Knowledge graphs are structured repositories of knowledge that store facts about the general world or a particular domain in terms of entities and their relationships. Owing to the heterogeneity of use cases that are served by them, there arises a need for the automated construction of domain-specific knowledge graphs from texts. While there have been many research efforts towards open information extraction for automated knowledge graph construction, these techniques do not perform well in domain-specific settings. Furthermore, regardless of whether they are constructed automatically from specific texts or based on real-world facts that are constantly evolving, all knowledge graphs inherently suffer from incompleteness as well as errors in the information they hold.
This thesis investigates the challenges encountered during knowledge graph construction and proposes techniques for their curation (a.k.a. refinement) including the correction of semantic ambiguities and the completion of missing facts. Firstly, we leverage existing approaches for the automatic construction of a knowledge graph in the art domain with open information extraction techniques and analyse their limitations. In particular, we focus on the challenging task of named entity recognition for artwork titles and show empirical evidence of performance improvement with our proposed solution for the generation of annotated training data.
Towards the curation of existing knowledge graphs, we identify the issue of polysemous relations that represent different semantics based on the context. Having concrete semantics for relations is important for downstream applications (e.g. question answering) that are supported by knowledge graphs. Therefore, we define the novel task of finding fine-grained relation semantics in knowledge graphs and propose FineGReS, a data-driven technique that discovers potential sub-relations with fine-grained meaning from existing polysemous relations. We leverage knowledge representation learning methods that generate low-dimensional vectors (or embeddings) for knowledge graphs to capture their semantics and structure. The efficacy and utility of the proposed technique are demonstrated by comparing it with several baselines on the entity classification use case.
Further, we explore the semantic representations in knowledge graph embedding models. In the past decade, these models have shown state-of-the-art results for the task of link prediction in the context of knowledge graph completion. In view of the popularity and widespread application of the embedding techniques not only for link prediction but also for different semantic tasks, this thesis presents a critical analysis of the embeddings by quantitatively measuring their semantic capabilities. We investigate and discuss the reasons for the shortcomings of embeddings in terms of the characteristics of the underlying knowledge graph datasets and the training techniques used by popular models.
Following up on this, we propose ReasonKGE, a novel method for generating semantically enriched knowledge graph embeddings by taking into account the semantics of the facts that are encapsulated by an ontology accompanying the knowledge graph. With a targeted, reasoning-based method for generating negative samples during the training of the models, ReasonKGE is able to not only enhance the link prediction performance, but also reduce the number of semantically inconsistent predictions made by the resultant embeddings, thus improving the quality of knowledge graphs.