CRII: III: Capturing Dynamism in Causal Relationships: A New Paradigm for Relationship Extraction from Text

文本挖掘在将大量非结构化文本数据转换为知识的方法方面取得了重要进展。然而，当前的关系提取范式有一个主要的局限性：它模拟了信息的快照，但未能捕捉到知识的基本对话和动态性质：相互冲突的发现、不一致的发现、反驳、矛盾、强化或确认，所有这些都随着时间的推移而变化。该项目旨在捕捉知识的基本动态，特别关注因果关系。尽管包括学术文章在内的许多文章都展示了表达因果关系的知识和关系，但这种关系不是静态的，可以随着时间的推移而随着条件的变化而变化。该项目的目标是从文本数据中识别因果知识的线索，量化因果关系的强度，并在不断变化的条件下建立其动态模型。最终，该项目旨在对从文本中提取的知识进行更全面的建模。由于文本数据被来自具有国家重要性的不同领域的研究人员和从业人员广泛使用，这些领域包括医学和卫生、经济学、公共政策、新闻，因此本项目的成果旨在为从业人员提供基础，以提供新的方法来理解大型文本数据集中存在的因果关系的演变性质。具体而言，将应用该项目制定的新方法来探索公共卫生数据，以确定气候、政治、经济条件的变化如何影响不同地理区域人口的身心健康。此外，将会有不同的教育活动作为这个项目的一部分-从这个项目中出现的和相关的主题将被包括在应用数据科学硕士课程的各种课程中；促进本科生研究，特别是招收来自代表性不足和经济上处于不利地位的社区的学生参与项目；组织一个研究研讨会，鼓励高中生参与STEM研究。项目活动包括开发一种新的因果关系提取模型，该模型利用了结合语义和语法线索的统一深度学习框架。这种方法将利用一个句子的关键语法特征，通过图形或树状模型来表示名词、动词和其他词性之间的语法关系。这项工作将确定句子是否具有表明因果关系的结构。此外，该模型的顺序组件将利用语义并识别句子中某些单词的影响，以表征文本中表达的因果关系的性质。这项任务将捕捉关系的强度（例如，使用“极有可能”，“肯定”等线索），任何支持或反对的证据（例如，“将导致”或“不导致”），并将识别条件线索（例如，“在场”）等。量化这些定性属性将导致本项目的第二个创新——因果距离。因果距离是一种时变度量，它将表示两个实体之间因果关系的大小，并通过随时间变化的条件或新证据修改自身来捕捉关系的动态。总的来说，本项目所取得的进展将进一步增强我们对新的计算方法的理解，这些方法需要挖掘和推理嵌入在大型文本数据集中的因果关系线索。该项目的成果，如数据集、源代码、最终软件、结果和出版物将通过可公开访问的url和在线代码存储库共享。此外，所有的项目资源和成果将在项目网站上公布。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Text mining made important advances in methods to convert vast and unstructured text data into knowledge. However, the current paradigm of relationship extraction has one major limitation: it models snapshots of information but fails to capture the fundamentally dialogic and dynamic nature of knowledge: conflicting findings, inconsistent discoveries, refutations, contradictions, reinforcements or confirmations, all changing over time. This project aims to capture such fundamental dynamics of knowledge, specifically focusing on causal relationships. Whereas numerous articles, including academic articles, present knowledge and relationships that express causality, such relationships are not static and can change over time due to changing conditions. The objective of this project is to identify cues of causal knowledge from text data, quantify the strength of the causal relationship, and model its dynamics over changing conditions. Ultimately, the project aims at modelling a more holistic view of the knowledge extracted from text. As text data is extensively used by researchers and practitioners from different domains of national importance, including, medicine and health, economics, public policy, journalism, the results of this project seek to provide the foundation to offer practitioners new ways to understand the evolving nature of the causal relationships present in large text datasets. Specifically, the novel approaches developed in the project will be applied to explore public health data to determine how changing climatic, political, economic conditions may affect the mental and physical health of the population in different geographic areas. In addition, there will be various educational activities as part of this project - emerging and related topics from this project will be included in the curricula of various courses in the applied data science master’s program; promote undergraduate research, specifically, recruit students to work in the project who are from underrepresented and economically disadvantaged communities; organize a research workshop to encourage participation of high school students in STEM research. The project activities include the development of a novel model of causal relationship extraction that leverages a unified deep learning framework combining both semantic and syntax cues. This approach will utilize the key syntactical features of a sentence represented by the grammar relationships between noun, verbs and other parts of speech through graphical or tree-like models. This work will determine whether the sentence features a structure that signals causality. Moreover, the sequential component of the model will utilize the semantics and identify the influence of certain words in the sentence to characterize the nature of the causal relationship expressed in the text. This task will capture the strength of the relationship (e.g., using cues like "extremely likely", "definitely"), any supporting or opposing evidences (e.g., "will lead to" or "does not lead to"), and will identify conditional cues (e.g., "in the presence of") etc. Quantifying such qualitative properties will lead to the second innovation of this project – causal distance. Causal distance is a time-variant metric that will denote the magnitude of causality between two entities as well as capture the dynamism of the relationship by modifying itself over time with changing conditions or new evidences. Collectively, the advances pursued in this projects will further enhance our understanding of the novel computational approaches needed to unearth and reason on cues of causal relationships embedded in large text data sets. The outcomes of this project, such as datasets, source code, final software, results and publications will be shared via publicly accessible URLs and online code repositories. Additionally, all the project resources and outcomes will be made available on the project website.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.