RI: Small: Robust Models for Sequence Labelling in Social Media Data

在过去的十年里，社交媒体平台对人们交流方式的影响越来越大；这些平台现在被认为是人们广泛使用的重要交流工具，人们不仅可以分享信息，还可以了解任何话题的最新事件。因此，通过这些平台运行的信息，由用户、公司、媒体和政治实体产生，对于理解时事、行为等非常相关，而这些数据的自动提炼具有很大的实用价值。当前的文本处理技术无法在社交媒体数据上准确地执行信息提取，因为这些复杂的算法是针对高度编辑的英文文本进行的，主题集很窄，例如Newswire数据中的主题。相比之下，社交媒体数据的语法流畅，词汇量非常大，主题不受限制，并且包括经常在同一文本中混合使用的多种语言。该项目解决了来自社交媒体来源的自动化处理所涉及的许多挑战。此外，研究小组将开发和发布新的注释数据，使这一方向的新研究成为可能。此外，这个项目将通过支持研究生和本科生来扩大对计算机科学的参与。这个项目的基本前提是，表征学习与语言和领域知识的更紧密耦合将允许模型通过提取每个单独文本中的所有相关语言抽象来学习任务，而不需要像端到端深度学习模型中典型的情况那样，需要令人望而却步的大量标记数据。该奖项将为序列标记任务设计强大的方法，可以通过双管齐下的方法分析社交媒体数据。首先，研究团队将研究社交媒体数据带来的挑战及其与预测表现的相关性。然后，研究人员将为序列标注任务设计新的模型架构，其中领域和语言知识监督学习过程。对建议模型的评估将包括来自不同社交媒体来源的数据。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In the last decade social media platforms have increased their impact on the way people communicate; these platforms are now considered an essential communication tool that people use broadly to share information, but also to get informed about the latest events on any topic. Consequently, the information running through those platforms, generated by users, companies, the media, and political entities, is extremely relevant to understand current events, behaviors, and more, and the automated distillation of this data is of great practical value. Current technology for text processing fails to perform information extraction accurately on social media data since these sophisticated algorithms have been trained on highly edited English text with a narrow set of topics, such as that in newswire data. In contrast, social media data has a fluid grammar, a very large vocabulary, unlimited topics, and includes multiple languages that are often mixed in the same text. This project addresses the many challenges involved in the automated processing from social media sources. Additionally, the research team will develop and release new annotated data that will enable new research in this direction. Furthermore, this project will address broadening participation in computer science by supporting graduate and undergraduate students, several of them from underrepresented groups in Computer Science.The underlying premise of this project is that a tighter coupling of representation learning with linguistic and domain knowledge will allow the models to learn the tasks by distilling all relevant linguistic abstractions in each single text, without requiring prohibitively large amounts of labeled data, as is typically the case in end-to-end deep-learning models. This award will design robust approaches for sequence labeling tasks that can analyze social media data with a two-pronged approach. First, the research team will study the challenges imposed by social media data and their correlation to prediction performance. Then the investigators will design new model architectures for sequence labeling tasks where domain and linguistic knowledge supervise the learning process. The evaluation of the proposed models will include data from different social media sources.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.