CAREER: Harnessing Interpersonal Common Sense for Social Grounding in Natural Language Processing

职业：利用人际常识来实现自然语言处理的社会基础

• 批准号: 2047232
• 负责人: Snigdha Chaturvedi
• 金额: $ 54.97万
• 依托单位: University of North Carolina at Chapel Hill
• 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=2047232&HistoricalAwards=false
• 关键词: CAREER; Harnessing; Interpersonal; Common; Sense

When interacting with other individuals, human behavior significantly depends on the relationship between the interactants. For example, when speaking to someone with a higher social status, people often exhibit language coordination by mimicking the linguistic style of the other speaker. Even children as young as 12 to 18-months old can adjust their behavior depending on who they are with. In other words, humans possess and employ interpersonal common sense: common-sense knowledge of the behavior acceptable in different interpersonal relationships, and use it in their day-to-day interactions. In contrast, computers lack this interpersonal common-sense knowledge. In order for computers to model the social behavior exhibited by humans and operate in a human-like manner, they need interpersonal common sense. This need to equip computers with this capacity has become even more important in recent times with technology, such as artificial conversational agents and robots, becoming increasingly pervasive in our day-to-day lives. The goal of this CAREER project is to instill interpersonal common-sense knowledge and reasoning capabilities in computers. To achieve this goal the project develops resources that store interpersonal common-sense knowledge together with techniques to leverage them for designing computer systems that are more aware of social dynamics prevalent in the human world. The project involves interdisciplinary efforts by researchers and students from within and outside of Computer Science. It includes developing interdisciplinary courses and seminars for graduate and undergraduate students in Computer Science, Linguistics, and Psychology. It also involves organizing workshops, demos and talks for attracting historically underrepresented minorities like women to computer science. This project develops technologies that will help Natural Language Processing (NLP) systems to acquire and incorporate interpersonal common sense in their functioning. The project’s efforts are divided into three thrusts. The first develops a knowledge base of human-annotated instances of interpersonal common-sense knowledge. Given a text excerpt containing interaction between two entities, the knowledge base will contain annotations about various facets of interpersonal inferences about the two entities. The second develops methods for continually and automatically expanding the knowledge base and generalizing to unseen situations. These methods are based on multi-task learning in order to automatically and jointly infer the various facets of interpersonal common sense, including those about unseen scenarios. The inferences, in turn, are also used to improve the methods in a semi-supervised framework. The third utilizes this interpersonal common-sense knowledge to improve NLP systems like those for dialog generation, summarization and information extraction for digital humanities. These NLP systems learn about interpersonal common sense from the knowledge base and utilize it while focusing on the entities mentioned in the text in order to improve the downstream task. Each thrust is accompanied by extensive and continual evaluations of the developed techniques. The research will result in publicly available resources, data and technologies for others in the field to use and train their systems on. Overall, this project pushes research in the direction of designing more socially cognizant and human-like NLP systems.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.

当与其他个体互动时，人类的行为在很大程度上取决于互动者之间的关系。例如，当与社会地位较高的人交谈时，人们经常通过模仿对方的语言风格来表现语言协调。即使是12到18个月大的孩子也能根据和谁在一起而调整自己的行为。换句话说，人类拥有并运用人际常识：关于不同人际关系中可接受的行为的常识性知识，并将其应用于日常交往中。相比之下，计算机缺乏这种人际交往的常识。为了让计算机模拟人类表现出的社会行为，并以类似人类的方式运作，它们需要人际常识。近年来，随着人工对话代理和机器人等技术在我们的日常生活中变得越来越普遍，为计算机配备这种能力的需求变得更加重要。这个CAREER项目的目标是向计算机灌输人际常识和推理能力。为了实现这一目标，该项目开发了存储人际常识性知识的资源，以及利用这些知识设计计算机系统的技术，使计算机系统更加了解人类世界中普遍存在的社会动态。该项目涉及计算机科学内外的研究人员和学生的跨学科努力。它包括为计算机科学、语言学和心理学的研究生和本科生开发跨学科课程和研讨会。它还包括组织研讨会、演示和讲座，以吸引女性等历史上代表性不足的少数群体参与计算机科学。该项目开发的技术将帮助自然语言处理（NLP）系统在其功能中获取和整合人际常识。该项目的工作分为三个重点。第一个开发了一个由人类注释的人际常识知识实例组成的知识库。给定包含两个实体之间交互的文本摘录，知识库将包含关于两个实体的人际推理的各个方面的注释。第二部分开发了持续自动扩展知识库并将其推广到未知情况的方法。这些方法是基于多任务学习，以自动和联合推断人际常识的各个方面，包括那些看不见的场景。反过来，这些推论也用于改进半监督框架中的方法。第三种是利用这种人际常识知识来改进NLP系统，如数字人文学科的对话生成、摘要和信息提取系统。这些NLP系统从知识库中学习人际常识，并在关注文本中提到的实体的同时利用它来改进下游任务。每次推进都伴随着对已开发技术的广泛和持续的评价。这项研究将产生可公开获得的资源、数据和技术，供该领域的其他人使用和训练他们的系统。总的来说，这个项目推动了研究的方向，设计更多的社会认知和类似人类的NLP系统。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。

项目成果

"Let Your Characters Tell Their Story": A Dataset for Character-Centric Narrative Understanding

• DOI: 10.18653/v1/2021.findings-emnlp.150
• 发表时间: 2021-09
• 作者: Faeze Brahman;Meng Huang-;Oyvind Tafjord;Chao Zhao;Mrinmaya Sachan;Snigdha Chaturvedi

Grounded Keys-to-Text Generation: Towards Factual Open-Ended Generation

接地键文本生成：迈向事实开放式生成

• DOI: 10.18653/v1/2022.findings-emnlp.547
• 发表时间: 2022
• 期刊: Proceedings of the Findings of the 2022 Conference on Empirical Methods in Natural Language Processing
• 作者: Brahman, Faeze;Peng, Baolin;Galley, Michel;Rao, Sudha;Dolan, Bill;Chaturvedi, Snigdha;Gao, Jianfeng
• 通讯作者: Gao, Jianfeng

NarraSum: A Large-Scale Dataset for Abstractive Narrative Summarization

NarraSum：用于抽象叙事摘要的大型数据集

• DOI: 10.18653/v1/2022.findings-emnlp.14
• 发表时间: 2022
• 期刊: Proceedings of the Findings of the 2022 Conference on Empirical Methods in Natural Language Processing
• 作者: Zhao, Chao;Brahman, Faeze;Song, Kaiqiang;Yao, Wenlin;Yu, Dian;Chaturvedi, Snigdha
• 通讯作者: Chaturvedi, Snigdha