Growing and Serving Large Open-domain Knowledge Graphs: Connecting Legibility and Knowledge Representation

Growing and Serving Large Open-domain Knowledge Graphs: Connecting Legibility and Knowledge Representation

Introduction:
In the realm of knowledge representation and machine learning applications, large open-domain knowledge graphs (KGs) play a crucial role. These KGs not only provide a repository of information but also serve as a foundation for various downstream tasks such as fact verification, fact ranking, related entities, and entity linking. Simultaneously, the concept of legibility in physical environments, specifically the distinction between warrens and plazas, offers a unique perspective on the quality and effectiveness of knowledge representation. By understanding the commonalities between these two concepts, we can enhance our understanding of KGs and their role in facilitating rich user experiences.

The Power of Knowledge Graph Embeddings:
Knowledge graph embeddings are instrumental in leveraging KGs for machine learning applications. By incorporating semantic annotations from the web, KGs are expanded with edges to open-domain web content, adding depth and relevance to the existing knowledge. This expansion enables better search and ranking capabilities, allowing virtual assistants to provide accurate and comprehensive answers to user queries. The continuous updating of KGs with new data from diverse sources ensures the correctness and completeness of facts at scale, making fact verification a seamless process.

Fact Ranking and Related Entities:
In a KG, entities can be associated with multiple facts, which can pose challenges when generating high-quality answers. For instance, when a virtual assistant is queried about a person's occupation, it needs to infer an importance-based ranking over the facts associated with that person in the graph. This ranking ensures that the generated answer is of high quality and relevance. Moreover, KGs empower virtual assistants to proactively provide users with information about related entities, enhancing the user experience by facilitating discovery and expanding knowledge.

The Role of Legibility in Knowledge Representation:
The concept of legibility, as illustrated by the warren/plaza analogy, sheds light on the quality of knowledge representation in KGs. Warrens, characterized by personalized and customized environments with emergent collaboration, represent a social environment where participants have limited visibility beyond their immediate context. In the context of KGs, warren-like representation implies an atmosphere of mystery, where narratives and character arcs can flourish. On the other hand, plazas, created by central planners, prioritize a global/big picture view of the knowledge graph. This overly-legible representation often overwhelms narratives and character arcs, leading to a loss of depth and intrigue.