CAREER: Data-Driven Document Generation

机器生成的语言越来越多地用于日常交互，从智能助手等消费者应用程序到简化复杂文档的辅助工具。这个CAREER项目旨在改进文本生成方法，以确保它们在生成超过几个句子的文档时能够一致地工作。研究者将研究基于机器学习的文本生成方法，旨在确保文档结构一致，适当引用人物和地点，并正确表达事实信息。该合同还将支持开源工具的开发，使其更容易被其他人使用，并将这些方法扩展到各种文本生成应用程序。这项研究将通过大学机器学习教学的新方法融入教育，为计算机科学中代表性不足的学生提供指导，并向可能没有接触过人工智能和数据科学作为研究领域的小学生提供服务。这个项目的重点是数据驱动的自然语言生成与深度学习。近年来，该领域在使用新的机器学习方法来学习基于人类示例生成文本方面取得了进展；然而，这项工作还远远达不到人类的水平，尤其是在生成文档级文本时。该项目将使用深度生成模型开发新的机器学习方法。其中包括使用深度隐马尔可夫模型学习和控制文档结构，通过生成参考和对齐来确定内容，以及通过自下而上的选择将内容聚合到文本中。其中每一个都结合了机器学习方法和自然语言分析的研究。该项目的目标是开发一个文本生成方法的开源实现，允许用户轻松地定位新领域并评估其系统的保真度和一致性。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。

Increasingly, machine-generated language is being used in everyday interactions, from consumer applications such as intelligent assistants to accessibility tools for simplifying complex documents. This CAREER project seeks to improve methods for text generation to ensure they work consistently when producing documents longer than a few sentences. The investigator will study machine learning-based methods for text generation that aim to ensure that document structure is consistent, that people and places are referred to appropriately, and that factual information is correctly expressed. The award will additionally support the development of open-source tools to make it easier for others to use and extend these approaches for a variety of text generation application. This research will be integrated into education through new methods for university teaching of machine learning, mentoring for underrepresented students in computer science, and outreach to primary-school students who may not have been exposed to artificial intelligence and data science as an area of study. The focus of this project is on data-driven natural language generation with deep learning. In recent years, the field has made progress on using new machine learning methods to learn to generate text based on human examples; however, this work is still far from human level particularly when generating document-level text. This project will develop new machine learning methods using deep generative models. These include learning and controlling document structure with deep hidden Markov models, content determination through generative reference and alignment, and aggregating content into text through bottom-up selection. Each of these combines research into machine learning methods with natural language analysis. The goal of the project is to develop an open-source implementation of a text generation methods that allow users to easily target new domains and assess the fidelity and coherence of their system.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.