CAREER: Building Creative Writing Assistants for Machine-in-the-Loop Storytelling

职业：为机器在环讲故事构建创意写作助手

• 批准号: 2046248
• 负责人: Mohit Iyyer
• 金额: $ 53.23万
• 依托单位: University of Massachusetts Amherst
• 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=2046248&HistoricalAwards=false
• 关键词: CAREER; Building; Creative; Writing; Assistants

This CAREER project focuses on building text generation systems that interact with people to improve their writing and also help them learn to write. Such “machine-in-the-loop” writing assistants are potentially transformative technologies for improving the writing quality and productivity of human authors, as well as providing new tools for writing pedagogy through cyberlearning applications. However, they have been relatively underexplored by the natural language processing community due to major difficulties in modeling, evaluation, and data collection. The technologies developed in this project address these challenges by (1) developing platforms that leverage existing online author communities to enable the design and evaluation of machine-in-the-loop writing assistants; (2) advancing text generation modeling to improve the quality of generated text; and (3) enabling assistants to rewrite and reorganize human-authored text through developments in automatic paraphrasing. In addition to aiding authors in online communities, the writing assistants developed through the project are deployed in K-12 classrooms to advance writing pedagogy. The project incorporates undergraduate students, including those outside of computer science, in natural language processing research, and provides significant outreach to underrepresented minorities.To make meaningful progress on the development of machine-in-the-loop writing assistants, the project includes a collaboration with Protagonist Labs, which runs online platforms for collaborative storytelling in both creative and pedagogical settings and already has incorporated systems built by the investigator’s team into user-facing interfaces. User interaction on such platforms allows fine-grained evaluation of novel text generation methods, which include neural language models that integrate context compression, context retrieval, and discrete latent variables into the generation process to improve overall coherence and relevance. In addition to producing new text, fully-featured writing assistants must also be able to rewrite user text into a specified form, such as a target writing style or suitability for a target audience. To this end, the project introduces new paraphrase generation models at a variety of units of text, including phrases, sentences, and paragraphs. This research effort aims to spur research into interactive text generation systems, and as such its outputs will include publicly-released pretrained models and open-sourced code.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.

这个CAREER项目的重点是建立与人们互动的文本生成系统，以提高他们的写作水平，并帮助他们学习写作。这种“机器在循环”的写作助手是潜在的变革性技术，可以提高人类作者的写作质量和生产力，并通过网络学习应用程序为写作教学提供新的工具。然而，由于建模、评估和数据收集方面的主要困难，自然语言处理社区对它们的探索相对不足。本项目开发的技术通过以下方式解决了这些挑战：(1)开发利用现有在线作者社区的平台，使机器在循环写作助手的设计和评估成为可能；(2)推进文本生成建模，提高生成文本的质量；(3)通过自动释义的发展，使助手能够重写和重组人类撰写的文本。除了帮助在线社区的作者之外，通过该项目开发的写作助手还被部署在K-12教室中，以推进写作教学。该项目将本科生（包括计算机科学以外的本科生）纳入自然语言处理研究，并为未被充分代表的少数民族提供重要的外展服务。为了在机器在循环写作助手的开发上取得有意义的进展，该项目包括与主角实验室的合作，该实验室在创造性和教学设置中运行在线平台，用于协作讲故事，并且已经将研究者团队构建的系统整合到面向用户的界面中。这些平台上的用户交互允许对新的文本生成方法进行细粒度评估，其中包括将上下文压缩、上下文检索和离散潜在变量集成到生成过程中的神经语言模型，以提高整体一致性和相关性。除了生成新文本之外，功能齐全的写作助手还必须能够将用户文本改写为指定的形式，例如目标写作风格或适合目标受众。为此，该项目在各种文本单元（包括短语、句子和段落）上引入了新的释义生成模型。这项研究旨在促进对交互式文本生成系统的研究，因此其产出将包括公开发布的预训练模型和开源代码。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

SLING: Sino Linguistic Evaluation of Large Language Models

SLING：大型语言模型的汉语言学评估

• DOI: 10.18653/v1/2022.emnlp-main.305
• 发表时间: 2022
• 期刊: Empirical Methods in Natural Language Processing
• 作者: Song, Yixiao;Krishna, Kalpesh;Bhatt, Rajesh;Iyyer, Mohit
• 通讯作者: Iyyer, Mohit

Overcoming Catastrophic Forgetting in Zero-Shot Cross-Lingual Generation

克服零样本跨语言生成中的灾难性遗忘

• DOI: 10.18653/v1/2022.emnlp-main.630
• 发表时间: 2022
• 期刊: Empirical Methods in Natural Language Processing
• 作者: Vu, Tu;Barua, Aditya;Lester, Brian;Cer, Daniel;Iyyer, Mohit;Constant, Noah
• 通讯作者: Constant, Noah

DEMETR: Diagnosing Evaluation Metrics for Translation

DEMETR：诊断翻译评估指标

• DOI: 10.18653/v1/2022.emnlp-main.649
• 发表时间: 2022
• 期刊: Empirical Methods in Natural Language Processing
• 作者: Karpinska, Marzena;Raj, Nishant;Thai, Katherine;Song, Yixiao;Gupta, Ankita;Iyyer, Mohit
• 通讯作者: Iyyer, Mohit