The objective: to build a robust knowledge graph that supports a large-scale customer facing question-answering system.
Context: Empirical experience from the presenter’s work supporting question-answering functionality in virtual assistants (notably Amazon Alexa) and, more recent, work at Compare the Market.
Covering: The talk will cover common pitfalls and proven solutions, including insights into what works - and what doesn’t - when deploying and scaling knowledge graphs in real-world environments.
- Use Case Design: What to consider when tailoring the graph to different domains, user profiles, machine-learning applications, and answer formats ranging from plain text and tables to graphical responses.
- Operationalising Development: Establishing high-level design principles, managing iterative development, addressing distributed ownership, handling change management and testing, and ensuring the system adapts to evolving data sources.
- Technical Challenges: We will explore trade-offs in accuracy, coverage, latency, expressivity and abstraction, and consider other potential requirements such as context-handling, semi-structured data and permissions levels.
- Internationalisation: Strategies for scaling globally including language support, units of measurement, and localised content.
Attendees will come away with a clear framework for designing knowledge graphs that are both scalable and adaptable, grounded in practical industry experience.