A considerable amount of biological data is preserved as physical documents, the legacy of former explorers, collectors, researchers, and others. Mobilizing data from handwritten documents has been considered particularly challenging, with well-known cases such as the manual transcription of specimen labels and herbarium sheets by museum staff, or crowdsourced transcription of data card collections through online platforms.
Here we present a pipeline of open-source software that can be used to
automatically transcribe handwritten text,
make it publicly available,
annotate it with e.g., scientific names,
extract names in Darwin Core Archive (DwC-A) for third-party reuse, and
automatically recognize named entities in the machine-readable text.
automatically transcribe handwritten text,
make it publicly available,
annotate it with e.g., scientific names,
extract names in Darwin Core Archive (DwC-A) for third-party reuse, and
automatically recognize named entities in the machine-readable text.
We based our use case on the correspondence of the German zoologist Karl Wilhelm Verhoeff, related to the Myriapoda collection held at the Musée Zoologique de Strasbourg.
The documents were processed with Transkribus (Muehlberger et al. 2019), a mostly open-source virtual research environment (OS VRE), which allows text in images to be converted into machine-readable text amenable to semantic enrichment. We achieved a character error rate as low as 5%, a remarkable result for handwritten material, as an accuracy higher than 95% for printed material is acceptable (Deutsche Forschungsgemeinschaft 2016). We then used Myriatrix (Martínez-Muñoz 2019), an instance of the Scratchpads OS VRE (Smith et al. 2011), to create bibliographic references, publish the full text, and annotate the correspondence with scientific names of myriapods. During the process, we added new scientific name spellings and combinations to the taxonomic backbone of Myriatrix and exported the full taxon classification in DwC-A via the Global Biodiversity Information Facility (GBIF) for reuse by the Global Names Architecture and its open-source tools (Patterson et al. 2016, Mozzherin et al. 2017).
As a next step we are planning to subject the corrected text from Transkribus to a specific text-preprocessing workflow combining natural language processing (NLP) and machine learning (ML) techniques (Lücking et al. 2021). This includes, inter alia, a multiple annotation approach for general and bioscientific term classification in order to detect the respective entities automatically. The workflow has been developed in the framework of the Specialized Information Service Biodiversity Research (Koch et al. 2017) to make biodiversity information available via a customized and (bio-)ontology-based semantic search engine (Pachzelt et al. 2021).
We recommend our comprehensive approach to natural history institutions seeking to efficiently digitize and mobilize the rich biological data present in their archival documents.
