Information Retrieval Service Aspects of the Open Research Knowledge Graph

Abstract

Information Retrieval (IR) takes a fresh perspective in the context of the next-generation digital libraries such as the Open Research Knowledge Graph (ORKG). As scholarly digital libraries evolve from document-based to knowledge-graph-based representations of content, there is a need for their information technology services to suitably adapt as well. The ORKG enables a structured representation of scholarly contributions data as RDF triples - in turn, it fosters FAIR (Findable, Accessible, Interoperable, and Reusable) scholarly contributions. This thesis has practically examined three different IR service aspects in the ORKG with the aim to help users: (i) easily find and compare relevant scholarly contributions; and (ii) structure new contributions in a manner consistent to the existing ORKG knowledge base of structured contributions. In the first part, it will evaluate and enhance the performance of the default ORKG “Contributions Similarity Service.” An optimal representation of contributions as documents obtains better retrieval performance of the BM25 algorithm in Elasticsearch. To achieve this, evaluation datasets were created and the contributions search index reinitialized with the new documents. In its second part, this thesis will introduce a “Templates Recommendation Service.” Two approaches were tested. A supervised approach with a Natural Language Inference (NLI) objective that tries to infer a contribution template for a given paper if one exists or none. And an unsupervised approach based on search that tries to return the most relevant template for a queried paper. Our experiments favoring ease of practical installation resulted in the conclusion that the unsupervised approach was better suited to the task. In a third and final part, a “Grouped Predicates Recommendation Service” will be introduced. Inspired from prior work, the service implements K-Means clustering with an IR spin. Similar structured papers are grouped, their in-cluster predicate groups computed, and new papers are semantified based on the predicate groups of the most similar cluster. The resulting micro-averaged F-measure of 65.5% using TF-IDF vectors has shown a sufficient homogeneity in the clusters.