CAREER: Neural Transcript Summarization and Induction of Document Structure

音频和视频爆炸式的传播范围和影响力，加上准确的字幕，正在扩大获取大量文本的途径。自动转录摘要可以从音频和视频记录的转录本中生成文本摘要。从远程医疗和远程医疗、金融服务、视频会议到播客和直播服务提供商，它为许多拥有大量成绩单的行业带来了希望。无论你是需要分享会议纪要，还是需要快速记录直播录音，使用文字摘要工具都是一种可行的方式，可以将原本需要大量劳动的将录音转换为文本摘要的任务外包出去。尽管从业者渴望总结各种各样的文本，但他们无法处理口语的复杂性。如果没有强大的摘要技术，用户可能会被大量可用的信息所淹没，而无法有效地确定重要的主题。本CAREER项目的研究目标是通过对摘要中的基本问题的研究，为自动摘要提供统一的方法框架，以提高内容选择和生产的效率。这项工作将利用深度神经网络和语言结构预测的力量来诱导抄本上的文档结构，并能够生成全面的摘要。研究计划的具体目标是：(a)揭示非正式的、冗长的自发演讲文本的结构，(b)生成全面的摘要，使用户能够轻松地浏览文本，以及(c)建立一个评估方案，结合内在和外在的措施来评估文本摘要的质量。该项目旨在解决转录摘要中的基本挑战，为现场从业者提供强大的，普遍可访问的摘要解决方案。研究计划将充分整合到协同教育计划中，让不同的学生学习者参与总结技术的探索。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。

The exploding reach and power of audio and video, combined with accurate captioning, is broadening access to large collections of transcripts. Automatic transcript summarization enables the production of textual summaries from transcripts of audio and video recordings. It holds promise for numerous industries that have large collections of transcripts, ranging from telehealth and telemedicine, financial services, video conferencing, to podcast and livestream service providers. Whether one needs to share the minutes of a meeting or quickly take notes of a livestream recording, using a transcript summarization tool is a viable way to outsource the otherwise labor-intensive task of turning voice recordings into textual summaries. Though practitioners are eager to summarize transcripts of various sorts, they cannot deal with the complexities of spoken language. Without robust summarization technology, users can become overwhelmed by the amount of information available and fail to effectively pinpoint topics of importance.The research goal of this CAREER project is to provide a unified methodological framework for automatic transcript summarization through investigation of fundamental problems in summarization to improve the efficiency of content selection and production of transcript summaries. The proposed effort will harness the power of deep neural networks and linguistic structure prediction to induce document structure on transcripts and enable the production of comprehensive summaries. Specific objectives of the research plan are to (a) uncover the structure of informal, verbose transcripts of spontaneous speech, (b) produce comprehensive summaries to allow users to navigate the transcripts with ease, and (c) establish an evaluation protocol that combines intrinsic and extrinsic measures to assess the quality of transcript summaries. The project seeks to address fundamental challenges in transcript summarization to provide robust, universally accessible summarization solutions to field practitioners. The research plan will be fully integrated into a synergistic education plan to engage diverse student learners in exploration of summarization technology.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.