Search Result Re-ranking via Ensemble Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing search engines often fail to accurately rank search results, leading users to miss relevant content, and fine-tuning these engines to improve accuracy can be time-consuming and may not be possible with proprietary or licensed systems.
Innovation Solution
A system and method that enhance search engine performance by generating and using ensemble scores for search results, based on embeddings of queries and modified chunks, without modifying the existing search engine configuration, and allowing the use of pre-existing fine-tuned machine learning models to incorporate domain-specific knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If search engines return more search results to improve coverage, then the quantity of information increases, but the accuracy of ranking decreases
Solution Approach 1:
The patent segments the search result processing into multiple independent scoring components (ensemble scores from multiple models, domain-specific scores, freshness scores, etc.) that can be calculated separately and then combined. This allows comprehensive evaluation of many results without compromising the precision of individual rankings, as each segment contributes a specific aspect of relevance assessment.
Solution Approach 2:
The patent introduces multiple scoring parameters beyond simple relevance matching, including domain-specific scores, freshness scores, and ensemble scores from multiple machine learning models. By changing and expanding the scoring parameters, the system can accurately rank a larger number of results by considering multiple dimensions of relevance simultaneously.
2Measurement precision
If search engines are fine-tuned to improve ranking accuracy, then the precision of search results improves, but the time and effort required increases significantly
Solution Approach 1:
The patent performs preliminary actions by pre-training multiple domain-specific machine learning models and pre-processing search results with various enhancement techniques (chunking, embedding generation, etc.) before the actual search query is executed. This preliminary preparation enables fast, accurate ranking during query execution without requiring time-consuming fine-tuning at query time.
Solution Approach 2:
The patent creates multiple copies of search results with different transformations (chunks, embeddings, domain-specific annotations) and evaluates them using multiple machine learning model copies. This copying approach allows comprehensive evaluation without modifying the original search engine, achieving high accuracy through parallel evaluation of multiple versions.
3Device complexity
If proprietary or licensed search engines are used, then the device complexity is reduced, but the ability to fine-tune for domain-specific knowledge is lost
Solution Approach 1:
The patent introduces an intermediary layer between the proprietary search engine and the user interface. This intermediary processes search results through multiple domain-specific machine learning models and scoring mechanisms without modifying the original search engine configuration. The intermediary acts as a mediator that adds domain-specific adaptability while preserving the simplicity of using existing search engines.
Solution Approach 2:
The patent creates a universal framework that can work with multiple different proprietary search engines and multiple domain-specific models simultaneously. The ensemble scoring system and domain-specific enhancement techniques are designed to be engine-agnostic, allowing the same approach to be applied across different search engines while incorporating various domain knowledges.
Data Source
AI summary
Systems and methods are described for enhancing performance of a search engine by using pre-existing fine-tuned machine learning models that include domain-specific knowledge. A computing device receives search results from a search engine based on a query, in which the search results include chunk identifiers and a first weight score associated with each chunk identifier. The computing device further determines a subset of search results, which includes a first predetermined number of top-ranked chunk identifiers in the search results. The computing device generates, for each chunk identifier in the subset of search results, one or more second weight scores. Then the computing device generates, via a machine learning model, ensemble scores for the chunk identifiers. Afterwards, the computing device determines a second predetermined number of top-ranked chunk identifiers based on the ensemble scores.