Composite Embedding Similarity Scoring With Multi-Level Granularity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems using individual vector embeddings for similarity comparisons face increased processing time and complexity due to the need for multiple similarity comparisons across different dimensions.
Innovation Solution
A system that generates composite embeddings by concatenating multiple initial embeddings, allowing for efficient similarity scoring through cosine similarity calculations, thereby reducing processing time and complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple individual vector embeddings are used for similarity comparisons, then measurement precision is improved, but processing time increases
Solution Approach 1:
The patent combines multiple individual embeddings (e.g., word embedding, character embedding, phrase embedding) into a single composite embedding by concatenation. This merging allows the system to perform similarity comparison across all dimensions simultaneously using one cosine similarity calculation, rather than performing multiple separate comparisons and aggregating results, thus reducing processing time while maintaining comprehensive measurement precision
2Measurement precision
If multiple individual vector embeddings are used for similarity comparisons, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent merges multiple embedding generation components and multiple similarity comparison operations into a unified composite embedding framework. By concatenating embeddings into a single vector and performing one similarity calculation, the system reduces the number of computational components and simplifies the overall architecture, thereby reducing device complexity while preserving the ability to measure similarity across multiple dimensions
3Loss of information
If multiple individual vector embeddings are used for similarity comparisons, then comprehensiveness of information is improved, but computational overhead increases
Solution Approach 1:
The patent combines multiple information-rich embeddings into a single composite embedding that preserves all dimensional information through concatenation. This approach maintains complete information from all embedding sources while reducing computational overhead by performing a single similarity comparison operation instead of multiple separate operations, thus lowering the computational burden
Data Source
AI summary
Systems and methods for selecting items from a pool of items based on comparisons of composite embeddings are disclosed. An anchor embedding is generated for a target string. The anchor embedding is a composite embedding including at least a first initial embedding and a second initial embedding. An item embedding is obtained for each item in a pool of items. Each item embedding is a composite embedding including a first initial item embedding and a second initial item embedding. A similarity score is generated by comparing the item embedding to the anchor embedding for each item in the pool of items and a set of items is selected from the pool of items. The set of items includes a predetermined number of items in the pool of items having a highest similarity score.