Knowledge Injection for Field of Research Classification and Scholarly Information Processing

This paper aims to address the need for efficient classification of schol arly articles into their respective fields considering the growing volume of scientific research. We address this by exploring the application of Semantic Web resources, such as DBpedia, to represent classes in deep learning models using knowledge graph embeddings. We construct a dataset comprising publications from 123 fields of research using openly available resources. Models are then trained using differ ent ways to semantically represent class labels via an automatic entity linking approach to DBpedia. We assess the impact of different publication meta data combinations, including titles, abstracts, authors and publishers, on the performance of these models. We find that general pre-trained knowledge graph embeddings suffer from a noise problem when applied to research classification tasks, with textual descriptions of DBpedia entities emerging as a more effective means of represent ing classes. Notably, we notice that titles and abstracts alone, without additional metadata features such as authors and publishers, provide the best-performing representation for publication metadata. This study fills a gap in the literature by demonstrating the effi cacy of deep learning methods and Semantic Web resources like DBpedia in the classification of scholarly articles across various fields of research. Our findings underscore the importance of considering the impact of different metadata fea tures on model performance, as well as using textual descriptions of DBpedia entities as class representations.