Shared Semantic Model for Neural Search Recall and Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In search scenarios, existing technologies face a disjunction in optimization targets between the recall and ranking stages, leading to suboptimal recall of highly matching data and inability to provide early warnings on low matching values, affecting the ranking and searching effect.
Innovation Solution
The implementation of a neural network model that shares a semantic representation model for both recall and ranking stages, ensuring consistent semantic representation and parameter distribution across both stages, thereby improving the ranking effect.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate semantic representation models are used for recall and ranking stages, then each stage can be optimized independently, but the optimization targets become disjoined leading to suboptimal recall and ranking effects
Solution Approach 1:
The patent merges the semantic representation models of the recall and ranking stages into a single shared model. This unified semantic representation model processes both recall queries and ranking queries through the same parameter distribution, ensuring consistent semantic understanding across both stages while improving matching value accuracy without proportionally increasing complexity
Solution Approach 2:
The shared semantic representation model serves multiple functions: it performs semantic encoding for recall operations and semantic encoding for ranking operations simultaneously. This multi-functional design allows a single model to support both recall and ranking tasks with consistent parameter distributions, resolving the contradiction between precision and complexity
2Reliability
If a shared semantic representation model is used for both recall and ranking stages, then consistent semantic representation is achieved, but the model must handle multiple optimization targets simultaneously
Solution Approach 1:
The patent segments the optimization process into distinct phases: the semantic representation model is trained independently with its own loss function and optimization targets, then the recall model and ranking model are trained separately with their respective optimization goals. This segmented training approach allows the shared semantic model to provide consistent representations while each downstream model adapts to its specific optimization targets
Solution Approach 2:
The semantic representation model is trained in advance as a preliminary step before training the recall and ranking models. This preliminary training establishes consistent parameter distributions and semantic understandings that both subsequent models can rely on, ensuring reliability while maintaining the flexibility of separate optimization stages
Data Source
AI summary
A search method based on a neural network model is provided. The neural network model includes a semantic representation model, a recall model, and a ranking model. The present disclosure relates to the field of artificial intelligence, and in particular to the technical field of search. An implementation of the method comprises: inputting a target search and a plurality of objects to be matched to the semantic representation model to obtain a first output of the semantic representation model; inputting the first output of the semantic representation model to the recall model, and obtaining at least one recall object matching the target search from the plurality of objects to be matched by using the recall model; and inputting a second output of the semantic representation model to the ranking model, and obtaining a matching value of each of the at least one recall object by using the ranking model.