RAPID: Augmented Intelligence for Accelerating Covid-Related Scientific Discovery

该项目将开发新的人工智能(AI)方法，以提高生物医学研究人员的生产力，并在新冠肺炎大流行的背景下加快科学发现。我们将每周更新我们广泛使用的Cord-19和SciSight资源，这是研究SARS-CoV-2的研究人员的关键资源，其他研究人员已经下载了超过10万次。我们还将从两个方面扩展这些资源，使其对医生和研究人员更有用。首先，我们将自动生成每篇论文的一句话摘要，以加快对快速变化的文学的轰动。其次，我们将自动提取广泛的实体(如疾病症状和研究挑战)和关系，以改进过滤和搜索。为了生成研究论文的一句话摘要，我们将使用两种新技术训练一个抽象的BART模型：1)在标题预测的辅助任务上进行联合训练；2)使用一组单句摘要进行微调，这些摘要是通过众包编辑同行评审来自OpenReview等网站的评论而生成的。我们将使用自动(Rouge)指标和用户偏好的组合来测试我们的一句话摘要生成。为了增加从研究论文中提取的实体和关系的数量，我们将使用数据编程技术进行引导，然后应用图-神经网络方法。我们将结合专家注释的数据和来自相关知识库的信息来评估我们的进展。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The project will develop new artificial intelligence (AI) methods to augment the productivity of biomedical researchers and accelerate scientific discovery in the context of the COVID-19 pandemic. We will issue weekly updates to our widely-used Cord-19 and SciSight resources, which are a critical resource for researchers studying SARS-CoV-2, having already been downloaded over 100,000 times by other researchers. We will also extend these resources to make them more useful to doctors and researchers in two ways. First, we will automatically generate one-sentence summaries of each paper to speed sensemaking of the rapidly changing literature. Second, we will automatically extract a broad range of entities (such as disease symptoms and research challenges) and relations to improve filtering and search. In order to generate one-sentence summaries of research papers, we will train an abstractive BART model, using two novel techniques: 1) co-training on the auxiliary task of title prediction, and 2) fine-tuning using a set of one-sentence summaries that we will generate by crowd-sourcing edits peer-review comments taken from sites such as OpenReview. We will test our one-sentence summary generation with a combination of automated (Rouge) metrics and user preference. In order to increase the number of entities and relations extracted from research papers, we will bootstrap with data-programming techniques then apply graph-neural-network methods. We will evaluate our progress using a combination of expert-annotated data and held out information from relevant knowledge bases.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.