III: Small: Intelligent Scientific Text Analytics with Knowledge-Augmented Abductive Reasoning

III：小：具有知识增强归纳推理的智能科学文本分析

• 批准号: 2234058
• 负责人: Meng Jiang
• 金额: $ 60万
• 依托单位: University of Notre Dame
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-06-01 至 2026-05-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2234058&HistoricalAwards=false
• 关键词: III; Small; Intelligent; Scientific; Text

Scientists are producing vast numbers of research articles and patents every year to advance our understanding of our world and the universe. Meanwhile, they are making a great effort to build tools to boost their productivity. While these tools are able to process scientific text, they are not endowed with intelligence to think or write like scientists to help their work. Natural language generation systems may generate some new statements that are fluent to read and hard to distinguish from human-written texts. However, existing systems are not as intelligent or reliable as working with human research assistants due to lack of reasoning abilities about scientific innovation. This project aims to enable comparative reasoning in a novel intelligent system of scientific text analytics, which is missing in existing systems. Comparative reasoning establishes the importance of something by comparing it against something else. Comparative reasoning plays a central role in scientific innovation and can be categorized as abductive reasoning in the context of artificial intelligence. This project will design and develop novel text generation approaches for scientific abductive reasoning and intelligent scientific text analytics. Moreover, this research will support the professional development of a cohort of PhD, undergraduate, and high school students.The technical aims of the project are divided into three thrusts. The first develops and compares natural language generation models based on a data-driven architecture and a novel architecture inspired and rooted in theories of abduction. These models will be evaluated on the tasks of comparative summarization and comparative argument generation in scientific domains. The second thrust designs retrieval-augmented approaches with heterogeneous knowledge sources such as tables, taxonomies, and knowledge graphs to improve the performance of scientific abductive reasoning models. Because retrieving and encoding every instance can be very time consuming, the third thrust builds knowledge memory networks that learns and manages distributed representations of scientific concepts and relations from the knowledge sources. They will accelerate the retrieval augmentation, when all the types of scientific source data are of large scale. Finally, these techniques will be integrated into a new artificial intelligence system that accurately generates explanatory sentences to automate comparative reasoning and assist scientific innovation.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.

科学家每年都会发表大量研究文章和专利，以增进我们对世界和宇宙的理解。与此同时，他们正在努力构建工具来提高生产力。虽然这些工具能够处理科学文本，但它们不具备像科学家一样思考或写作的智能来帮助他们的工作。自然语言生成系统可能会生成一些易于阅读且难以与人类编写的文本区分开的新语句。然而，由于缺乏科学创新的推理能力，现有系统不像人类研究助理那样智能或可靠。该项目旨在在科学文本分析的新型智能系统中实现比较推理，这是现有系统所缺少的。比较推理通过将某事物与其他事物进行比较来确定其重要性。比较推理在科学创新中发挥着核心作用，可以归类为人工智能背景下的溯因推理。该项目将设计和开发用于科学溯因推理和智能科学文本分析的新颖文本生成方法。此外，这项研究还将支持一批博士生、本科生和高中生的专业发展。该项目的技术目标分为三个主旨。第一个模型开发并比较了基于数据驱动架构和受溯因理论启发和植根的新颖架构的自然语言生成模型。这些模型将根据科学领域的比较总结和比较论证生成的任务进行评估。第二个主旨是利用表格、分类法和知识图等异构知识源设计检索增强方法，以提高科学溯因推理模型的性能。由于检索和编码每个实例可能非常耗时，因此第三个推动力构建了知识记忆网络，该网络从知识源学习和管理科学概念和关系的分布式表示。当所有类型的科学源数据规模都很大时，它们将加速检索增强。最后，这些技术将被集成到一个新的人工智能系统中，该系统可以准确地生成解释性句子，以自动进行比较推理并协助科学创新。该奖项反映了 NSF 的法定使命，并通过利用基金会的智力价值和更广泛的影响进行评估，认为值得支持审查标准。

项目成果

Data-Centric Learning from Unlabeled Graphs with Diffusion Model

使用扩散模型从未标记图进行以数据为中心的学习

• 发表时间: 2023-12
• 期刊: NeurIPS
• 作者: Liu, Gang;Inae, Eric;Zhao, Tong;Xu, Jiaxin;Luo, Tengfei;Jiang, Meng
• 通讯作者: Jiang, Meng

Completing Taxonomies with Relation-Aware Mutual Attentions

通过关系感知的相互关注完成分类

• 发表时间: 2023-07
• 期刊: KDD
• 作者: Zeng, Qingkai;Zhang, Zhihan;Lin, Jinfeng;Jiang, Meng
• 通讯作者: Jiang, Meng

Pre-training Language Models for Comparative Reasoning

用于比较推理的预训练语言模型

• DOI: 10.18653/v1/2023.emnlp-main.763
• 发表时间: 2023-01
• 期刊: EMNLP
• 作者: Yu, Mengxia;Zhang, Zhihan;Yu, Wenhao;Jiang, Meng
• 通讯作者: Jiang, Meng

IfQA: A Dataset for Open-domain Question Answering under Counterfactual Presuppositions

IfQA：反事实预设下的开放域问答数据集

• DOI: 10.18653/v1/2023.emnlp-main.515
• 发表时间: 2023-01
• 期刊: EMNLP
• 作者: Yu, Wenhao;Jiang, Meng;Clark, Peter;Sabharwal, Ashish
• 通讯作者: Sabharwal, Ashish

Auto-Instruct: Automatic Instruction Generation and Ranking for Black-Box Language Models

自动指令：黑盒语言模型的自动指令生成和排序

• DOI: 10.18653/v1/2023.findings-emnlp.659
• 发表时间: 2023-01
• 期刊: EMNLP
• 作者: Zhang, Zhihan;Wang, Shuohang;Yu, Wenhao;Xu, Yichong;Iter, Dan;Zeng, Qingkai;Liu, Yang;Zhu, Chenguang;Jiang, Meng
• 通讯作者: Jiang, Meng