CAREER: Knowledge-Rich Neural Text Comprehension and Reasoning

职业：知识丰富的神经文本理解和推理

• 批准号: 2044660
• 负责人: Hanna Hajishirzi
• 金额: $ 54.98万
• 依托单位: University of Washington
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-09-01 至 2026-08-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2044660&HistoricalAwards=false
• 关键词: CAREER; Knowledge; Rich; Neural; Text

Enormous amounts of ever-changing knowledge are available online in diverse textual styles (e.g., news vs. science text) and diverse formats (knowledge bases vs. web pages vs. textual documents). This proposal addresses the question of textual comprehension and reasoning given this diversity: how can artificial intelligence (AI) help applications comprehend and combine evidence from variable, evolving sources of textual knowledge to make complex inferences and draw logical conclusions? Recent advances in deep learning algorithms, large-scale datasets, and industry-scale computational resources are spurring progress in many Natural Language Processing (NLP) tasks, including question answering. Nevertheless, current models lack the ability to answer complex questions that require them to reason intelligently across diverse sources and explain their decisions. Further, these models cannot scale up when task-annotated training data are scarce and computational resources are limited. Our results will give rise to the next generation of question answering and fact checking algorithms that offer rich natural language comprehension using multi-hop and interpretable reasoning even when annotated training data is scarce. With a focus on textual comprehension and reasoning, this research will integrate capabilities of symbolic AI approaches into current deep learning algorithms. It will devise hybrid, interpretable algorithms that understand and reason about textual knowledge across varied formats and styles, generalize to emerging domains with scarce training data (are robust), and operate efficiently under resource limitations (are scalable). Toward this end, this research will focus on four transformative research initiatives: (1) defining a general-purpose formalism to promote data comprehension through knowledge-rich neural representations, (2) devising an interpretable, multi-hop inference and reasoning engine, (3) developing robust and scalable algorithms to demonstrate generalizable domain and device adaptation, and (4) building applications and datasets in question answering and fact checking tasks that will have lasting general-purpose utility.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.

大量不断变化的知识可以在网上以不同的文本风格（例如，新闻与科学文本）和多种格式（知识库与网页与文本文档）。该提案解决了文本理解和推理的问题，考虑到这种多样性：人工智能（AI）如何帮助应用程序理解和联合收割机证据从可变的，不断发展的文本知识来源，使复杂的推理和得出逻辑结论？ 深度学习算法、大规模数据集和行业规模计算资源的最新进展正在推动许多自然语言处理（NLP）任务的进展，包括问答。然而，目前的模型缺乏回答复杂问题的能力，这些问题需要他们在不同的来源中进行智能推理并解释他们的决策。 此外，当任务注释的训练数据稀缺且计算资源有限时，这些模型无法扩展。我们的研究结果将产生下一代的问题回答和事实检查算法，即使在注释的训练数据很少的情况下，这些算法也可以使用多跳和可解释的推理来提供丰富的自然语言理解。 这项研究的重点是文本理解和推理，它将把符号人工智能方法的能力整合到当前的深度学习算法中。它将设计出混合的、可解释的算法，这些算法可以理解和推理不同格式和风格的文本知识，推广到具有稀缺训练数据的新兴领域（鲁棒性），并在资源限制下有效运行（可扩展性）。为此，本研究将重点关注四项变革性研究计划：（1）定义通用形式体系以通过知识丰富的神经表示来促进数据理解，（2）设计可解释的、多跳的推理和推理引擎，（3）开发鲁棒的和可扩展的算法以展示可推广的域和设备适配，以及（4）在问答和事实核查任务中构建具有持久通用用途的应用程序和数据集。该奖项反映了NSF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。

项目成果

InSCIt : Information-Seeking Conversations with Mixed-Initiative Interactions

InSCIt：具有混合主动交互的信息寻求对话

• DOI: 10.1162/tacl_a_00559
• 发表时间: 2023
• 期刊: Transactions of the Association for Computational Linguistics
• 影响因子: 10.9
• 作者: Wu, Zeqiu;Parish, Ryu;Cheng, Hao;Min, Sewon;Ammanabrolu, Prithviraj;Ostendorf, Mari;Hajishirzi, Hannaneh
• 通讯作者: Hajishirzi, Hannaneh

When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories

• DOI: 10.18653/v1/2023.acl-long.546
• 发表时间: 2022-12
• 作者: Alex Troy Mallen;Akari Asai;Victor Zhong;Rajarshi Das;Hannaneh Hajishirzi;Daniel Khashabi

Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks

• DOI: 10.18653/v1/2022.emnlp-main.340
• 发表时间: 2022-04
• 作者: Yizhong Wang;Swaroop Mishra;Pegah Alipoormolabashi;Yeganeh Kordi;Amirreza Mirzaei;Anjana Arunkumar;Arjun Ashok;Arut Selvan Dhanasekaran;Atharva Naik;David Stap;Eshaan Pathak;Giannis Karamanolakis;H. Lai;I. Purohit;Ishani Mondal;Jacob Anderson;Kirby Kuznia;Krima Doshi;Maitreya Patel;Kuntal Kumar Pal;M. Moradshahi;Mihir Parmar;Mirali Purohit;Neeraj Varshney;Phani Rohitha Kaza;Pulkit Verma;Ravsehaj Singh Puri;Rushang Karia;Shailaja Keyur Sampat;Savan Doshi;Siddhartha Mishra;Sujan Reddy;Sumanta Patro;Tanay Dixit;Xudong Shen;Chitta Baral;Yejin Choi;Noah A. Smith;Hannaneh Hajishirzi;Daniel Khashabi

Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?

• DOI: 10.18653/v1/2022.emnlp-main.759
• 发表时间: 2022-02
• 期刊: ArXiv
• 作者: Sewon Min;Xinxi Lyu;Ari Holtzman;Mikel Artetxe;M. Lewis;Hannaneh Hajishirzi;Luke Zettlemoyer

Self-Instruct: Aligning Language Models with Self-Generated Instructions

自指导：使语言模型与自生成的指令保持一致

• DOI: 10.18653/v1/2023.acl-long.754
• 发表时间: 2023
• 期刊: ACL
• 作者: Wang, Yizhong;Kordi, Yeganeh;Mishra, Swaroop;Liu, Alisa;Smith, Noah A.;Khashabi, Daniel;Hajishirzi, Hannaneh
• 通讯作者: Hajishirzi, Hannaneh