Research Process Knowledge Graph Extraction from Publications

We present a digital workflow for creating knowledge graphs (KG) describing research processes by extracting relevant entities and relations from the text of publications, associating them with publication metadata and exporting the output as Resource Description Framework (RDF) triples adhering to Linked Data standards. We extract three types of entities: (i) Activities – i.e. actual research processes or steps thereof, like an archeological excavation, an anthropological study, an experiment, etc. carried out by the researchers-authors of the paper which appear in text as entities of variable length with complex lexico-syntactic structure; (ii) Methods – i.e. procedures employed by researchers to carry out research activities, like an algorithm or a specific technique, which appear in text as named entities of variable length and (iii) Goals – i.e. the research tasks that were addressed by the researchers through their activities which appear in text as entities of variable length with mostly fixed lexico-syntactic structure. In addition, we extract two types of relations: (i) employs(Activity, Method) among activities and the methods they employ and (ii) hasObjective(Activity, Goal) linking research activities with their goals. We experiment with Convolutional Neural Networks (CNN), transformer-based binary classifiers; a two-stage pipeline implementation comprising a transformer-based text classifier, which predicts whether a sentence contains the entities sought, in tandem with a transformer-based entity recognizer for finding the boundaries of the entities inside the sentences that contain them; and Large Language Model (LLM) prompting techniques. For training/evaluation we use manually created datasets annotated specifically for the tasks of entity and relation extraction. We evaluate our methods along four dimensions discussing the differences in results based on: (i) classification method, (ii) linguistic characteristics of each entity type, (iii) processing granularity (token-based, entity-based partial and entity-based strict) and (iv) KG creation where we explore the capabilities of the output through indicative SPARQL queries. The entire workflow is ontology-driven, based on Scholarly Ontology, specifically designed for documenting scholarly work.