BIGDATA: IA: Hype Cycles of Scientific Innovation

科学发现是一个基于合作、评估和共识的集体过程。但是，科学研究的巨大扩张使得我们很难判断哪些智力努力促成了集体进步。更好的识别和跟踪科学运动的模型，从大量的文章、书籍、赠款和专利中识别和跟踪危及科学和学术工作的科学运动，对于提高我们国家在科学和工业方面取得进步的能力至关重要，有助于政策制定者、资助机构、风险资本家以及科学家和学者自己。我们的项目通过收集和分析大量科学和其他学术文章、书籍、赠款和专利中的文本来研究创新。通过观察语言的微妙模式以及它们如何随时间变化，我们可以描述和预测知识创新的集体趋势在何时何地出现和下降；哪些科学思想和运动产生了对工业和卫生保健至关重要的转化知识；以及这种集体智力运动的关键驱动因素是什么。这项工作有助于确定哪些领域的创新潜力最大，哪些领域最容易接受新发现的到来，以及资源影响最大的时间和地点。该项目基于研究人员从1990年至2016年编制的美国研究活动文本的大型数据集，包括科学文章、拨款和专利，并通过消除这些数据集中提到的个人的歧义和链接。该项目使用主题建模和其他自然语言处理算法，通过绘制表征这些语料库的子语言来识别这些语料库中的不同智力运动，并应用方法来验证和评估这些运动。这使得研究人员可以使用潜在增长混合模型和k谱聚类等工具来模拟运动轨迹，确定与每次运动相关的机制，它们如何随时间上升和下降，以及它们如何转化为工业或健康应用。例如，该项目研究了智力运动的轨迹如何成为其环境和竞争性研究努力的函数，并展示了它如何依赖于关键资源（例如，资金、认可、新招募和学员、支持的社会网络或知识本身的连贯性）的时间和规模。在这种情况下，这项工作是重要的第一步，揭示了创新的配方是一个集体的、偶然的过程。

Scientific discovery is a collective process based on collaboration, assessment and consensus. But the enormous expansion of scientific research makes it difficult to tell which intellectual efforts forge collective advances. Better models for identifying and tracking scientific movements from the vast collections of articles, books, grants, and patents that compromise scientific and academic work is crucial to improving our nation's ability to make advances in science and industry, helping policy makers, funding agencies, and venture capitalists as well as scientists and scholars themselves. Our project studies innovation by collecting and analyzing texts in vast collections of scientific and other scholarly articles, books, grants and patents. By looking at the subtle patterns of language and how they change over time, we can describe and predict where and when collective trends of knowledge innovation emerge and decline; which scientific ideas and movements result in translational knowledge key to industry and health care; and what the key drivers are for such collective intellectual movements. This work is helping identify where the potential is greatest for innovation, which fields are most primed and receptive to the arrival of new discoveries, and the times and places where resources have the greatest influence.The project is based on a large dataset of texts the researchers have compiled on US research activity from 1990-2016, including scientific articles, grants, and patents, and by disambiguating and linking mentions of individual people across these datasets. This project uses topic modeling and other natural language processing algorithms to identify distinct intellectual movements in these corpora by drawing on the sub-languages that characterize them, also applying methods to validate and evaluate these movements. This allows the researchers to model the trajectories of movements over time with tools like latent growth mixture modeling and k-spectral clustering, identifying the mechanisms associated with each movement, how they rise and fall over time, and how they translate into industrial or health applications. For example, the project studies how the trajectory of intellectual movements is a function of their environment and competing research efforts, and show the ways in which it depends on the timing and magnitude of key resources (e.g., money, recognition, new recruits and trainees, social networks of support, or the coherence of the knowledge itself). In such a fashion, this work is an important first step in unraveling the recipe for innovation as a collective, episodic process.