CRII: RI: Learning a Timely Semantic Resource from Social Media Data

文本挖掘和自然语言处理研究中的一个关键挑战是，单一的含义可以用许多不同的方式来表达，即释义。从信息检索、信息提取和自然语言生成到IBM的Watson、Google的Knowledge Graph等等，大量的释义资源已经取得了稳步进展，其应用也有了显著的增长。本研究旨在创造更好的释义习得技术和更大规模的语义资源，可用于社会科学、国家安全等相关领域的各种自然语言处理任务和社交媒体数据分析。一个潜在的应用例子是文本简化，它自动将复杂的文本重组成更简单的语言，供儿童或有阅读障碍的人使用。本研究的技术创新集中在句子对之间单词和短语级别的对齐的联合建模，以应对从非正式数据源(如社交媒体)提取语义知识的挑战，这些数据源存在的数量非常大，而不是像以前的工作那样仅存在于正式来源，如News swire。模型设计通过图形模型和神经网络两种方法扩展了多实例学习，并可以灵活地允许探索不同的假设和对单词或短语的重要性进行建模。建模的进步可以推广到其他自然语言理解任务，这些任务需要根据给定上下文中的词级组成或词义来分析句子，以及自然语言生成任务，这些任务受益于学习哪些单词和短语应该删除或重新措辞。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。

One key challenge in text mining and natural language processing research is that a single meaning can be expressed in many different ways, i.e., paraphrases. There has been steady progress towards large paraphrase resources, and a significant increase in its applications: from information retrieval, information extraction, and natural language generation to IBM's Watson, Google's Knowledge Graph, and many more. This research aims to create better paraphrase acquisition techniques and larger scale semantic resources, which could be of great use in various natural language processing tasks and social media data analytics in social science, national security, and other related fields. One example of potential applications is text simplification, which automatically rephrases complex texts into simpler language for children or people with reading disabilities.The technical innovation of this study focuses on joint modeling of word- and phrase-level alignments between sentence pairs to address the challenges of extracting semantic knowledge from informal data sources (such as social media), which exist in very large quantities rather than just formal sources, such as newswire as per previous work. The model design extends multiple instance learning via two methods, a graphical model and neural network, and can flexibly permit the exploration of different assumptions and models the importance of words or phrases. The modeling advancements can be generalized to other natural language understanding tasks, which require analyzing sentences based on word-level composition or word meaning in a given context, and natural language generation tasks that benefit from learning what words and phrases to remove or rephrase.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.