Expandable Dialogue System Training via Weak Supervision

The patent applies dynamics by transitioning from static constrained grammars to dynamic neural network-based parsing. The system uses trained neural networks that can adaptively parse utterances into structured representations, allowing the system to handle diverse meanings without requiring exhaustive predefined grammatical constraints. This dynamic approach enables the system to learn and generalize from training data, expressing meanings that would be impossible to capture with fixed grammatical rules.

The patent applies copying by enabling non-expert administrators to create new dialogue patterns through weak supervision without requiring deep understanding of system internals. Administrators can annotate training data and supervise learning processes, effectively copying and adapting system capabilities to new domains. The system learns from these annotated examples, allowing expansion to new functionalities while maintaining stability through the underlying neural network architecture and training frameworks.

The patent applies self-service by enabling the system to expand its own functionality through weakly supervised learning. Non-expert administrators can annotate training data and supervise the learning process without requiring significant engineering intervention. The neural network automatically learns from these annotations, allowing the system to self-expand to new domains and functionalities. This dramatically reduces the need for engineering resources compared to traditional approaches requiring manual pattern definition and system reconfiguration.