A Library Perspective on Supervised Text Processing in Digital Libraries: An Investigation in the Biomedical Domain

Digital libraries that maintain extensive textual collections may One way to explore a digital library's content is to apply natural want to further enrich their content for certain downstream applications, language processing methods, e.g., identify central entities (e.g., e.g., building knowledge graphs, semantic enrichment of the Person Albert Einstein), their relationships (e.g., Albert Einstein documents, or implementing novel access paths. All of these applications was born in Ulm), and classify documents as belonging to require some text processing, either to identify relevant classes (e.g., descriptive articles). The extraction of semantic relationships entities, extract semantic relationships between them, or to classify between named entities is already used in several digital documents into some categories. However, implementing reliable, library projects for different purposes, e.g., constructing a biomedical supervised workflows can become quite challenging for a digital knowledge graph from scientific papers like SemMedDB [18], library because suitable training data must be crafted, and reliable harvesting leader boards of how computer science methods perform models must be trained. While many works focus on achieving the on benchmarks [17], harvesting scientific information as done highest accuracy on some benchmarks, we tackle the problem from in SciGraph [44], enabling graph-based discovery systems in digital a digital library practitioner. In other words, we also consider tradeoffs libraries [20], or enriching library content like newspapers as done between accuracy and application costs, dive into training data in the Swiss-Luxembourgish impresso [10].