III: Medium: Collaborative Research: Athena: Learning-oriented Search with Personalized Learning Flows

雅典娜项目将开发一种名为“搜索即学习”的技术，这是一套鼓励和支持学习的搜索技术，而不仅仅是简单的文档查找。为了学习，搜索者必须接触既新颖又易懂的信息。因此，在核心，Athena将通过建模几个重要因素来支持学习：（1）覆盖主题的文档之间的知识联系，（2）用户对该主题的当前知识状态，（3）用户可能从文档中获得的知识类型，以及（4）用户成功参与文档所需的知识。雅典娜项目将涉及两种类型的端到端系统，这两种系统都将建模和利用学习者的知识状态（LSK）：一个LSK感知搜索引擎和一个LSK感知问答系统。雅典娜系统将引导用户通过一个主题，并在以前遇到的信息和主题网络中捕获的主题结构的上下文中找到相关信息。该团队将使用标准措施以及一系列涉及人类受试者的研究来评估雅典娜。如果雅典娜项目成功，它将使人们更容易使用搜索引擎和相关技术来了解复杂的主题，其中有许多相互关联和相互依赖的子主题应该考虑。鉴于搜索是网络内外最常见的在线活动之一，Athena及其技术将对试图学习这些主题的搜索者产生重大影响。Athena使用称为学习流图（LFG）的数据结构实现“搜索即学习”。LFG包括表示子主题的节点（例如，概念）和表示子主题之间的关系的顶点（例如，一个子主题是理解另一个子主题的基础）。Athena利用LFG对上述不同因素进行建模。它使用LFG中节点之间的概率分布来建模：（1）用户的知识状态，（2）从信息项中获得的潜在知识，以及（3）用户成功参与信息项所需的先决条件知识。Athena团队将开发从结构化、半结构化和非结构化资源（例如，课程大纲、目录、书籍索引、知识库、查询日志），用于将LFG集成到搜索和问答模型中的算法，以及用于基于搜索行为重新估计LFG和用户的知识状态的算法（例如，查询、点击、跳过、停留时间等）。结构化文本数据以找到通过它的最佳学习路径是非常有趣的，尽管大多数现有的工作都集中在提取信息以填充“知识库”中的插槽，这是一个更细粒度的任务。LFG表示还提供了对更大主题的一种解释，与可解释系统的广泛兴趣相联系。Athena的工作将扩展文本表示、神经方法（包括注意力技术）、查询和主题建模、上下文文本摘要以及理解人类复杂搜索活动方法的最新技术水平。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

The Athena project will develop technology called "search as learning," a set of search technologies that encourage and support learning rather than just simple document finding. In order to learn, searchers must engage with information that is both novel and understandable. Therefore, at the core, Athena will support learning by modeling several important factors: (1) the knowledge connections between documents covering a topic, (2) a user's current state of knowledge on that topic, (3) the types of knowledge a user is likely to gain from a document, and (4) the knowledge required for a user to successfully engage with a document. The Athena project will involve two types of end-to-end systems, both of which will model and leverage the learner's state of knowledge (LSK): an LSK-aware search engine and an LSK-aware question answering system. The Athena systems will guide a user through a topic and find relevant information in the context of previously encountered information and the topic structure captured in a web of topics. The team will evaluate Athena using standard measures as well as a series of studies involving human subjects. If the Athena project is successful, it will make it easier for people to use search engines and related technologies to learn about complex topics, where there are numerous interrelated and dependent subtopics that should be considered. Given that search is among the most common online activities on and off the Web, Athena and its technologies will have a substantial impact on searchers trying to learn such topics.Athena enables "search as learning" using a data structure referred to as a Learning Flow Graph (LFG). An LFG comprises nodes that represent sub-topics (e.g., concepts) within a given domain and vertices that represent relations between sub-topics (e.g., one sub-topic being foundational to understand another). Athena leverages LFGs to model the different factors mentioned above. It uses probability distributions across nodes in an LFG to model: (1) a user's knowledge state, (2) the potential knowledge gains from an information item, and (3) the prerequisite knowledge required for a user to successfully engage with an information item. The Athena team will develop algorithms for generating LFGs from structured and semi- and unstructured resources (e.g., course syllabi, tables of contents, book indices, knowledge bases, query logs), algorithms for integrating LFGs into search and question-answering models, and algorithms for re-estimating LFGs and a user's knowledge state based on search behaviors (e.g., queries, clicks, skips, dwell times, etc.). Structuring textual data to find the optimal learning paths through it is of great interest, though most existing work has focused on extracting information to fill slots in a "knowledge base," a much finer grained task. The LFG representation also provides a type of explanation of a larger topic, connecting to the broad interest in explainable systems. The Athena work will extend the state of the art in text representation, neural approaches including attention techniques, query and topic modeling, contextual text summarization, and understanding human approaches to complex search activities.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.