RI: Small: Print and Probability - A Statistical Approach to Analysis of Clandestine Publication

这个打印和概率项目开发了新的机器学习和计算机视觉技术，以推断出数千名书籍和小册子印刷商的身份，这些印刷商的身份已经躲避了大约500年。在现代之前，图书贸易往往是危险而隐秘的。由于害怕受到迫害和惩罚，1473年至1800年间，印刷商拒绝在所有已知书籍和小册子中署名的约占四分之一。然而，现在超过13万本书已经被早期英语在线图书（EEBO）项目数字化，这个时代印刷工具的缺陷和变化可能是识别这些打印机的关键。一旦单个金属铅字损坏，就会产生独特的邮票。由于字版属于特定的印刷商，因此损坏字版的印痕可以作为识别成千上万种秘密出版物的印刷商的指纹。打印和概率项目自动检测和跟踪这些独特的损坏类型，以发现有关书籍历史的新信息。在这个项目中开发的方法可以推广到其他重要的任务和领域，例如，数字取证和作者归属。此外，打印和概率项目将以多学科的方式培养学生，让他们参与多个领域的合作。通过开发视觉异常检测的新技术，打印和概率项目检测产生一致畸变的损坏字体。基于这些损坏的类型提取，该项目开发了打印机和损坏的字母形式识别的概率模型，允许在规模上直接推断打印机。该框架还将其他证据来源纳入识别模型——最重要的是，个别出版社排字工人的拼写、标点和间距习惯，他们的独特做法使他们能够在集合中的所有文本页面上进行聚类和自动归因。该项目集成了一种自动排字归属的新方法，开发了一种用于打印机识别的统计模型，该模型利用了由稀有书籍学者手工编制的相同证据来源，但其规模和速度前所未有。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。

This Print and Probability project develops novel machine learning and computer vision techniques to infer thousands of book and pamphlet printers whose identities have eluded scholars for roughly 500 years. Before the modern era, the book trade was often dangerous and secretive. For fear of persecution and punishment, printers between 1473-1800 declined to attach their names to about a quarter of all known books and pamphlets. However, now that over 130,000 books have been digitized by the Early English Books Online (EEBO) project, defects and variations in the printing tools of this era may hold the key to identifying these printers. Once an individual piece of metal type is damaged, it creates unique stamps. Since typesets belonged to specific printers, impressions of damaged type can thus serve as the fingerprints to identify the printers of tens of thousands of clandestine publications. The Print and Probability project automatically detects and tracks these unique pieces of damaged type in order to uncover new information about the history of books. The methods developed in this project could be generalizable to other important tasks and domains - for example, digital forensics and authorship attribution. In addition, the Print and Probability project will train students in a multidisciplinary way, engaging them in collaboration across multiple fields.By developing new techniques for visual anomaly detection, the Print and Probability project detects damaged letterforms that create consistent aberrations. Based on these damaged type extractions, the project develops probabilistic models of both printer and damaged letter form identifications that allow direct inference of printers at scale. This framework also incorporates other sources of evidence into the identification model - most significantly, the spelling, punctuation, and spacing habits of individual press-house compositors, whose distinctive practices lend themselves to clustering and automatic attributions across all pages of text in the collection. Integrating a new method for automatic compositor attribution, this project develops a statistical model for printer identification that leverages the same sources of evidence compiled manually by scholars of rare books, but at a scale and speed never before possible.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.