Query Matching Model Generalization via Cross Attention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional search engine techniques fail to satisfy users' demands for quick and precise information acquisition in the era of information explosion, as they lack the ability to generalize queries effectively, leading to suboptimal accuracy and recall rates in search and question-answering systems.
Innovation Solution
A method and apparatus for generalization processing using a pre-trained query matching model based on a cross attention model, which determines candidate queries through literal, semantic, and query rewriting matching to generate a generalized query, improving query matching accuracy and comprehensiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional search engine techniques are used, then the system structure remains simple, but the accuracy and recall rate of information acquisition deteriorate
Solution Approach 1:
The patent segments the query processing into multiple independent modules: query rewriting module, candidate query determination module (with literal matching, semantic matching, and query rewriting sub-modules), and query matching model. Each module handles a specific aspect of query generalization, improving accuracy while maintaining manageable system complexity through functional decomposition
Solution Approach 2:
The patent applies preliminary action by pre-training the query matching model offline using large-scale query data before deployment. The model learns query representation and matching patterns in advance, enabling fast and accurate query generalization during online operation without requiring complex real-time processing
2Quantity of substance
If traditional search techniques are used, then the system remains simple to operate, but the recall rate of search results deteriorates
Solution Approach 1:
The patent merges multiple query matching approaches (literal matching, semantic matching, and query rewriting) into a unified processing framework. These different matching strategies work together to generate comprehensive candidate query sets, improving recall rate by capturing diverse query variations while presenting a unified interface to users
Solution Approach 2:
The query matching model serves multiple functions: it performs literal matching, semantic matching, query rewriting, and candidate query selection within a single unified model. This multi-functionality increases the comprehensiveness of search results without requiring separate specialized systems for each function
3Measurement precision
If query generalization is not applied, then the system remains fast to execute, but the accuracy of query responses deteriorates
Solution Approach 1:
The query matching model is pre-trained offline on large-scale query data to learn optimal matching patterns and query representations. This preliminary training enables the model to perform accurate query generalization rapidly during online execution, achieving high response accuracy without sacrificing processing speed
Solution Approach 2:
The patent replaces traditional rule-based query expansion mechanisms with a learned neural network model. The model automatically learns query transformation patterns from data, eliminating the need for manual rule engineering and enabling fast, accurate query generalization through differentiable optimization
Data Source
AI summary
The present disclosure provides a generalization processing method, apparatus, device and computer storage medium, and relates to technical field of artificial intelligence and specifically to a deep learning technique. A specific implementation solution is: determining a set of candidate queries in a query library that are similar to a requested query in at least one of a literal matching manner, a semantic matching manner and a query rewriting manner; determining a generalized query corresponding to the requested query from the set of candidate queries by using a pre-trained query matching model; wherein the query matching model is obtained by pre-training based on a cross attention model. The generalization for the requested query can be achieved according to the present disclosure.