Contextualized Token Retrieval for Fast Accurate Document Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for identifying relevant documents to a query are resource-intensive and often fail to deliver the accuracy they promise, leading to inefficiencies and suboptimal performance.
Innovation Solution
A system that generates relevance scores using only retrieved token vectors of candidate documents, rather than all token vectors, to improve efficiency and accuracy by employing neural networks to process queries and documents, and using similarity measures to identify relevant documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods use all token vectors to evaluate token-level interactions between queries and documents, then measurement precision is improved, but use of energy and device complexity worsen
Solution Approach 1:
The patent segments the document processing into two distinct phases: (1) an initial filtering phase that processes only a small subset of retrieved token vectors to identify candidate documents, and (2) a subsequent evaluation phase that processes all token vectors for the identified candidates. This segmentation allows the system to avoid the computational burden of processing all documents at full detail, thereby reducing overall FLOPS while maintaining accurate identification of relevant documents through the two-stage approach.
Solution Approach 2:
The patent applies partial action by initially processing only a partial set of token vectors (the retrieved subset) to identify candidate documents, rather than processing all token vectors from all documents. This partial processing significantly reduces computational complexity and energy consumption. The system then applies excessive action by processing all token vectors specifically for the identified candidate documents to ensure accurate evaluation and maintain high measurement precision in the final relevance scoring.
2Measurement precision
If conventional methods evaluate token-level interactions between queries and documents, then measurement precision is improved, but productivity worsens
Solution Approach 1:
The patent segments the document retrieval process into an initial fast filtering stage using retrieved token vectors and a subsequent detailed evaluation stage using all token vectors. This segmentation enables the system to quickly narrow down the candidate document set, thereby improving productivity and speed. The detailed evaluation is then applied only to the reduced candidate set, ensuring measurement precision is maintained without the full computational cost applied to all documents.
Solution Approach 2:
The patent uses partial action in the initial stage by evaluating only retrieved token vectors to rapidly identify candidate documents, significantly improving processing speed. Then excessive action is applied in the second stage by evaluating all token vectors for the identified candidates to ensure high measurement precision in the final results, achieving both speed and accuracy.
3Measurement precision
If conventional methods process all token vectors of documents, then measurement precision is improved, but device complexity worsens
Solution Approach 1:
The patent segments the memory processing requirements by initially loading and processing only the retrieved token vectors, which occupy minimal memory space. After identifying candidate documents through this low-memory initial phase, the system then loads and processes all token vectors specifically for the candidate documents in a second phase. This segmentation dramatically reduces peak memory requirements compared to loading all document token vectors simultaneously, while still achieving high measurement precision through the comprehensive evaluation of candidate documents.
4Measurement precision
If conventional methods use resource-intensive processing to identify relevant documents, then measurement precision may be improved, but productivity and reliability worsen due to resource constraints
Solution Approach 1:
The patent segments the processing into two reliable stages: an initial stage using retrieved token vectors that consumes minimal resources and consistently identifies candidate documents, and a second stage using all token vectors for comprehensive evaluation of candidates. This segmentation ensures the system maintains reliable and consistent performance under resource constraints, as neither stage exceeds available computational or memory resources, while still achieving high measurement precision through the thorough evaluation of identified candidates.
Data Source
AI summary
Methods, systems, and apparatuses, including computer programs encoded on computer storage media, for identifying relevant documents to a query by using only retrieved token vectors of candidate documents rather than all token vectors of the documents. That is, by using only retrieved token vectors of candidate documents rather than all token vectors of the documents, the described techniques dramatically increase the speed and accuracy of identifying relevant documents to a query.