Exploring Large-Scale Digital Archives – Opportunities and Limits to Use Unsupervised Machine Learning for the Extraction of Semantics

Abstract

The current excitement in regards to machine learning has spurred enthu-siasm amongst collection holders and historians alike to rely on algorithms to re-duce the amount of manual labor required for management and appraisal of largevolumes of non-structured archival content. The Digital Humanities and commer-cial archival software promote out-of-the-box tools for auto-classification, but is theadoption of machine learning as straightforward as it is currently presented in boththe popular press and the Digital Humanities literature? This chapter brings a senseof pragmatism to the debate by giving an overview of both possibilities and limitsof machine learning to extract semantics from large collections of digitized textualarchives. Two methods have gained substantial popularity: Topic Modeling (TM)and Word Embeddings (WE). This chapter introduces these non-supervised ma-chine learning methods to the community of historians, based on an experimentalcase-study of digitized archival holdings of the European Commission (EC).