Cross-Token Attention Summarization for Storage Efficiency

The patent extracts and processes only the most salient features and representations from documents rather than processing entire documents. The cross-token attention model and cross-utterance attention model selectively attend to important tokens and utterances, extracting key information while discarding redundant data, thereby reducing storage requirements while maintaining predictive accuracy.

The patent segments documents into smaller units (tokens, utterances) and processes them hierarchically through multiple attention mechanisms. The cross-token attention operates at the token level, cross-utterance attention at the utterance level, and document-level attention at the document level, allowing selective processing that reduces overall storage needs while preserving important information.

The patent performs preliminary processing through cross-token attention and cross-utterance attention mechanisms that pre-compute and store condensed representations of documents. These pre-processed representations are then used in subsequent training stages, eliminating the need to repeatedly process entire documents and significantly accelerating training speed while maintaining predictive accuracy.

The hierarchical segmentation into token-level, utterance-level, and document-level processing allows the system to train on progressively condensed representations. This multi-stage approach with segmented processing reduces the computational burden in later training stages while preserving the ability to maintain high predictive accuracy through the detailed token and utterance representations.

The patent applies local quality by using attention mechanisms that selectively focus on important tokens and utterances while maintaining detailed representations where needed. The cross-token attention assigns different weights to different tokens based on their importance, preserving detailed information for salient tokens while using more condensed representations for less important tokens, thus maintaining information completeness while improving storage efficiency.

The patent changes the parameter of representation density dynamically through the attention mechanisms. Different parts of the document are represented at different levels of detail based on their importance, with the cross-token and cross-utterance attention models adjusting the effective representation density to optimize both storage efficiency and information completeness.