Résumé:
Knowledge Graphs are employed to refer to entities and their relationships in a semantic, struc- tured form that enhances information organization and retrieval. However, due to their nature and size, traditional Information Retrieval techniques typically do not work for use against knowledge graphs.
This thesis proposes a cutting-edge Information Retrieval approach based on Knowledge Graph Embeddings and Relational Graph Convolutional Networks. Dense vector representations of entities are initially acquired using KGE models to encode semantic relations. These embeddings are afterwards fine-tuned using R-GCN to incorporate structural and relation knowledge from the graph. The learned embeddings are used for retrieving relevant entities or concepts by semantic similarity.
The system is evaluated across several scenarios, namely: the evaluation of embedding and similarity search techniques using standard metrics, the evaluation of queries using a semantic similarity module based on Bio_ClinicalBERT and the exploitation of the Knowledge Graph in a question-answering use case. Results indicate that adding KGE to R-GCN improves retrieval quality in a way it is possible to have better and more contextual search results.
Keywords: Information Retrieval, Knowledge Graph, Knowledge Graph Embeddings, Rela-
tional Graph Convolutional Network, Semantic Search, Medical Knowledge Graphs.
