Dialogue State Tracking Using Dual Dynamic Graph Neural Network

The patent segments the slot feature learning process into multiple independent feature dimensions. Instead of learning a single aggregated feature, the system learns multiple feature vectors (e.g., semantic feature, positional feature, contextual feature) separately for each slot, then combines them through attention mechanisms. This segmentation allows the system to capture different aspects of slot features independently, improving measurement precision without excessive complexity increase.

The patent transitions from single-dimensional slot feature representation to multi-dimensional feature space. By introducing additional feature dimensions (semantic, positional, contextual, hierarchical), the system can distinguish between slots with the same indicating words through their different feature vectors. This dimensional expansion enables precise differentiation while maintaining manageable complexity through efficient feature combination strategies.

The patent segments the multidomain dialogue tracking into domain-specific slot tracking modules. Each domain has its own slot feature learning and state tracking mechanism, allowing independent optimization for each domain. This segmentation enables precise tracking within each domain while managing overall system complexity through modular architecture.

The patent implements a universal feature extraction and combination framework that works across multiple domains. The attention mechanism and feature combination strategies are domain-agnostic, allowing the same core architecture to handle different domains with their specific slot features. This universality reduces redundant complexity while maintaining high measurement precision across all domains.

The patent performs preliminary feature extraction and encoding before the main distinction process. By pre-processing slot features into comprehensive feature vectors that include semantic, positional, and contextual information, the system prepares all necessary information in advance. This preliminary action ensures that when values with the same indicating words need to be distinguished, all relevant features are already available, minimizing information loss during the distinction process.

The patent introduces attention mechanisms as intermediary components that selectively weigh and combine multiple slot features. These attention mechanisms act as mediators that preserve important information from different feature dimensions while filtering out redundant or conflicting information. The attention weights dynamically adjust to retain the most informative features, reducing overall information loss while maintaining reliability in distinguishing values with the same indicating words.