Contextualising Conversational AI

Conversational AI has evolved from simple rule-based systems to sophisticated large language models capable of engaging in complex dialogues. However, despite significant advances in fluency and coherence, these systems often struggle with fundamental aspects of human communication, particularly in how they handle context as a pragmatic factor. This thesis investigates how conversational AI systems can better understand, represent, and respond to the multiple layers of context that shape human dialogue. Through an examination of uncertainty management, perspective integration, and context representation, we explore both the potential and limitations of current approaches to context-aware conversational AI. We begin with a theoretical foundation grounded in Grice's Cooperative Principle, establishing how pragmatic maxims influence current AI implementations. The first research direction explores how conversational AI models handle uncertainty, examining both the expression and consistency of uncertainty across different contexts. Through multilingual experiments and analysis of explanation triggers, we find significant misalignments between models' internal confidence levels and their expressed certainty, particularly in non-English languages where models maintain high confidence despite degraded performance. The second direction investigates how systems incorporate individual perspectives, examining the integration of subjective information, user-specific adaptation mechanisms, and cultural considerations. Research into task-oriented dialogue systems reveals challenges in balancing factual and subjective knowledge, while analysis of multilingual datasets shows that current approaches often preserve Anglo-centric biases rather than representing diverse cultural viewpoints. The final direction examines how conversational AI can be grounded in world and personal contexts through structured representation mechanisms. We explore Graph-based approaches for maintaining dialogue memory and context across interactions, demonstrating promise for both transparency and long-term coherence in conversational systems. The findings show that while conversational AI has made substantial progress, current approaches often fail to capture the nuanced ways humans use context to shape communication, from expressing appropriate uncertainty to adapting responses based on cultural backgrounds. We conclude that context should be treated not as an optional enhancement but as a fundamental component of conversational AI design, requiring research that bridges theoretical understanding of human communication with practical implementation in AI systems.