，可供新闻内容提供商使用，以根据他们如何互动和阅读在线新闻文章，对不同类型的读者进行更深入的了解。更重要的是，潜在的成果还包括这些出版商如何发现和支持有阅读困难的用户。这种进步有可能改善约7%至15%的阅读障碍人群的在线体验。该项目直接解决了使这种新型大规模数据在不同环境中有用和可用的多重挑战，并可能在大规模信息管理，机器学习和人机交互方面做出许多关键的智力贡献。第一个重大挑战是对原始大规模数据进行建模和处理，以产生鲁棒的注意力信号。数据非常嘈杂，分析和理解具有挑战性，不同格式的新闻文章，不同的内容类型，由不同的用户群体使用。其次，该项目将通过使用新的深度学习技术来大规模理解语言和注意力之间的相互作用，从而做出独特的贡献。这种建模可以扩展当前的自然语言处理（NLP）技术，以改进用于在新闻文章的上下文中分析语言和叙述的方法，并且更广泛地用于阅读任务。最后，将需要结合实地研究和大规模数据分析，以了解可能在阅读方面有困难的用户的注意力模式。一项实地研究将从已知有阅读障碍的用户那里收集数据，以开发一种技术，该技术可以帮助从大型交互数据集中的注意力数据估计用户的阅读困难。因此，这个项目将通知工作，结合自然语言处理和人机交互，建议可能的路径，支持这些用户在网上阅读任务。为了支持进一步的研究和再现性，该项目中使用的数据，软件和模型尽可能提供给其他研究人员。该项目的成果将通过研究论文和讲座传播给学术界和工业界的观众，并通过项目网站。该奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的知识价值和更广泛的影响审查标准进行评估的支持。

Media websites now capture intricate measures of engagement from millions of readers. These measures, such as in-page scrolling and viewport position, can help us understand patterns of user attention beyond simple measures like time spent on page. This exploratory project will transform the use of this under-utilized attention data by showing how such data, mostly in the context of online news articles, can lead to better understanding of user behavior online, and help capture (and reason about) user attention at scale. However, such data is extremely noisy and challenging to analyze. The researchers will explore developing new techniques for data analysis to reason about the connection between language, text, and attention on the Web. Using these data, and novel analysis techniques, the project will also explore how to identify users with reading difficulties, and potentially support them in online reading tasks. The potential outcomes of this project are algorithms and methods that can be used by news content providers to develop a refined understanding of different types of readers, according to how they interact and read news articles online. More importantly, potential outcomes also include advancements in how such publishers can detect, and support, users with reading difficulties. Such advancement has the potential to lead to improvements in online experiences for the estimated 7% to 15% of the population who suffer from reading disabilities. This project directly addresses multiple challenges in making this new type of large-scale data useful and usable in different settings, and is likely to result in a number of key intellectual contributions in large-scale information management, machine learning, and human computer interaction. The first significant challenge is in modeling and processing the raw large-scale data to result in a robust attention signals. The data is extremely noisy and challenging to analyze and understand, with news articles of different formats, different content types, used by different groups of users. Second, the project will make a unique contribution by using novel deep learning techniques to understand the interaction between language and attention, at scale. Such modeling can extend current Natural Language Processing (NLP) techniques to improve methods for the analysis of language and of narrative in the context of news articles, and more broadly for reading tasks. Finally, a combination of field studies and large-scale data analysis will be required to understand the attention patterns of users who may have difficulties in reading. A field study will collect data from users who are known to have reading disabilities, in order to develop a technique that can help estimate a user's reading difficulty from the attention data in a large interaction dataset. This project will thus inform work that combines NLP and Human-Computer Interaction to suggest possible paths for supporting such users in online reading tasks. In order to support further research and reproducibility, the resulting data, software, and models used in this project available to other researchers, as possible. The project results will be disseminated via research papers and talks to be academic and industry audiences, and through the project website.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.