PhD Dissertation Proposal by LO Pei Chi | Connecting The Dots with Knowledge Graph Enrichment

In recent years, knowledge enriched information retrieval applications are widely popular. These include semantic search using knowledge graphs, query answering with knowledge graphs, and knowledge graph-enhanced recommendation. While each of these applications involves its own task and set of assumptions, they are all based on the same high level objective: improving integration of unstructured content with a knowledge graph to enrich the retrieval tasks such that the task results can be more complete with the knowledge graph closing the information gap in the given content, or be more structured with the knowledge graph providing the connections between entities in the results. In this presentation, we choose to focus on content understanding tasks enriched with knowledge graphs. The tasks include entity linking, contextual path retrieval and contextual path generation.

We divide this presentation into two parts. In the first half of the presentation, we focus on our two primary objectives: Entity Linking and Contextual Path Retrieval/Generation. The first research objective addresses the entity linking task of augmenting an input unstructured text with knowledge, i.e., linkage of named entity mentions from the input text to entity entries of some knowledge base or knowledge graph. Our second research objective explores the use of knowledge graph to offer explainable connections between input entities mentioned in the same unstructured text also known as the context. We define the explainable results to be the relevant paths or subgraphs in the knowledge graph that connect the input entities.

In the second half of the presentation, we introduce our two supporting components: Knowledge Graph Embedding and Human Intelligence. Knowledge graph embeddings aims to learn the low-dimensional representations of entities and relations from a known knowledge graph. In our research, we seek to improve the existing knowledge graph embedding methods by incorporating contextual knowledge. Human intelligence, on the other hand, provides human knowledge that does not exist in the existing knowledge graphs. We explore an interactive approach to collect labels from human annotators by modeling the tasks assigned to the annotators, as well as by modeling the expertise of the annotators.

Pei-Chi LO is a PhD candidate advised by Professor Ee-Peng LIM in the School of Information System, Singapore Management University. Her research interests include knowledge graph embeddings, information retrieval, and crowdsourcing.