EAGER: Machine learning of discourse structure for personalized online tutoring

教育中的技术既提供了巨大的机遇，也带来了巨大的挑战。通过在线授课，我们可能能够接触到比以前更广泛的学生。但是，虽然技术允许教学接触到更多的学生，但它也有可能失去面对面教学提供的丰富互动和反馈。要想真正有效，课程既需要丰富的内容，也需要个性化的互动和反馈。因此，该项目旨在为将更丰富的交互恢复到在线学习体验提供技术基础，特别是通过讨论板和同伴评分等自然语言内容。该项目将开发机器学习技术，以发现自然语言文本中的高级结构，包括语篇结构(文本之间的关系)、主题结构和语义结构。为了发现这种结构，该项目调查了光谱学习方法的使用。这些方法依赖于分解观测矩的矩阵或张量，能够有效地学习潜在结构，并且没有局部最优。尤其令人感兴趣和挑战的是开发出易于处理的方法，可以从未标记或弱标记的数据中学习潜在结构。当前的许多自然语言处理技术在非正式语言方面存在困难，并且不能提取更高层次的结构；我们的目标是开发频谱方法来解决这些弱点。该项目将试图从大型在线课堂的讨论板上获取的数据中学习这种高水平的潜在结构。我们的想法是，未来的工作可以利用这种理解来帮助学生个性化访问这种丰富的自然语言内容的方式：例如，通过帮助聚焦搜索，找到相关和连贯的内容，合成学生问题的答案，促进在线讨论中的富有成效的行为，以及支持自我和同行评分。最终目标是为学生提供个性化的界面，方便他们学习新材料。

Technology in education provides both tremendous opportunities and tremendous challenges.  By delivering instruction online, we may be able to reach a much wider range of students than previously possible.  But, while technology allows instruction to reach more students, it also risks losing the rich interaction and feedback that in-person instruction provides.  To be truly effective, courses need both rich content and personalized interaction and feedback. So, this project seeks to provide a technical foundation for restoring richer interaction to the online learning experience, particularly through natural language content such as discussion boards and peer grading.The project will develop machine learning techniques for discovering high-level structure in natural language text, including discourse structure (relationships among pieces of text), topic structure, and semantic structure.  To discover this structure, the project investigates the use of spectral learning methods.  These methods, which rely on factoring a matrix or tensor of observed moments, are able to learn latent structures efficiently and without local optima.  Of particular interest and challenge is to develop tractable methods that can learn latent structure from unlabeled or weakly-labeled data.  Many current NLP techniques have difficulty with informal language and do not extract higher level structure; our goal is to develop spectral methods to address these weaknesses. The project will attempt to learn this high-level latent structure in data taken from discussion boards of large online classes.  The idea is that future work could use this understanding to help personalize the way students access this sort of rich natural language content: e.g., by helping to focus searches, find relevant and cohesive content, synthesize answers to student questions, promote productive behaviors in online discussions, and support self and peer grading.  The eventual goal is to provide personalized interfaces to students that facilitate their learning of novel material.