EXP: Development of Human Language Technologies to Improve Disciplinary Writing and Learning through Self-Regulated Revising

写作和修改是学习的重要组成部分，然而许多大学生毕业时并没有表现出对学术写作的改善或掌握。本项目探索通过整合自然语言处理方法、数据可视化和用户界面的最佳实践以及当前教学理论的修订环境来提高学生学术写作的可行性。这种环境将支持和鼓励学生发展写作和修改所必需的自我调节技能，包括目标设定、写作策略的选择和进度的自我监控。作为一种学习技术，环境可以大规模地应用，从而提高包括英语学习者在内的不同学生群体的写作水平。此外，该项目的研究生多学科培训侧重于增加网络学习研究和开发的多样性。计划分三个阶段进行调查。首先，为了分析学生的复习行为数据，我们进行了一系列的实验来研究学生之间的互动和复习写作环境的变化。其次，收集到的数据形成了开发端到端系统的黄金标准，该系统可以自动提取学生草稿之间的修订，并确定每次修订的目标。考虑了多种提取算法，包括基于语义相似度度量和深度学习方法的短语对齐。为了确定修订的目标，从金标准中训练监督分类器。研究了一组不同的特征和已确定目标的表示（例如，粒度、范围）。除了“提取-分类”管道外，还开发了一种替代联合序列标记模型。序列的标记用于识别修订目标，序列被突变以产生可能的句子对齐更正，用于修订提取。编写环境被迭代地改进，通过频繁的用户研究来增加界面原型。第三，一个完整的端到端系统，集成了最成功的组件模型，部署在大学水平的写作课程。通过多个作业跟踪学生的进度。

Writing and revising are essential parts of learning, yet many college students graduate withoutdemonstrating improvement or mastery of academic writing. This project explores the feasibilityof improving students' academic writing through a revision environment that integrates naturallanguage processing methods, best practices in data visualization and user interfaces, andcurrent pedagogical theories. The environment will support and encourage students to developself-regulation skills that are necessary for writing and revising, including goal-setting, selectionof writing strategies, and self-monitoring of progress. As a learning technology, the environmentcan be applied on a large scale, thereby improving the writing of diverse student populations,including English learners. Additionally, the project's multidisciplinary training of graduatestudents is focused on increasing diversity in cyberlearning research and development. Three stages of investigation are planned. First, to analyze data on students' revision behaviors,a series of experiments are conducted to study interactions between studentsand variations of the revision writing environment. Second, the collected data forms the goldstandard for developing an end-to-end system that automatically extracts revisions betweenstudent drafts and identifies the goal for each revision. Multiple extraction algorithms areconsidered, including phrasal alignment based on semantic similarity metrics and deep learningapproaches. To identify the goal of a revision, a supervised classifier is trained from the goldstandard. A diverse set of features and the representations of the identified goals (e.g.,granularity, scope) are explored. In addition to the "extract-then-classify" pipeline, an alternativejoint sequence labeling model is also developed. The labeling of sequences is used torecognize revision goals and the sequences are mutated to generate possible corrections ofsentence alignments for revision extraction. The writing environment is iteratively refined,augmenting the interface prototyping through frequent user studies. Third, a completeend-to-end system that integrates the most successful component models is deployed incollege-level writing classes. Student progress is tracked across multiple assignments.