Knowledge-centered conversational agents with a drive to learn

We propose a framework for adaptive knowledge-centered conversational agents (KCCAs) that acquire knowledge from conversations, evaluate its quality, and are internally driven to become more knowledgeable. Unlike service agents trained for predefined tasks and based on unidirectional interactions, KCCAs treat people as diverse, social sources of information across open domains. Each KCCA is equipped with an episodic Knowledge Graph (eKG): a structured, evolving memory that organizes knowledge communicated in dialogue into RDF triples and nested named graphs. This information is accumulated in a database that stores claims, related perspectives, and original sources along with contextual and temporal information about the dialogues. This representation enables KCCAs to manage subjective knowledge, detect conflicting viewpoints, and reason about people’s beliefs over time. The eKG also supports domain-independent assessment of knowledge and interaction quality. Structural patterns in the graph can be associated with knowledge dimensions such as correctness, completeness, redundancy, interconnectedness, and diversity. Applying these patterns to a KCCA's memory results in targeted areas for quality improvement that can be regarded as the agent's knowledge needs. Moreover, tracking the evolution of the eKG provides insight into knowledge exchange during dialogue and correlates with human judgments of interaction quality. KCCAs are also equipped with task-agnostic dialogue capabilities for knowledge acquisition, such as dialogue management, natural language generation, and natural language understanding. To actively pursue its knowledge needs, KCCAs employ reinforcement learning to adapt their communicative strategies. To express their current knowledge, they aggregate information in a perspective-aware manner, allowing for nuance where needed. To ingest new knowledge, they transform natural language into RDF triples to be added to the eKG. We demonstrate the KCCA framework in three real-world applications, including: 1) Type 2 diabetes lifestyle support, 2) Personal timeline reconstruction, and 3) Counter-speech generation. These use cases highlight the strengths of graph technologies for dialogue agents, particularly due to their structure and connectivity properties. We further extend the KCCA framework to include embodied agents, developing tools for capturing and creating multimodal interactions. Together, these contributions provide a dynamic, adaptive, and socially aware foundation for Hybrid Intelligence systems capable of long-term, knowledge-driven interaction.